Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

A Note-Taking Agent Ran a GitHub Board: Over Six Weeks, 8 Cards Went Overdue and 9 Were Forgotten

Ideas from long coding-agent sessions disappear when the context gets compressed or the terminal closes. The author tried five places to store them and settled on GitHub Issues with a Projects board: the agent creates task cards, and the human closes them. After six weeks, 8 newsletter cards were two weeks overdue and 9 hadn't been touched in over a month, because nobody was keeping the board in order. That means a board the agent writes to needs a daily check.

AI-processed from Habr AI; edited by Hamidun News
A Note-Taking Agent Ran a GitHub Board: Over Six Weeks, 8 Cards Went Overdue and 9 Were Forgotten
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The agent opens an issue and moves its status with the gh project item-edit command
The owneryou and owner:claude labels split tasks between the human and the agent
Key metric #3The most neglected card sat untouched for 42 days, and a closed issue got stuck in "waiting"

You spend half a day working with a coding agent. You notice a bug you can't get to right now, and you come up with a new feature. The agent agrees: "We should also..."

Then the context gets compressed, the terminal closes, and nobody remembers that "later." The author of a Habr write-up decided to log these loose ends on a GitHub Projects task board: every "later" becomes an issue (a task card), the agent acts as a secretary and creates the cards, and the human closes them. Six weeks later, the board had turned into a junkyard: 8 newsletter cards were two weeks overdue, 9 cards hadn't been touched in over a month, and one sat idle for 42 days.

The fault lies not with the agent or the human, but with the absence of anyone responsible for keeping the board current.

This problem isn't limited to programmers. Any extended work with an agent, whether writing, analytics, or reviewing someone else's project, leaves a similar tail of deferred tasks.

Five places for "later"

Scrolling back through the chat. The default option. It works until the context is compressed or the terminal or chat is closed. With agents, that's a few hours at most.

A notes file in the repository. This is NOTES.md, CLAUDE.

md, or the agent's own memory. It's good for facts about the project: which flags the build uses, what quirks the deploy script has. It doesn't work for tasks, because a line has no status or deadline, and nobody is responsible for closing it.

The author went further and built ai-tree: a tree of each session's ideas on a local page. Each branch has a status: untouched, in progress, done. A request in CLAUDE.

md to "keep the tree up to date" didn't work in long sessions, because the agent forgot. The working version reads the session log as an observer outside the agent. The author's conclusion: asking the agent to "remember it yourself" is pointless.

Claude Code's built-in task list. It's enabled with the environment variable CLAUDE_CODE_ENABLE_TODO_TOOLS=1. Tasks are stored in ~/.claude/tasks/<session>/, one JSON file per task, and a new session sees the old list. The author didn't go back to it: the list lives inside the agent, you can't find your own plans there, and another person won't look there.

Jira. If this is company work, the question is settled: the process is documented and the fields are agreed on. The agent only needs rules for maintaining tickets or a few examples, and it will create tickets that look like everyone else's.

GitHub Issues plus a Projects board. This is the author's personal setup. The owner:you label marks what only a human can do: register an app, make a decision, publish. The owner:claude label marks code and documentation. The agent creates a card, sends a link, and moves the status with gh project item-edit. The human only closes the issue, and closed means "done." In the next session, the agent reads the board through gh, so the task survives the session.

The approach has an unexpected bonus: the board can be public. An outside user found an issue about migrating the author's MCP server to a new revision of the specification and left three detailed comments, including a correction drawn from their own project's experience. That doesn't happen in Jira.

A board with no owner

The downside shows up on the same project's marketing board. The card for an article published the day before is still "In Progress." A closed issue sits in "waiting." Eight newsletter cards are overdue, and nine haven't moved in over a month. Nobody made a mistake: the agent created everything that was asked for, and the human closed everything they remembered.

Hence the rule: a board the agent writes to needs a daily routine that the agent performs without being asked. The agent acts as a project manager, except it doesn't do the work itself; it watches the board. In the morning, it opens the board and reports what's approaching its deadline and what's overdue. Then it closes the cards that the session has already finished (this happens more often than you'd think) and asks about everything untouched for three weeks: is this still needed? The author already has a morning briefing from a scheduled headless agent. She wants to build the board check into it as a separate "janitor" that changes nothing silently. She hasn't refined this part yet.

Miro didn't help. The agent can draw there through an MCP mind map, timeline, and impact/effort matrix, and it looks prettier. But the copies reproduced stale statuses from the GitHub board, and the author didn't want to move work to another environment just for nicer views.

Two kinds of "later"

The author divides ideas into two types. Tasks ("do this") need a status, an owner, and a closing date. Issues, boards, Jira, and the built-in task list were made for them. Brainstorming reports ("we could go this way") need a map: what grew out of what, and what was set aside. ai-tree and mind maps were made for those. Most tools cover only the first type, and the author keeps losing the second.

What to take away

The material is subjective and rests on a single project. But the conclusion applies more broadly: if an agent writes to a tracker, whether GitHub, Linear, or Jira via MCP, build a daily reconciliation into the process. Without it, any board gradually turns into a junkyard. The author invites readers to share their experience: does your storage survive a session, does the board stay current without you, and does anyone have a ready-made daily cleanup task?

Story

Key Questions Answered

What are the key details on Five places for "later"?

Scrolling back through the chat. The default option. It works until the context is compressed or the terminal or chat is closed. With agents, that's a few hours at most. A notes file in the repository. This is NOTES.md, CLAUDE.md, or the agent's own memory. It's good for facts about the project: which flags the build uses, what quirks the deploy script has. It doesn't work for tasks, because a line has no status or deadline, and nobody is responsible for closing it.

What are the key details on A board with no owner?

The downside shows up on the same project's marketing board. The card for an article published the day before is still "In Progress." A closed issue sits in "waiting." Eight newsletter cards are overdue, and nine haven't moved in over a month. Nobody made a mistake: the agent created everything that was asked for, and the human closed everything they remembered. Hence the rule: a board the agent writes to needs a daily routine that the agent performs without being asked.

What are the key details on Two kinds of "later"?

The author divides ideas into two types. Tasks ("do this") need a status, an owner, and a closing date. Issues, boards, Jira, and the built-in task list were made for them. Brainstorming reports ("we could go this way") need a map: what grew out of what, and what was set aside. ai-tree and mind maps were made for those. Most tools cover only the first type, and the author keeps losing the second.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…