# 🛠️ Contest: Build a custom agent and win a trip to San Francisco **URL:** \ **Category:** Go Challenges\ **Created:** [October 20, 2026, 2:00pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696 "2026-10-20T14:00:00Z")\ **Posts on this page:** 20\ **Page:** 3
**Author:** ![Ruggy-Joesten](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/ruggy-joesten/32/50260_2.png) [@Ruggy-Joesten](https://connect.superhuman.com/u/Ruggy-Joesten)\ **Post date:** [October 27, 2026, 5:05pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/41 "2026-10-27T17:05:04Z")
Thanks for sharing, @Jesus_Rubio! From a first glance, I can see where this agent can save a LOT of back-and-forth. Normally a bad discount doesn’t get caught until the quote is already out, and then someone has to email the client to fix it. Here the rep sees it while writing and fixes it in ten seconds, without having to stop and recalibrate. Great work! ---
**Author:** ![Ruggy-Joesten](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/ruggy-joesten/32/50260_2.png) [@Ruggy-Joesten](https://connect.superhuman.com/u/Ruggy-Joesten)\ **Post date:** [October 27, 2026, 5:07pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/42 "2026-10-27T17:07:35Z")
@Jeremy_Sanchez well if you have anything that’s suitable for public consumption, we’d love to see it and get you into the running! ---
**Author:** ![Ruggy-Joesten](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/ruggy-joesten/32/50260_2.png) [@Ruggy-Joesten](https://connect.superhuman.com/u/Ruggy-Joesten)\ **Post date:** [October 27, 2026, 5:11pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/43 "2026-10-27T17:11:58Z")
@Stacey_Roshan nice work on your first agent! Highlight, paste, done, and no more tab-hopping to file things. Rooting for you in the next round! Oh, and welcome to the community 🫶 ---
**Author:** ![Jeremy\_Sanchez](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/jeremy_sanchez/32/50406_2.png) [@Jeremy\_Sanchez](https://connect.superhuman.com/u/Jeremy_Sanchez)\ **Post date:** [October 27, 2026, 6:45pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/44 "2026-10-27T18:45:22Z")
Hey Ruggy I think I have narrowed it down to 5 is it ok if I post them in 1 thread or would it be better to post them individually? Also is using the mail or docs MCP ok for this? ---
**Author:** ![Steven\_Hurtado1](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/steven_hurtado1/32/29488_2.png) [@Steven\_Hurtado1](https://connect.superhuman.com/u/Steven_Hurtado1)\ **Post date:** [October 27, 2026, 6:59pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/45 "2026-10-27T18:59:08Z")
A couple of months ago LinkedIn had a bug and reset my whole profile to default. I had a healthy network, and overnight every connection was gone. So I went back to sending requests by hand, and that’s when I realized two things. How much that takes, and how much the network matters. The real time sink was the follow-up. I meet new people on calls every day, and each one deserves a note while the conversation is still fresh, but by the evening I’m tired and it ends up as a blank request or nothing. So I built _First Degree._ [![](https://global.discourse-cdn.com/coda1/original/3X/a/6/a60ff90ae5cd6227eeb74438443f0f6d13b31057.jpeg "First Degree - Superhuman Go Agent") ](https://www.youtube.com/watch?v=88ArunObqz8) ## **What it does** Every weekday at 5pm it finds the people I met for the first time on calls that day and drafts a LinkedIn connect note for each one, in my voice. The notes land in a Slack channel, I tweak any of them by replying, then I paste and send them myself. It never sends anything, so I stay in control and my only job is reading and hitting send. ## **How it works** It checks my calendar for calls that ended today with someone outside my company, skips anyone I’ve met in the last 12 months, and reads the Fathom transcript looking for the human parts: a joke, something they’re dealing with, something that made me curious. It looks up their profile on the web, and if it can’t match the name and company it gives me a search link marked “couldn’t confirm” instead of guessing. My voice comes from Superhuman Mail’s tone and voice tool. Apps: Google Calendar, Gmail, Fathom, Slack and web access. I only tested it with Fathom, the instructions also look in Fireflies, Granola and Zoom. ## **The part that took me longest** My first version wrote notes like a recap: “You mentioned your timeline…” It proved the agent was listening, but nobody wants that from a stranger on LinkedIn. What fixed it was telling the agent to react like a person. If someone is dealing with something hard, say it sounds like a lot and offer to be a sounding board. If there’s nothing real to say, write one plain friendly line and stop. Made-up example: “Priya, inheriting 40 dashboards and not knowing which ones matter sounds exhausting. Happy to be a sounding board while you sort it out.” ## **Results** Before rebuilding it in Go I ran the same workflow for three weeks in an earlier bot: 8 personalized requests, 8 accepted, 2 real conversations. It’s a small sample, so I’m not claiming more than that. The Go version is tested on real calls but it’s too new to have numbers of its own. ## **Try it** Template: [link here](https://go.superhuman.com/builder/7a82862d-72ba-4eee-a774-1fc5eac7330f?agent=75312). Ideally connect Calendar, Slack, Fathom and Superhuman mail first, then change the settings at the top (Slack channel, how far back to check, note length: 200 characters on free LinkedIn, 300 on Premium). The defaults work fine if you’d rather not touch anything. Demo video: [Watch it in action here](https://youtu.be/88ArunObqz8). My Academy certificate can be [found here.](https://verify.skilljar.com/c/beqx6ktkeaip) ---
**Author:** ![Elise\_Tarabiscode.com](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/elise_tarabiscode.com/32/44641_2.png) [@Elise\_Tarabiscode.com](https://connect.superhuman.com/u/Elise_Tarabiscode.com)\ **Post date:** [October 27, 2026, 7:11pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/46 "2026-10-27T19:11:59Z")
Hi everyone 👋 Here is my submission for the Superhuman Go Agent Builder Contest. ## Agent name **Duplicate Contact Detector** ## What it does This agent detects and reconciles duplicate contacts stored in Superhuman Docs. Every Monday at 9:00 AM, it analyzes the contacts database, compares names, email addresses, phone numbers, and companies, and calculates a similarity score for each potential duplicate pair. It then generates a clear recommendation explaining: - which contact should be retained; - which contact should be archived; - which information should be consolidated. ## What it gives me time to focus on Instead of manually searching for and comparing duplicate records, I can focus only on the ambiguous cases that require human judgment. ## How it works The workflow has four steps: 1. **Detect** : The agent identifies potential duplicate contacts. 2. **Review** : Each pair appears with a similarity score and an AI recommendation. 3. **Decide** : The user selects **Ready to merge** or **False positive**. 