Claude Code Kanban

Tasks belong where the work happens

My to-do app kept filling up with long briefs that nothing ever closed. The agent doing most of my work couldn't see it, and I spent my energy re-explaining context. So I moved the tasks next to the agent.

• One file is the truth. The board is a JSON file on disk. Claude can read it, update it and check it against the real state of my projects.
• Claude does the legwork. It researches each card (repos, mail, Slack, call notes, the live site), rewrites the card to what's actually left, and drafts the follow-up message.
• I keep the decisions. Nothing leaves my machine unless I press Approve on that exact message.
• Small cards, few priorities. Titles of two to five words, one checkable "done when" criterion, three priorities a day at most.

Four small pieces around one file

You talk to Claude in the chat ("add: call the accountant Thursday"), press buttons in the panel, and the routines run on their own. All three go through the same script, so the board never gets corrupted by two writers.

A card

{
  "id": "k31",
  "title": "Port badges to the template",
  "desc": "Live on the first site. Reproduce it in the shared template; no question for the partner.",
  "status": "todo",            // inbox · todo · doing · done · later
  "day": "2026-10-06",         // null = no day yet
  "bonus": false,              // true = "if I have time", never a priority
  "project": "~/DEV/directory-app",
  "done_when": "A new site built from the template shows its badges, tested locally",
  "needs_human": false,        // true = only I can confirm it's done
  "proposal": null,            // { to, text, status: pending | approved | sent }
  "researched": "2026-10-05T12:52:20-03:00"
}

What each status means

The approval loop

1. A routine finds that a card is blocked on someone, writes the message and saves it on the card as a proposal.
2. The card shows To approve with the exact text and two buttons: Approve and Reject.
3. The next run sends only what I approved, then rewrites the card into the follow-up ("Follow up with X on Y", two working days later).
4. Several short messages to the same person get merged into one message with a bullet per topic. A long message on a single topic stays on its own.

The daily rhythm

The times avoid Claude's weekday peak hours (5 to 11 am Pacific), when usage limits drain faster. Convert that window to your own time zone before picking yours.

Start with your data, not the kanban

The board is the easy part. What makes it work is that Claude can already see my world. Before I built any of this, I connected and searched each source from Claude Code until it could answer questions about my week on its own:

• My notes. A folder of Markdown files (my "Life OS") with projects, people and decisions.
• Email. Who's waiting on me, what I promised.
• Call notes. Fathom summaries and transcripts, where most commitments are made.
• Calendar. What's coming and what needs prep.
• Slack. Where my partner and I work day to day.
• ClickUp. The team's tasks, read only.
• Project folders and live sites. To check whether something is actually done.

Without these, the routines can only shuffle cards. With them, Claude can tell you that a task is already done, what's really left, and who you're waiting on.

Then sort the work into three lanes

The kanban is where the three lanes meet: decisions show up as cards to approve, the agent's work shows up as cards that move on their own, and repeating cards tell you which agent to build next.

Help me map my data before building a task system.
1. List the tools where my work and commitments live (notes, email, calls,
   calendar, chat, team tools, code). For each, check whether you can
   already read it from Claude Code; if not, tell me how to connect it.
2. Search each connected source for my open commitments this week and
   show me what you found, so I can see what you're missing.
3. Sort what you found into three lanes: decisions only I can make,
   actions you can take on your own, and recurring missions that deserve
   their own agent. Ask me where you're unsure.

Build your own

Once your data is connected, don't copy my files. Paste these prompts into Claude Code one at a time, answer its questions, and let it build a version shaped around your tools and habits. Each step works on its own, so you can stop after step 1 and still have something useful.

1. The board and its script

   Pick a folder you already use for notes. The script is the single writer of the board.

