Planning With Files

Planning with Files

Use this pattern for any task where you can't complete everything in one pass: long refactors, migrations, multi-feature buildouts, research tasks, or anything with more than ~5 tool calls. File-based state survives context resets and gives you (and the user) a running record of what's been done.

Place planning files in docs/operational/ if the harness is linked, or in a .planning/ directory at the project root — whichever exists.

Files to create at the start

task_plan.md

Written before any implementation. Contains:

# Task: <short title>

## Goal
One sentence: what does "done" look like?

## Scope
What's in and what's explicitly out.

## Steps
1. Step one (estimate: small/medium/large)
2. Step two
3. ...

## Open questions
- Question 1 (unblocks: step 3)

## Success criteria
- [ ] Criterion A
- [ ] Criterion B

findings.md

Appended to as you work. Never edited retroactively.

# Findings

## <ISO timestamp>
**Context:** What you were investigating.
**Finding:** What you learned.
**Impact:** Does it change the plan? If yes, add a note to task_plan.md.

progress.md

Updated after each step completes:

# Progress

## Completed
- [x] Step 1 — <one-line summary> (commit: <sha>)

## In progress
- [ ] Step 2 — started <timestamp>

## Blocked
- Step 3 — waiting on: <reason>

## Next action
<Exact next thing to do when resuming>

Workflow

1. Read existing task_plan.md (if present) — don't start a second plan
   for the same task.

2. Write task_plan.md → get user approval on scope before doing anything.

3. For each step:
   a. Write findings to findings.md before and after each discovery.
   b. Do the work (code, commands, searches).
   c. Mark step complete in progress.md; record the commit SHA.
   d. If blocked, record the blocker in progress.md and surface it.

4. When context resets: read progress.md first. Resume from "Next action".

5. When all steps are complete:
   - Verify all success criteria in task_plan.md.
   - Archive or delete the planning files (they're operational docs, not
     permanent — see docs/operational/README.md for the lifecycle).

Rules

• One plan per task. If a task_plan.md already exists, read it and continue — don't overwrite a plan in progress.
• Write findings before you forget them. Append to findings.md immediately after each discovery; don't reconstruct from memory later.
• Keep progress.md truthful. Only mark a step complete when its commit is on the branch. "In progress" is not the same as "done."
• Surface blockers immediately. A step that's blocked doesn't become "in progress" — it goes to the Blocked section with the reason.
• Never truncate. Append to findings.md; don't trim old entries to save space — they're the audit trail.

When NOT to use this pattern

• Single-function edits or quick fixes (< 5 tool calls).
• Well-scoped tasks the agent can complete in one continuous context window.
• Tasks where the user is present and can provide all context interactively.