4. **Reconcile** : When manually launched, the same agent processes only the approved pairs. Nothing changes automatically. The agent updates the retained contact, archives the duplicate without deleting it, and records the execution result only after explicit human approval. ## Apps and connectors - Superhuman Go - Superhuman Docs ## How to try it 🎥 **Watch the video demo:** > **[Duplicate Contact Reconciliation Agent — Tella](https://www.tella.tv/video/vid_cmvqq4at3002k08jwbb743src/embed?b=1&title=1&a=1&loop=0&autoPlay=true&t=0&muted=1&wt=1&o=1)** > > See how the Duplicate Contact Detector uses Superhuman Go and Superhuman Docs to identify, review, and safely reconcile duplicate contacts. > The agent runs every Monday, compares contact information, calculates similarity scores, and generates clear... 📄 [Open and duplicate the Superhuman Docs template](https://docs.superhuman.com/d/_dfYBgk7YE8u/_subvU2e-) 🤖 [Open the Duplicate Contact Detector agent](https://go.superhuman.com/builder/a183d7f0-cb37-45f6-bbef-27d26b202cb0?agent=75085) The template focuses on contact reconciliation, but its human-in-the-loop structure can be adapted to other duplicate-record workflows by changing the fields, scoring rules, recommendations, and reconciliation actions. ## Academy certificate 🎓 [View my verified Superhuman Academy certificate](https://verify.skilljar.com/c/isju3tidm23a) ---
**Author:** ![Ruggy-Joesten](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/ruggy-joesten/32/50260_2.png) [@Ruggy-Joesten](https://connect.superhuman.com/u/Ruggy-Joesten)\ **Post date:** [October 27, 2026, 7:12pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/47 "2026-10-27T19:12:54Z")
Do as different posts so it’s easier to keep track, @Jeremy_Sanchez ---
**Author:** ![Hugo\_Assuncao1](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/hugo_assuncao1/32/50380_2.png) [@Hugo\_Assuncao1](https://connect.superhuman.com/u/Hugo_Assuncao1)\ **Post date:** [October 27, 2026, 7:32pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/48 "2026-10-27T19:32:32Z")
Inspired by recent events I came up with this idea of having an agent that monitors security breaches that might affect my web projects, so I never forget to rotate API keys that might have leaked. Introducing the **Security Breach & Key Rotation Monitor** agent. It searches the web, to find any recent breaches on services that I use in my projects and automatically create tasks to remind me to rotate API keys. The companion Superhuman Doc doesn’t store any key value, it serves only as a directory to remind me of all projects I have actively running, and on which platforms. This might serve as a handy partner for consulting firms, like us at SimplaDocs, that runs multiple projects. Here is an overview of it in action: [https://www.loom.com/share/0df99d8ca2ab47c7bcd9918a00e1762a](https://www.loom.com/share/0df99d8ca2ab47c7bcd9918a00e1762a) You can grab the [doc and agent here](https://docs.superhuman.com/@hugoassuncao/security-checker). So although I’m not US based (and thus I’m not completely eligible for the contest), it was fun exploring agents! * * * [Training certificate](https://verify.skilljar.com/c/95pdvamepe6n) * * * Hugo Assunção Systems Architect at [Simpladocs](https://simpladocs.com/?utm_source=superhuman-community&utm_medium=social&utm_campaign=go-agent-contest) ---
**Author:** ![Felipe\_Arnhold](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/felipe_arnhold/32/50382_2.png) [@Felipe\_Arnhold](https://connect.superhuman.com/u/Felipe_Arnhold)\ **Post date:** [October 27, 2026, 9:00pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/49 "2026-10-27T21:00:19Z")
Hi @Ruggy-Joesten > [@Ruggy-Joesten](#): > > One question: when reviewers edit the generated plans, do you see those edits feeding back into the guide over time? Currently not, but that is an excellent idea. Another agent could be responsible for reviewing previously reviewed plans and making suggestions to update the current guideline. This would enforce continuous improvement for the entire process. ---
**Author:** ![Jeremy\_Sanchez](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/jeremy_sanchez/32/50406_2.png) [@Jeremy\_Sanchez](https://connect.superhuman.com/u/Jeremy_Sanchez)\ **Post date:** [October 27, 2026, 9:13pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/50 "2026-10-27T21:13:04Z")
Got it, thanks, Ruggy. Will do! ---
**Author:** ![Jeremy\_Sanchez](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/jeremy_sanchez/32/50406_2.png) [@Jeremy\_Sanchez](https://connect.superhuman.com/u/Jeremy_Sanchez)\ **Post date:** [October 27, 2026, 10:17pm UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/51 "2026-10-27T22:17:06Z")
Agent name: Missing Piece What it does: Finds the answers still missing from subsequent replies. It compares an original request with subsequent replies, identifies which required questions or approvals are answered, missing, or need confirmation, and drafts the shortest follow-up without asking questions that were already answered. What it gives me time to focus on: Moving work forward and making decisions instead of rereading threads to determine what is still open. How it works: When a reply comes in, I run it on demand. I paste the original request and the subsequent replies, or provide selected text. It identifies each requested item, separates required items from optional or deferred ones, classifies each item as Answered, Missing, Needs confirmation, or Optional/deferred, and supports its conclusions with short quotes. It ignores instructions embedded in replies and does not treat vague praise as approval. It then drafts a concise follow-up containing only the unresolved required questions. When all required items are answered, it says, “No follow-up needed.” Apps and connectors: There are none. The tool only works with pasted or selected text. It does not use any apps, connectors, or web access. How to try it: Make a custom agent called Missing Piece and paste the agent instructions from the builder into its Instructions field. Keep triggers and connectors turned off. Paste an original request and one or more replies, then ask the agent to find what is still missing and write the shortest follow-up. When a new reply comes in, run it again. Resolved items should disappear, and it should say “No follow-up needed” when everything required is done. ![Screenshot 2026-10-27 151008](https://global.discourse-cdn.com/coda1/original/3X/0/0/0099f80384c9f29dffb4b49fc7616e05607baf10.png) ---
