agentleFS
SdamirsaSdamirsa/dot-claude-iff/.claude/skills/plan-task/SKILL.md

plan-task

Skill2 starsChanged 42 days agoApache-2.0View on GitHub

Create and maintain resume-proof task files under .claude/tasks/. Invoke on "plan this", "let's plan", "break this down", or any task likely to span more than one session, AND on "resume", "where were we", "continue the task", "pick up <task>". NOT for trivial single-step asks that finish inside one exchange; those need no task file.

13 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use ship, handoff and 1 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • AWSSkill

    ship

    Finalize dev work on an issue branch — update docs, push, and open a PR.

    Also in 3 projects that use this

  • MicrosoftSkill

    handoff

    Compact the current conversation into a handoff document for another agent to pick up.

    Also in 4 projects that use this

  • GoogleSkill

    code-review

    Adversarial systems, concurrency, memory-model, and architecture code review protocol. Use this skill whenever asked to review code, PRs, diffs, or branches in Filament.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

12 rules. 0 are common, 6 are distinctive, and 6 are very specific to this project.

  • 0

    Common

    Most of the 11 edited versions we read keep these.

  • 6

    Distinctive

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    After every meaningful step, update the task file's ## Checkpoint block:

    line 56
  2. Distinctive:

    Update the checkpoint BEFORE long or risky operations, not only after, and in the same step set the journal pointer to match:

    line 65
  3. Distinctive:

    Use the same text you just wrote as Next action.

    line 72
  4. Very specific:

    Read .claude/STATUS.md to identify the active task file(s).

    line 103
  5. Distinctive:

    Read the task file: Goal, Plan, Checkpoint, NEEDS-HUMAN.

    line 104
  6. Very specific:

    If any ## NEEDS-HUMAN item gates the next action, resolve it with the user now (you are in the main session, asking is allowed here), then …

    line 110
  7. Distinctive:

    When every Plan box is checked and each step's done-evidence has been verified (not assumed):

    line 118
  8. Very specific:

    Update .claude/STATUS.md: remove the task from ## Active tasks, refresh ## Next steps.

    line 127
  9. Very specific:

    Every task file has a status line, a Definition of done, one backticked Test command, and a journal entry under its milestone (statectl.py …

    line 136
  10. Distinctive:

    Every Plan step names its done-evidence.

    line 138
  11. Very specific:

    Never delete a task file; archive it under .claude/tasks/archive/YYYYMMDD/.

    line 139
  12. Distinctive:

    BEFORE long or risky operations, not only after (crash-safety).

    line 141

“Common” is measured: kept by at least 6 in 10 of the 11 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

plan-task: resume-proof task planning

The conversation is not durable; the task file is. This skill turns any multi-step piece of work into a file under .claude/tasks/ that a future session (yours, another agent's, or a session after a crash) can pick up without reconstruction. If losing this conversation would lose progress, the task needs a file.

Creating a task

  1. Copy .claude/tasks/_template.md to .claude/tasks/YYYYMMDD-<slug>.md (date = today, slug = short kebab-case, e.g. 20260720-backfill-q2.md). One file per task. The file name without .md is the task's id in the journal (20260720-backfill-q2).
  2. Fill the top sections:
    • Status line: _Created <today> · Status: todo_, kept current as the work moves (todo, doing, done, blocked). The console and the phase exit checks read it.
    • Goal: 1 to 2 sentences that include the definition of done. If you can't state what must exist or pass for the task to close, you aren't ready to plan it, ask.
    • Definition of done and Test: the condition that closes the task, and on the **Test:** line exactly one backticked command that proves it, such as python3 -m pytest tests/test_x.py -q, run from the repo root without a shell. Leaving the build phase reruns that command; a Test that cannot fail proves nothing.
    • Context: why now; link the Project log entries and .claude/research/ files that motivated it, so a resumer doesn't re-derive the rationale.
    • Plan: a checklist of verifiable steps. Each step names its done-evidence: the file that will exist, the test that will pass, the output that will match. A step whose completion can't be checked from disk isn't a step, split or sharpen it.
  3. Register the task in the journal under its milestone, in the same step:

    python3 .claude/tools/statectl.py task <id> --title "<name>" --status todo --milestone <mid>

    For a new body of work, register the milestone first (python3 .claude/tools/statectl.py milestone <mid> --title "<what it delivers>") and name it on the file's Milestone: line. Leaving the plan phase checks that every open task file is registered this way (checkctl.py phase-exit --from plan).

  4. Register the task in .claude/STATUS.md under ## Active tasks (link the file).
  5. An idea that does not belong to this milestone is not a task: park it with python3 .claude/tools/statectl.py proposal add "<idea>" --source <human|agent:<name>|issue#N>.

