Plan First

Plan First

Write the plan before doing the work. The todo tool is your external memory - anything you record there stays visible on every future turn, even after context compaction, and lets the user see exactly where you are in a multi-step task.

When to plan

Trigger planning when ANY of these are true:

• The request has 2+ distinct goals ("add X and update Y").
• The request has change + verify steps ("fix the bug and run the tests").
• You'll need to touch 3+ files.
• You're starting exploratory work where you don't yet know all the steps.

Skip planning for trivial single-step asks: "what does this file export?", "show me line 40", "rename this variable". No plan needed when the whole request is one tool call.

Workflow

1. Draft the plan. Call todo with action: "add" and an items array containing one short, verifiable step per entry. Each step should be something you can clearly tell is "done" or "not done". Prefer 3–8 items; fewer if the task is small, more only if it's genuinely complex.

{
  "action": "add",
  "items": [
    "Read foo.ts and bar.ts to locate the bug",
    "Write a failing test that reproduces it",
    "Apply the fix",
    "Run the test suite",
    "Update CHANGELOG.md"
  ]
}

2. Start one item. Call todo with action: "start" and the id of the first item:

{ "action": "start", "id": 1 }

Only one todo may be in_progress at a time - this is enforced by the tool. Trying to start a second one while another is active will return an error.

3. Do the work. Use other tools (read / bash / edit / write / grep / etc.) to execute the current item. Stay focused on the one thing that's in_progress.

4. Park in review for verification. When the change is written but you haven't yet confirmed it works, move the item to the review column:

{ "action": "review", "id": 1, "note": "fix applied - need to run tests" }

review can only be entered from in_progress, and at most one item may be in review at a time (separate limit from in_progress - you can have one in_progress and one review simultaneously, e.g. tests running on item A while you start coding item B). The note is optional but helpful as a reminder of what still needs verifying.

5. Verify. Actually check the outcome matches what the step promised:

• "Run tests" → the last bash call must have shown the tests passing.
• "Apply fix" → the edit or write succeeded and the change is visible in the file.
• "Read foo.ts" → you've actually read it, not just searched for it.

If verification reveals more work is needed, move the item back to in_progress with action: "start", fix it, and return to review.

6. Complete. Once verified:

{ "action": "complete", "id": 1 }

From review, the note is optional - the review step was the verification parking, so the tool trusts that verification happened. You can add a final note ("note": "all 47 tests pass") if it's worth recording.

If verification is immediate (e.g., you ran the tests as part of doing the work and don't need a separate parking step), you can skip review and go straight from in_progress to complete - but the tool will then require a note describing what verified the outcome:

{ "action": "complete", "id": 1, "note": "bash call in previous turn showed 47/47 passing" }

This is the forcing function: either the plan shows you parked for verification (review step), or you spell out what evidence makes you confident it's done.

7. Next item. Call start on the next pending todo. Repeat.

8. Handle obstacles. If you can't make progress - missing info, broken env, failing test you can't diagnose, unclear spec - don't silently move on. Call block with a note explaining the blocker:

{ "action": "block", "id": 3, "note": "The test harness requires FOO_TOKEN which isn't in the env" }

Then surface the blocker to the user. They can unblock you or repivot.

9. Adapt as you learn. The initial plan won't always be right:

• New work surfaces mid-task: add more items.
• A step turns out to be unnecessary: complete it with a note like "not needed, handler already supports this".
• The whole direction needs to change: clear and start over.

Anti-patterns

• Don't mark items complete you haven't verified. "I think that's done" is not complete. Either park it in review, verify, then complete; or complete directly with a note spelling out the evidence.
• Don't skip planning because "this will only take a second." If you were wrong, the plan costs you nothing and saves a derailed session.
• Don't start a second item while one is in_progress. Finish (or review, or block) the first. The tool will reject the second start.
• Don't stack items in review. One review slot; verify and complete (or start again to revise) before parking another.
• Don't claim you're "done" while in_progress, review, or pending todos remain. Either finish them or mark them blocked and explain why. The harness will catch and re-prompt you if you try - save the round trip by self-checking first.
• Don't over-decompose. "Open file", "read line 10", "close file" is too fine-grained. One verifiable outcome per todo.
• Don't forget to add items that emerge mid-task. A plan that doesn't reflect reality is worse than no plan.

Full example

User: "Please add a rate-limiter to the /api/search endpoint and write tests for it."

Good opening move:

{
  "action": "add",
  "items": [
    "Read src/api/search.ts to find the handler entry point",
    "Design the rate-limiter shape (sliding window, per-IP)",
    "Implement src/lib/rate-limit.ts",
    "Wire it into the search handler",
    "Add unit tests for rate-limit.ts",
    "Run the test suite and verify green"
  ]
}

Then { "action": "start", "id": 1 } and begin.

After reading the file:

{ "action": "complete", "id": 1, "note": "entry point at searchHandler() in search.ts:42" }
{ "action": "start", "id": 2 }

For later items that need verification (like running tests), park first:

{ "action": "review", "id": 6, "note": "invoked npm test; waiting on output" }

...then after seeing green output:

{ "action": "complete", "id": 6 }

...and so on until the final complete, at which point you tell the user the task is done.

Quick reference