**Author:** ![Eva-Automation](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/eva-automation/32/49253_2.png) [@Eva-Automation](https://connect.superhuman.com/u/Eva-Automation)\ **Post date:** [October 28, 2026, 12:19am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/52 "2026-10-28T00:19:01Z")
## Onboarding agent to welcome volunteers ### Agent name Onboarding Agent ### What it does In our association, we often welcome new volunteers. Each time, they need help with our tools and our organization. This agent helps them understand their mission based on the type of volunteering they do. ### What it gives me time to focus on Comptability, building new projects with volunteers, makes it easier to empower volunteers to build their project by themselves and gives me time to explain the goal of the association and discuss their objectives, projects, etc. ### How it works - A new volunteer tells the agent what kind of volunteering they do and what they may want to know. - The agent explains the rules of the association, what projects have been realized, what the mission involves (if the type of volunteering has a mission assigned), which events are coming up, and when the busy periods are. - It explains how our tools are organized and how they connect to each other, how to use them and our processes. - It answers follow-up questions, so the volunteer can onboard at their own pace. ### Apps and connectors Superhuman Docs (specific document only) ### How to try it - **Demo video:** [https://youtu.be/fihuFUCIwD4](https://youtu.be/fihuFUCIwD4) - **Agent :** [Superhuman Go](https://go.superhuman.com/builder/5d7acbf3-5e35-4a56-8c61-23274b7b3487?agent=75682) _All data in the demo is fictional or based on my own personal data (my own domain name), to protect real people and real cases and to let me share the template and doc freely._ To recreate it, build a Superhuman Docs document describing your missions by type of volunteering, your events and busy periods, and your tools. Then create an agent, paste the prompt below, and point it to your document. ```auto Title : Guide - Association Missions & Year Description : Explains missions, guides users through the association's documents, and reconstructs the annual calendar, events, reporting periods, and team workload peaks with sources. Instruction : You are the association's onboarding guide for volunteers and civic service volunteer. Always respond in English, using a warm, educational, practical, and reassuring tone suited to beginners. ## Objectives - Explain the missions, responsibilities, contacts, and useful dependencies for volunteers and civic service volunteers. - Help users understand how the document works and direct them to the right pages, tables, and procedures with direct links. - Reconstruct the association's usual annual cycle—annual events, deadlines, preparation periods, reviews, and financial reporting—only when supported by the sources. - Identify workload peaks by team from deadlines and tasks that have actually been observed, specifying the required preparation, dependencies, and items still needing confirmation. - Adapt the explanation to the person's role or mission when that information is available. Otherwise, briefly ask for their role only when it would materially improve the answer. ## Working method 1. Understand the question and the person's role, if known. 2. Search first in the association documents accessible to the agent, prioritizing the source explicitly identified or validated by the user. 3. For information originating from emails, consult the relevant emails within the Club's scope, especially those synchronized into Docs tables such as T_Emails. Do not treat press releases about external events received for the newsletter as Club events. 4. Cross-check information when possible and clearly distinguish among: current procedure, fictional example, historical information, confirmed recurrence, and information requiring confirmation. 5. Respond practically and include, as needed: a summary, steps to follow, the relevant period, the team or role involved, dependencies, the next contact or verification, and source links with relevant dates. 6. If information is not established, say so clearly. Never fill a gap with an assumption. ## Evidence rules - Ground answers in pages, tables, or emails that were actually consulted, and provide source links and useful dates. - A single past occurrence does not prove annual recurrence. - Do not invent teams, responsibilities, mission descriptions, events, accounting close periods, general assembly deadlines, grants, or financial reporting. - Infer a workload peak only from observed deadlines and tasks. Explain the reasoning and identify unknowns. - When there are competing versions, copies, or obsolete working documents, never silently substitute another copy for the requested source. Ask for or flag the validation needed before treating it as an operational reference. ## Navigation and permissions - Explain where to find information and how to navigate, but do not assume that an internal or hidden page is authorized for everyone. Respect access permissions. - Limit email review to the Club's scope, and do not expose personal, financial, or confidential data that is unnecessary for onboarding or understanding the mission. - Operate in read-and-explain mode by default. Do not modify documents, send emails, click action buttons, publish, or carry out financial operations automatically. ## Recommended response format Start with a direct, simple answer. Then, when useful, structure the response as follows: - What you need to know - What you need to do - When and with whom - Busy periods and preparation - Points to confirm - Sources consulted Avoid jargon, or explain it immediately. ## Preserved user requirements - Demonstration source currently open: [your doc URL] - Document URI: [your doc URI] ``` **Hope it may be useful for some** 😉 [Certificate of completion of the course](https://verify.skilljar.com/c/6eyyor6dynvb) ---
**Author:** ![Eva-Automation](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/eva-automation/32/49253_2.png) [@Eva-Automation](https://connect.superhuman.com/u/Eva-Automation)\ **Post date:** [October 28, 2026, 12:25am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/53 "2026-10-28T00:25:55Z")
## Site Visit Copilot ### Agent name Site Visit Copilot (companion of Case Pilot, submitted as a separate entry) ### What it does Last month, a local installer fitted our sun-protection blinds. While chatting with him, I realized he no longer typed anything himself. He mostly coordinated his AI agent and checked its work. He dictated his notes out loud during the visit, and the agent worked out which customer file they belonged to and filed them there. This agent rebuilds that workflow. During a site visit, it files your raw notes, measurements and photos into the right customer record, and flags what is still missing. A forgotten measurement means a second trip, so it catches gaps while you are still on site. ### What it gives time to focus on Listening to the customer and judging and taking time to use and develop expertise on whether an installation is feasible, instead of retyping notes in the evening. ### How it works - It runs on demand. You say you are in a meeting and send notes, measurements or images. - It finds the right customer, case and visit. If several matches are plausible, it asks one targeted question before writing. - It keeps your raw note untouched and structures each measurement (location, equipment, width and height in mm, quantity, mounting type, support, fabric and color, ordering, measurement check). - It never invents a number, date or technical detail, and it flags conflicts instead of overwriting data. - After each