A well-formed plan step versus a vague one:

- [ ] Reingest June exports (done when: output/weekly/2026-06-*.parquet exist and
      `pytest tests/test_ingest.py -q` passes)         # verifiable
- [ ] Handle June data                                  # not a step, no evidence

While working

After every meaningful step, update the task file's ## Checkpoint block:

  • Last completed: the step just finished, verbatim from the Plan.
  • Next action: the exact command to run, or the exact file + edit to make. "Continue the migration" is not a next action; uv run scripts/ingest.py --month 2026-06 is.
  • State files: the paths that hold progress (outputs, caches, staging files), so a resumer can verify state instead of guessing it.
  • Updated: today's date.

Update the checkpoint BEFORE long or risky operations, not only after, and in the same step set the journal pointer to match:

python3 .claude/tools/statectl.py pointer "<next concrete action>"

Use the same text you just wrote as Next action. The task file is what a resumer reads once they have found it; the pointer is what a COLD session sees first, printed by the SessionStart hook before anyone has opened a task file at all. If the operation crashes the session, both must already describe how to recover: a checkpoint update with no matching pointer call leaves the cold-start orientation pointing at stale work even though the task file itself is fine.

Questions for the human do not block the work. Open them in the needs-human queue:

python3 .claude/tools/statectl.py need open --title "<question>" \
  --category <provide-input|decide|review|unblock-env|approve-release|system-blocker> \
  --context "<the situation in plain language: what is going on, why it matters, what it blocks>" \
  --action "<the one concrete thing the human should do>" \
  --band <SEV0|SEV1|SEV2|SEV3> --note "<options + your recommendation>"

It prints an id (band defaults from category if you omit --band). Record that same id, the question, options, your recommendation, and what it blocks in the task file's ## NEEDS-HUMAN table, then finish all independent steps. Two places hold one fact on purpose: the queue is what a cold session and the console surface without anyone opening this file; the table is what a resumer sees in context once they do. The rules for when to ask versus proceed are in .claude/protocols/human-gates.md.

The task file is the single source of truth for progress. Never let the conversation hold state the file doesn't.

Resuming

On "resume", "where were we", "continue the task", or at the start of any session with an active task:

  1. Read .claude/STATUS.md to identify the active task file(s).
  2. Read the task file: Goal, Plan, Checkpoint, NEEDS-HUMAN.
  3. Verify the checkpoint against reality before trusting it. Checkpoints can be stale: a crash mid-write, an operation that half-completed, a manual change since. Check: do the State files exist? Does the last completed step's done-evidence still hold (files present, tests pass, output matches)? If reality disagrees with the checkpoint, reality wins, correct the checkpoint first, then proceed.
  4. If any ## NEEDS-HUMAN item gates the next action, resolve it with the user now (you are in the main session, asking is allowed here), then close the matching queue entry: python3 .claude/tools/statectl.py need resolve <id> --answer "<answer>". Items that don't gate anything can wait for the next natural boundary.
  5. Continue from Next action.

Closing

When every Plan box is checked and each step's done-evidence has been verified (not assumed):

  1. Fill ## Outcome: what shipped, and the Project log entry that records it.
  2. Change the status line to done, record it (statectl.py task <id> --status done), and move the file to .claude/tasks/archive/YYYYMMDD/ (create the dated folder; add a README line there: what, archived, why, replacement). The build exit check still finds an archived task's Test there.
  3. Append a deliverable or milestone entry to .claude/Project-log.jsonl if one doesn't already exist for this work.
  4. Update .claude/STATUS.md: remove the task from ## Active tasks, refresh ## Next steps.

Closing normally happens inside /project-memory at session end (the user types it; suggest it at a natural boundary), but close immediately if the task finishes mid-session: a done task lingering as active misleads the next resumer.

Rules

  • One file per task. A task that grows a second goal becomes two files.
  • Every task file has a status line, a Definition of done, one backticked Test command, and a journal entry under its milestone (statectl.py task <id> --milestone <mid>).
  • Every Plan step names its done-evidence; unverifiable steps get split or sharpened.
  • Never delete a task file; archive it under .claude/tasks/archive/YYYYMMDD/.
  • Checkpoint updates set the journal pointer in the same step (statectl.py pointer ...), BEFORE long or risky operations, not only after (crash-safety).
  • Parked questions go to both places: the needs-human queue (statectl.py need open) and the task file's ## NEEDS-HUMAN table, same id in both.
  • On resume, verify the checkpoint against reality before acting on it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.