   Create a personal kanban for me as a JSON file plus a small Python CLI
   (standard library only) that is the only thing allowed to write it.
   Statuses: inbox, todo, doing, done, later. Each card has: id, title (2-5 words),
   desc, status, day (ISO date or null), project (folder path), done_when,
   needs_human, bonus (an "if I have time" flag), comments, proposal
   ({to, text, status: pending|approved|sent}), researched (timestamp).
   Commands: add, move, plan <id> <day>, edit (incl. --clear-comments,
   --researched, --bonus), archive (never delete), propose new|approve|reject|sent,
   list, show, week. Use a file lock so two writers can't corrupt it.
   Ask me where the file should live and which statuses I actually want first.

   Copy prompt

2. The panel inside Claude Code

   Mods are Claude Code plugins that can draw a pane, a band above the prompt or a status line. Claude has a built-in skill for writing them, and hot-reloads them while you iterate.

   Use the plugin-authoring skill to build a Claude Code mod called "kanban"
   that reads my board file every few seconds and shows it in a pane:
   - Day view: Todo, Doing, Done today, plus an "If time" section underneath.
   - Week view: the seven days with their cards, then cards with no day.
   - Clicking a card opens it in place with buttons that call my board CLI
     (Doing, Done, a button per remaining day of the week, Later, If time, Close = archive).
   - A pending proposal shows its text with Approve and Reject buttons.
   - A /kanban command opens the pane.
   Square borders, bold uppercase labels, small status icons like Linear.
   Then package it in a local marketplace so it stays installed across sessions.

   Copy prompt

3. The routines

   Scheduled tasks run while the Claude app is open. A run that was due while the app was closed starts at the next launch.

   Create a scheduled task that runs on weekdays outside peak hours and:
   1. Sends each proposal I approved on the board, exactly as approved, then
      marks it sent and rewrites the card as a follow-up two working days later.
   2. Researches up to 8 open cards not researched in 3 days, read-only
      (my notes, the project repo, mail, Slack, call notes, the live site),
      then rewrites each card: short title, 1-3 sentence description,
      checkable "done when", comments cleared. No phase numbers or file names.
   3. Drafts a message when a card is blocked on someone and saves it as a
      pending proposal. Merge short messages to the same person into one.
   4. Never sends anything I didn't approve, never deletes, publishes or deploys.
   End with a five-line summary of what changed and what waits for me.

   Copy prompt

4. Permissions

   In auto mode, Claude Code may refuse to create a scheduled task that sends messages on its own. If you agree with that trade-off, allow that one tool in ~/.claude/settings.json:

   "permissions": {
     "allow": ["mcp__scheduled-tasks__create_scheduled_task"]
   }

   Then open the task in the sidebar and press Run now once, so it asks for the connectors it needs (Slack, mail, call notes) while you're there.

Adapt it before you trust it

This setup fits how I work: I dictate everything, I work with one main partner on Slack, and I only want to make decisions. Yours will be different. Settle these with Claude before you let anything run on a schedule:

• Where your tasks really come from. Calls, email, Slack, WhatsApp, a team tool. Only connect the sources you read anyway.
• What Claude may send, and to whom. Start with nothing sent automatically. Add one channel at a time, always behind an approval.
• Which tool keeps team tasks. Keep your personal board personal. If your team lives in ClickUp, Linear or Monday, let Claude read those to propose cards, not to edit them.
• How you talk. If you dictate, tell Claude how your speech-to-text spells names, and to match them by sound to people it already knows.
• The language of each output. Mine: cards in French, messages to my English-speaking partner in English, ready to send as is.
• How many priorities a day. Three works for me. The "If time" lane catches the rest without guilt.

Treat the routines like a new assistant. Read what they did for the first week, correct them in plain words, and only widen what they're allowed to do once they stop surprising you.

The first version of my assistant

This is version one of my personal assistant, and it's deliberately simple: one file, one script, one panel, a few routines. It already covers what I need most. I make the decisions, Claude does the legwork, and nothing goes out without my approval.

Later versions will grow from the same board. Next comes a mission control for the third lane, with dedicated agents running the recurring missions, plus a way to capture, approve and get briefed from my phone. I'll probably build that on ClaudeClaw. The board stays the source of truth throughout, so nothing has to be rebuilt.