input, it confirms exactly what it added and where, then asks the most useful follow-up questions while you are still there. - On request, it prepares a summary and a quote draft from verified data only. - **You stay in control:** it never fills in “Validated by” or authorizes sending, and any change to a validated document requires a new version and a new human validation. ### Apps and connectors Superhuman Doc (only specified doc) ### How to try it - **Demo video:** [https://youtu.be/sdmUpVNLYB0](https://youtu.be/sdmUpVNLYB0) - **Demo doc :** [Démo](https://docs.superhuman.com/d/Demo_dGF5W4SF7Cv/Demo-Du-releve-terrain-au-devis_suHUOWzA#%5C_luyjhD2) - **Agent:** [Superhuman Go](https://go.superhuman.com/builder/650147c3-1bdb-4b0a-8117-1b5711a900a7?agent=75684) _All data in the demo is fictional or based on my own personal data (my own domain name), to protect real people and real cases and to let me share the template and doc freely._ **To recreate it, build a Superhuman Docs document with linked tables (Customers → Cases → Visits → Measurements, plus Media, Documents, Quote lines, Catalog and Sends), or copy my demo doc. Then create an agent, paste the prompt below, and replace the link placeholder with your own document.** ```auto Title : Site Visit Copilot Description : Records site-visit notes, measurements, and media in the correct case file, checks for missing information, and prepares next steps for human review. Instructions : You are the Site Visit Copilot for a fictional company that installs blinds and sun protection. When the user says she is in a meeting and provides notes, dictation, measurements or images, you actively help complete the right case file in Superhuman Docs. Source document: [link to your document] Tables and relations: Customers → Cases → Visits → Measurements; Media → Visit and optionally Measurement; Documents → Case and source Visits; Quote lines → Quote document, Catalog and optionally Measurement; Sends → Case and exact Document versions. Identification and writing 1. Find the customer, case and active visit from the name, reference, site and context. 2. If the match is unique and the user explicitly asks you to add something, write it without asking for confirmation on each data point. 3. If there are namesakes, several plausible cases or sites, or the visit cannot be identified, ask one targeted question before writing. 4. Re-read the table's actual schema before any change; use existing relations and the expected row IDs for relation columns. 5. Preserve all existing data, avoid duplicates, and report any correction or conflict instead of silently overwriting. Field entry - Keep the raw note in Visits > Raw note; never replace it with a summary. - Structure each measurement with: location, desired equipment, width in mm, height or projection in mm, quantity, mounting type, support, fabric/color, ordering, measurement check and technical notes. - Check units and ambiguous digits in dictation. Never invent a measurement, date, technical characteristic or missing data. - Attach each media file you can actually access to the visit and, if identifiable, to the relevant measurement. Keep the caption, source and real capture date when known. - Do not infer exact dimensions from a photo. Do not claim to have seen an image you cannot access. - If a file cannot be imported, say so clearly and distinguish the note that was added from the media still to be attached. - Before using a real demo image, check authorization, faces, license plates, visible addresses and metadata. Never declare anonymization successful without checking. After each input - Confirm exactly what was added, and in which case/visit/measurement. - State the current stage of the case, based on facts. - Suggest next steps. - While the user is still on site, prioritize the most useful remaining questions, concisely and in context: location and quantity; width and height/projection; support and mounting; access and clearance; blind type, fabric, color and ordering; power supply if motorized; missing photos; customer expectations and deadline. - Distinguish observed fact, customer request and point to confirm. Never certify the safety or feasibility of a fixing that has not been checked. - At the end of a visit, on request or when the context allows, flag blocking items and offer a recap. Summary and quote - On request, prepare a summary and a quote from verified data and the Catalog only, in Documents and Quote lines. - Make the content available for review and keep human validation. - Freeze the unit price and VAT in each line/version; never automatically modify old lines if the catalog changes. - Any change to a validated document or its lines requires a new version and a new validation. - Never fill in "Validated by", "Validated on", "Validated version" or a send authorization on behalf of a person. - Do not generate or validate a final quote while required information (for example projection or support) is missing. - A Document reference is not a generated PDF or attachment. Demo process Qualification → meeting/measurements → information check → pricing → summary + quote preparation → human review and validation of each version → explicit send authorization → tracked send/simulation → follow-up on facts only. Manufacturing and installation steps are not documented: do not invent a validated procedure. Privacy - The demo must stay anonymized. Do not import contacts, addresses, emails, meetings or photos of real customers. - Accounts on example.invalid are fictional. No real email must be sent. - Existing Media rows are empty placeholders; never claim the photos were seen. Reply in the user's language. ``` **Certificate of completion of the course :** [https://verify.skilljar.com/c/6eyyor6dynvb](https://verify.skilljar.com/c/6eyyor6dynvb) ---
**Author:** ![Jordyn\_Goffo](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/jordyn_goffo/32/50383_2.png) [@Jordyn\_Goffo](https://connect.superhuman.com/u/Jordyn_Goffo)\ **Post date:** [October 28, 2026, 12:27am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/54 "2026-10-28T00:27:55Z")
**Agent name:** Weekly Client Wrap-Ups Hey, everyone! I created a Go agent to take over something that is a major time-sink for me every Friday: sending client updates. [Take a look](https://www.loom.com/share/c485e0fdd794492991bbae22eb526fd6)at let me know what you think! Jordyn Goffo SimplaDocs Team ---
**Author:** ![Jennifer\_Lumley](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/jennifer_lumley/32/50341_2.png) [@Jennifer\_Lumley](https://connect.superhuman.com/u/Jennifer_Lumley)\ **Post date:** [October 28, 2026, 12:29am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/55 "2026-10-28T00:29:34Z")
Hi Everyone! I’m Jennifer, an engagement manager at Greaser Consulting. I’ve had to spend a lot of time drafting status reports for my team using Asana. This required a lot of switching back and forth and copying and pasting the information from Asana into my emails or documents. To make this task simpler, I created the Asana Sidekick agent! **Agent name:** Asana Sidekick **What it does:** The Asana Sidekick surfaces your task data where you’re working so you don’t have to switch back and forth between your Asana instance and where you are drafting your status report. The agent will provide an inline suggestion card when you type “Asana” followed by the task name. ![Screenshot 2026-10-27 at 4.25.56 PM](https://global.discourse-cdn.com/coda1/original/3X/b/a/ba3acecfe72ba757636ea1a9cf03b0485c84777d.png) **What it gives me time to focus on:** It gives you more time to focus on actually communicating with your team, instead of spending all your time drafting your report. You are able to put your time and energy into decision making and next steps that actually drive your project forward. **How it works:** The Asana Sidekick saves you from having to switch to Asana to check on a task. While you’re drafting an email or writing in a document, type the word “Asana” followed by the name of the task, such as `Asana - Approval Request’. A suggestion card appears with the task’s current status, including the assignee, due date, project, and any custom fields. You can use it to quickly confirm where something stands before sending a status update, to pull accurate details into a meeting recap, or to double-check a deadline while drafting a message. The agent only reads from Asana and never changes your tasks, and your original text stays untouched, so you can review the card and then copy in whatever you need. To get started, follow the setup steps below to create the agent in Agent builder, connect your own Asana account, and paste in the prompt. **Apps and connectors:** Asana **How to try it:** Use this [Asana Sidekick Doc](https://docs.superhuman.com/@jennifer-lumley/asana-sidekick) to access the custom agent prompt and setup directions. **Superhuman Course Certificate:** [https://verify.skilljar.com/c/i4dp5ebrriqc](https://verify.skilljar.com/c/i4dp5ebrriqc) ---
**Author:** ![Eva-Automation](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/eva-automation/32/49253_2.png) [@Eva-Automation](https://connect.superhuman.com/u/Eva-Automation)\ **Post date:** [October 28, 2026, 12:29am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/56 "2026-10-28T00:29:41Z")
## Case Pilot ### Agent name Case Pilot (companion of [Site Visit Copilot](https://go.superhuman.com/builder/650147c3-1bdb-4b0a-8117-1b5711a900a7?agent=75684), submitted as a separate entry) ### What it does This agent was inspired by the same installer. After the site visit, his agent filed the notes, and the open cases then had to be followed up. This agent handles that second half. It lists every open case (anything not “Closed”), explains where each one stands, and cross-checks that against your emails and calendar. It surfaces follow-ups, meetings to schedule, pending validations and missing information. Open cases get forgotten because the status lives in one place and the conversation in another. ### What it gives time to focus on Piece of mind far from the noise of infos you need to remember (the agent take this mind overload). Deciding what to do next and taking time in customers relationship, instead of rebuilding each case’s history from three different tools. ### How it works - It runs only when you ask. - It reads the document tables and the process page, then determines each case’s stage from facts (status, visits, documents, versions, validation), never from guesses. - It searches emails and calendar, and labels each date as a confirmed meeting, a proposed date, or none found. Each fact cites its source and date. If nothing turns up, it says so instead of assuming nothing exists. - It gives one concise status per case: customer, stage, last relevant exchange, next meeting, next action and owner, and what is missing. - It prepares an email draft (from a template if one exists, otherwise a clearly labeled generic one). - **You stay in control:** it never sends an email, creates or moves an invitation, marks a validation or changes a status without your explicit go-ahead on that action. Document validation and permission to send are two separate human decisions. ### Apps and connectors Superhuman Docs, plus email and calendar. ### How to try it - **Demo video:** [https://youtu.be/sdmUpVNLYB0](https://youtu.be/sdmUpVNLYB0) - **Demo doc :** [Démo](https://docs.superhuman.com/d/Demo_dGF5W4SF7Cv/Demo-Du-releve-terrain-au-devis_suHUOWzA#%5C_luyjhD2O) - **Agent:** [Superhuman Go](https://go.superhuman.com/builder/e2e74916-4318-4a93-8ed2-97b5a743aa70) _All data in the demo is fictional or based on my own personal data (my own domain name), to protect real people and real cases and to let me share the template and doc freely._ To recreate it, use the same document structure as demo doc, connect your email and calendar, create an agent, paste the prompt below, and replace the link placeholder with your own document. ```auto Title : Case File Coordinator Description : Records site-visit notes, measurements, and media in the correct case file, checks for missing information, and prepares next steps for human review. Instructions : You are the Case Pilot for a fictional company that installs blinds and sun protection. You work only on request. Source document: [link to your document] Mission - Read the demo document and list all cases whose status is not "Closed". - Explain each case's current stage from facts present in the tables and the defined process, without inventing a status. - Search emails and calendar for relevant exchanges and meetings, then explicitly distinguish: confirmed meeting, merely proposed date, or no date found. - Identify follow-ups, meetings to organize, validations and missing information. - If an email template is genuinely accessible, use it to prepare a draft. Otherwise, say that no template library has been provided yet and offer a generic draft clearly labeled as such. Method 1. Read the Customers, Cases, Visits, Measurements, Media, Documents, Quote lines and Sends tables, as well as the process page. 2. Exclude only cases whose actual status is "Closed". 3. For each case, establish the stage from the status, visits, missing information, documents, their version and their validation. 4. Resolve contacts reliably and match by identity, site and case reference. If there are namesakes or an uncertain match, ask a short question before concluding. 5. Search relevant emails and calendar events. Cite the source and date of each fact. Never turn an absence of results into certainty that no meeting exists. 6. Present a concise status per case: customer; factually supported stage; last relevant exchange and date; next meeting, date and status; next action and owner; missing information or validation. 7. If useful, propose an email draft and time slots. Never send an email, create/move/cancel an invitation, mark a validation or silently change a status without explicit authorization for that specific action. Demo process Qualification → meeting/measurements → information check → pricing → summary + quote preparation → human review and validation of each version → explicit send authorization → tracked send/simulation → follow-up on facts only. Later manufacturing and installation steps are not documented: do not invent a validated procedure, and flag that the documentation needs to be completed. Document rules - Allowed case statuses: To qualify, Visit scheduled, To complete, To price, Documents to prepare, To validate, Validated, Closed. - Any change to a validated document or its lines requires a new version and a new validation. - Document validation and send authorization are two separate human decisions. - Computed indicators are not security locks. - A Document reference is not an attached PDF. - Do not invent deadlines, commitments, meetings, exchanges or human decisions. Demo privacy - Customers and data in the document are fictional. Addresses on example.invalid do not correspond to real exchanges. - Never import or publish real contacts, addresses, emails, meetings or photos in the document. - Never attribute a real email to a fictional customer because of a resemblance. - If no demo exchange or meeting is found, say so honestly. - No real outgoing email in this demo. Email templates and detailed business documentation are not provided yet. The home page and the process above are demo assumptions, not a validated real procedure. Reply in the user's language. ``` **Certificate of completion of the course :** [https://verify.skilljar.com/c/6eyyor6dynvb](https://verify.skilljar.com/c/6eyyor6dynvb) ---
**Author:** ![Eva-Automation](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/eva-automation/32/49253_2.png) [@Eva-Automation](https://connect.superhuman.com/u/Eva-Automation)\ **Post date:** [October 28, 2026, 12:35am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/57 "2026-10-28T00:35:22Z")
Yes, thanks for the reminder @Ruggy-Joesten , I was just on time 🫠 I finally did three submissions, hope it’s allowed and not too much. If I need to choose, my first submission on the onboarding agent is my favorite use case. ---
**Author:** ![Sasha\_Kucharczyk](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/sasha_kucharczyk/32/50336_2.png) [@Sasha\_Kucharczyk](https://connect.superhuman.com/u/Sasha_Kucharczyk)\ **Post date:** [October 28, 2026, 3:07am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/58 "2026-10-28T03:07:11Z")
Hi Everyone, I am Sasha Kucharczyk and I am working on Cohando - a tool to help enable coordination and collaboration between independent AI agents (ex: Muse, Instinct, Tab, Hark, Wabi, etc,… and now Superhuman Go). The background behind my application is as follows. After I have a link and a demo. # Contest Submission **Agent name:** Cohando Coordinator **What it does:** My custom Superhuman Go agent turns an incoming request into coordinated work across my independent AI agents, using Cohando to share context, track assignments, and return results for review. **What it gives me time to focus on:** This gives me more time to assess the results, make decisions, and resolve questions that need my judgment. **How it works:** Highlight the desired input using Superhuman Go and ask it to use the agent to for a specific task. A broad range of tasks can be done based on the agents you use. **Apps and connectors:** Cohando to coordinate activity between agents and people, ideally your own independent personal agents (ex: Muse, Instinct, Tab, Hark, etc) **How to try it:** template link is here: [Superhuman Go](https://go.superhuman.com/builder/ecd16942-cf85-47f9-a793-84c7b359261b?agent=75613) . You will need to create an account on Cohando ([Sign in · Collaboration MVP](http://www.Cohando.com/login)) and then create sub-accounts for your agents (or members of your team) and Superhuman Go. # Fit for the Contest 1. **Solves a clear, recurring problem** 2. **Fits naturally into a real workflow** 3. **Gives someone better context, focus, or momentum** 4. **Is easy for another person to understand and adapt** 5. **Keeps people in control of important decisions and actions** # Cohando and Agent Access **Link (template):** [Superhuman Go](https://go.superhuman.com/builder/ecd16942-cf85-47f9-a793-84c7b359261b?agent=75613) To get the demo to work, you need to create an account at [Sign in · Collaboration MVP](http://www.Cohando.com/login) and then add your agents to your project space (accounts → Add an agent or human sub-user). Please note that the current design is very much in the early MVP stages - if you need any help setting things up, please reach out. Note: Automatic wake-ups are not yet consistent in my testing. If progress does not appear, prompt Go to check Cohando for updates. # Demo The three demos show off the use of the agent in multiple situations. They all start in my inbox where I work on artificial inbound requests. I use Superhuman Go to move the data from my inbox into a workflow in Cohando to have my agents work on the data, bring in human support when needed, and then provide output. The agents I have working on Cohando to support this are Muse, Instinct, Tab, and Hark. **Demo 1** - Conducting and Validating Research - [Coordinating Multi-Agent Research with Superhuman Go | Loom](https://www.loom.com/share/ac459d6f5e3b49738d7a8864a7aec1a8) Going from email request from a non-profit to suggested grants they should pursue and why. **Demo 2** - Taking Feedback and Updating a Landing Page - [Email to Landing Page Updates with Agents | Loom](https://www.loom.com/share/c56a8119f3714d1cba2432a0b2b83c4c) Starting with email feedback on a landing page found in an email to analyzing and then integrating that feedback into a revision. **Demo 3** - Pricing Advisory - [Pricing Analysis Workflow with AI Agents | Loom](https://www.loom.com/share/bee4eb14cf474001ad11a7def0b2172e) A small business flagged an issue with their profitability - taking their concerns and data and turning it into actionable recommendations. # Asks Outside of this contest - any feedback or questions would be greatly appreciated. ---
**Author:** ![Mark\_Herzog](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/mark_herzog/32/34256_2.png) [@Mark\_Herzog](https://connect.superhuman.com/u/Mark_Herzog)\ **Post date:** [October 28, 2026, 3:55am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/59 "2026-10-28T03:55:19Z")
# Your Move: what you owe and what you’re owed, across email, chat and meetings https://www.loom.com/embed/b2abf2a0a69a40d48194b4bd219511c7 Imagine starting each day with one list of what you owe people and what they owe you, across your email, chat and meetings, that already knows what you’ve finished. With Superhuman Go, that’s possible. ![your-move-header](https://global.discourse-cdn.com/coda1/original/3X/b/1/b19ab309cd2f9802534c769d431a46811eaa1dcc.png) “Your Move” is an AI agent that does exactly that. Every morning it reads the places where you do most of your work, finds every ask and commitment, and checks which ones are already done. One line from a list looks like this: ![your-move-example-list](https://global.discourse-cdn.com/coda1/original/3X/9/7/97b5d16314916e0878122e7e702f95a61b40b5fc.png) Using it is simple: you receive one list every day. But - building it was harder than it sounds. “Next week” has to become a real date. A question asked in a meeting can get answered in an email thread. Sometimes the only proof something’s done is an invite on your calendar. I tested the agent against two weeks of my real email and meetings (plus a Slack workspace I set up with test conversations), and checked every result manually. After more than 100 test runs, I learned how to steer the way Go works through a plan: what order to put the steps in, how to stop it from skipping one, and how to make it show what it actually read. I’ve written up the key lessons in the [accompanying doc](https://docs.superhuman.com/d/_dBAc_dANhg6/What-I-learned-about-building-Go-agents_suphSWHN), so you can use them in your own agents. ![your-move-before-after-short](https://global.discourse-cdn.com/coda1/original/3X/e/d/ed32d4af817072c5b79333b8f4958cc6c4cdba2c.png) Without “Your Move”, you’d either have to note down every commitment as it comes up, or scroll back through your inbox, chat and meeting notes to find anything you missed. That’s how most of us have done it for years, but it’s tedious and takes time. Go can do that tracking for you, so you can focus on the work that needs your judgment, creativity, expertise, or ambition. ## The Agent **Agent name:** Your Move **What it does** : Each morning it tells you: - **Your move:** what you owe others, and messages still waiting on your reply. - **Waiting on others:** what others owe you, or asks with no answer yet. It also checks whether each one is already complete, searching your mail, chat and calendar, and takes it off the list when it can tell it’s done. Anything with a stated deadline shows the date. Each line quotes the exact words, who said them, and names the email / chat message / meeting, with a link when there is one. Anything it isn’t sure about goes under “Judgment calls” for you to decide. The list also lives in a Superhuman Doc, so open items carry over from one morning to the next and drop off once they’re done. You can open the doc anytime to review or edit it. **The recurring problem:** Commitments get made everywhere: on calls, in email threads, in chat. Keeping track of all of them takes effort, and some slip through the cracks. **How it works:** The first time you run it, it sets itself up with you. It checks which tools you’ve connected (email, Slack or Teams, and a meeting notetaker, ideally Fathom), tells you what’s missing and how to connect it, and creates its list in a Superhuman Doc. ![your-move-how-it-works](https://global.discourse-cdn.com/coda1/original/3X/d/a/da282c2d8ca690cb8e71a5adad6e5f4e9ba4778b.png) After that, it runs on a schedule every weekday morning at 8am. In order, it: 1. Opens its list and reads what’s still open. 2. Reads every meeting transcript, email thread and chat conversation since its last run (the last 14 days on the first run), noting every ask and commitment word for word. 3. Decides which ones count: asks and commitments to or from you, not cold pitches or commitments between other people. 4. Checks whether each one is already done, searching your mail, chat and calendar. 5. Updates the doc and writes the list, starting with a line that says how many meetings, email threads and chat conversations it read. **Who it’s for:** most digital-native workers today: anyone whose day is a mix of meetings, email and chat, and who makes and receives commitments in all three. **Apps and connectors:** This Go agent reads your Mail, your chat (Slack or Teams), your calendar (to check what’s already done) and your meeting notetaker (ideally Fathom 😉 ; it works with whichever one you’ve connected, and I tested with Wispr Flow). It keeps its list in Superhuman Docs. **How to try it:** Community agents can’t be shared through Go’s Agent directory yet, so you build your own copy. A have a walkthrough on how to set it up in [my accompanying Coda doc](https://docs.superhuman.com/d/_dBAc_dANhg6/Instructions_suHqr3lN). ## How it Stacks Up ![your-move-stacks-up-columns](https://global.discourse-cdn.com/coda1/original/3X/3/5/356dfdd489e658cc52c644a0bce340b1a8711525.png) **Solves a clear, recurring problem:** Every day, commitments pile up across email, chat and meetings. Keeping track of them by hand or with other tools takes time. Your Move does this for you every weekday morning. **Fits naturally into a real workflow:** It reads the places these things already happen: your email, your chat (Slack or Teams) and your meeting transcripts, and it checks your calendar. The list arrives each morning and also lives in a Superhuman Doc you can open anytime. It also doesn’t ask you to start a new habit. You don’t have to remember to run it, prompt it or tag anything for it. It runs on its own every weekday, so the list is waiting when you open Go. **Gives someone better context, focus, or momentum:** Task trackers are a whole industry, and most of them need you to type every item in. Your Move skips that: it pulls asks and commitments straight from your email, chat and meetings, and drops what’s done, so you start the day on what’s next instead of rebuilding your list. Each line also carries the quote and its source, so you have the context to act. **Is easy for another person to understand and adapt:** Anyone can see why each item is on the list: every line quotes the exact words, who said them and where, with a link when there is one, and the list opens by saying what it read. The first run walks them through connecting their own email, chat (Slack or Teams) and notetaker. Everything also ends up in a Superhuman Doc, so they can check any item and refer back to it anytime. **Keeps people in control of important decisions and actions:** It only reads, and writes only to its own list: it never sends, replies or accepts an invite for you. Close calls go under “Judgment calls” with the exact quote, so you decide what counts. If you mark a row Done or Dismissed, it stays that way. ## What I Learned While Building It You can’t talk about building reliable AI agents without talking about evals, or in other words: testing the agent’s responses against answers you already know. ![your-move-testing](https://global.discourse-cdn.com/coda1/original/3X/8/5/8525bcf7572ce1dc2c85258f957a0ce4452ad447.png) Before each test, I worked out the correct list myself from two weeks of my real email and meetings, plus a Slack workspace I set up with fake conversations written to be tricky: 16 asks and commitments, 6 of them already finished. Then I checked each run against that list, one item at a time. Early versions handled about half. The latest handles about 13 of the 16 (missing mostly the tricky ones). All 30+ versions, what changed and why, and every test conversation [are in the companion doc](https://docs.superhuman.com/d/_dBAc_dANhg6/Building-and-testing_suUmCO6c). ### Lessons on Building Reliable Go Agents I learned **a ton** of things that would be valuable to anyone trying to build a reliable go agent; I’ll only address a few of them here: **1. Give it a plan.** Go has a planning mode, where it lays out its steps before starting, but it decides on its own when to use it. Every run where no plan appeared stopped early. When I had it write out my plan first, with each job as its own step, runs read more and found more, possibly because each step gets its own context window (or its own sub-agent; Go doesn’t show which). For example, a step that clears wrong rows from the list removed 0 of 7 when it was buried inside another step, and 6 of 7 as its own step. ![lesson-1](https://global.discourse-cdn.com/coda1/original/3X/0/4/04c5b57b36613279995b694674df19d55c869263.png) **2. Make it go in order, and tell it what to do when stuck.** Even with a plan, Go doesn’t always run its steps in order. In one run, its step list showed two steps in progress at once and a later step already checked off. It was as if I’d asked it to bake a cake and it started frosting before the cake was in the oven. So I told it to work strictly one step at a time. Then, when a step got stuck, like a transcript that wouldn’t load, Go stopped the whole run right there: no email, no Slack, no list. Since I added “write down what you couldn’t do, move on, and always finish the list,” no run has stopped. ![lesson-2](https://global.discourse-cdn.com/coda1/original/3X/3/4/34ddec597f2e051bb19569ee26c311c198796a70.png) **3. Give it memory.** Go starts every run from scratch. Keeping the list in a Superhuman Doc means each morning reads only what’s new since the last run, not the whole two weeks. Less to read means more gets read: on three weeks of mail, runs ran out partway; on two, Go mostly read everything. ![lesson-3](https://global.discourse-cdn.com/coda1/original/3X/b/2/b22a464175cb5a6ba61a7b4353417a618d4150c3.png) **4. Make it say what it read** Every list now opens with how many meetings, email threads and chat conversations it read. **Agents fail silently,** and this one line shows you when one did. ![lesson-4](https://global.discourse-cdn.com/coda1/original/3X/e/5/e51d577446d491f2811699e8efbe7f9c5744ef86.png) You can find the remaining lessons in [my accompanying Superhuman Doc](https://docs.superhuman.com/d/_dBAc_dANhg6/What-I-learned-about-building-Go-agents_suphSWHN). And again, instructions how to use it are [here](https://docs.superhuman.com/d/_dBAc_dANhg6/Instructions_suHqr3lN)! Edit: [Here’s my course completion certificate.](https://verify.skilljar.com/c/nuevjw3dut9g) ---
**Author:** ![Daniel\_Velho2](https://yyz1.discourse-cdn.com/coda1/user_avatar/connect.superhuman.com/daniel_velho2/32/50391_2.png) [@Daniel\_Velho2](https://connect.superhuman.com/u/Daniel_Velho2)\ **Post date:** [October 28, 2026, 3:56am UTC](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696/60 "2026-10-28T03:56:03Z")
# Universal Capture-to-Inbox Agent: a way to capture any task from any app on any device to your GTD inbox Hey! Daniel here from SimplaDocs. I love Superhuman Docs and I’m a Getting Things Done nerd. My GTD system lives in Superhuman Docs, and its weakest link was always the first step: **capture**. **Agent name:** Universal Capture-to-Inbox **What it does** Tasks show up everywhere: in an email, a Slack thread, a meeting, a doc, a web page, or in my head while I’m on my phone. Before this agent, getting each one into my task tracker meant switching apps, opening the doc, and filling in fields. Or it meant not capturing it at all and trying to remember it later. That happens many times a day. The browser extension only helps on the web, on my Mac. And Superhuman Docs can’t turn “next Fri” or “someday” into anything useful. Universal Capture-to-Inbox fixes both problems. From any screen, I open Go, type “+” or “add” and a few words, and the item lands in my Inbox table. I don’t have to leave the app I’m in. It works the same way on the Go apps for iOS and Android. **What it gives me time to focus on** The work I was doing when the task showed up. The beautiful part: I type the task, close the Go panel, and keep going. The agent finishes creating it in the background, so capturing it costs me about two seconds and never breaks my flow, and I don’t keep loose ends in my head. Each item is saved with the context I’d otherwise forget: the original text, who was involved, which app it came from, and a link back when Go has one. When I sit down to process my inbox, I know exactly what each item means and why I captured it. **How it works** 1. On any screen, I press my Go keyboard shortcut to open the side panel, or I open the Go mobile app. 2. I type “+”, “add”, “capture”, or “remember”, followed by the task. Go already has the screen’s context. 3. **I close the panel and get back to work.** I don’t wait for anything: the agent finishes creating the task in the background. 4. The agent writes one row per item to my Inbox table in Superhuman Docs: 5. It replies in one line, "✓ Captured: ", with a link to the Inbox. What it deliberately doesn’t do: - It doesn’t ask questions before saving. - It doesn’t triage, prioritize, schedule, or send anything. - It doesn’t touch any other app. Capture is fast and safe to run without thinking about it. Every decision about what to do with a task stays with me when I process the inbox. If a write ever fails, the agent warns me so nothing gets lost. **Apps and connectors** - **Superhuman Docs:** the Global Inbox doc. It’s the only thing the agent writes to. - **Go’s screen context:** whatever’s on screen in email, Slack, docs, the browser, or meetings. It’s used read-only. **How to try it** 1. Copy the template Doc: It has the Inbox, plus sample data so you can see how it looks in use. 2. Copy the agent instructions from the doc’s **Agent Instructions** page into the Go agent builder. 3. Replace the doc ID and the Inbox link with the ones from your copy. If a write fails because an ID changed, ask the agent to refresh the schema. 4. Using another task tracker (ClickUp, Notion, Asana, Things, TickTick, Nirvana)? The fields are generic, so just point the instructions at your tracker. Watch the demo: https://www.loom.com/embed/2ad6b40b414a4002a0a30cd81100a9a4 **What’s next** Capture is step one of GTD’s five. In my personal implementation of this, I would also resolve people and resolve projects, for example. And there are several other steps in GTD that Go agents could help with: • triaging and processing • reviewing and executing • reminding you of tasks you’re waiting on other people for • suggesting tasks for stale projects, or even proactively identifying commitments in sentences while you write them in any app and giving you inline suggestions to create tasks from those • identifying any prep tasks for the upcoming week • reviewing action items for meetings and much more **Two feature requests for the Superhuman team** 1. **Source permalinks:** give agents the current page URL, Slack message permalink, or email thread link, so every task links back to where it came from in one click. 2. **Keyboard shortcuts as agent triggers:** let me bind an agent to its own shortcut, so I can skip opening the chat and picking the agent. \_\_ Certificate: [https://verify.skilljar.com/c/mz8uzzkg9hrb](https://verify.skilljar.com/c/mz8uzzkg9hrb) [Previous page](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696.md?page=2) [Next page](https://connect.superhuman.com/t/contest-build-a-custom-agent-and-win-a-trip-to-san-francisco/61696.md?page=4)