| name | to-plan |
| description | Turn the current conversation context into a concrete implementation plan for immediate execution. Use when the user wants to plan quick implementation after a grilling session, or when saying "let's plan this" or "make a plan" for a change that doesn't need a full PRD. |
To Plan
Turn conversation context into a concrete, ordered implementation plan — persisted as a file that tracks progress so you can resume across sessions.
When to use
- You've just finished a
grill-with-docs session and want to move straight to building
- The change is small-to-medium and doesn't warrant a full PRD → issues → TDD pipeline
- The user says "let's plan this" or "make a plan" and intends to implement right away
When NOT to use
- The change is large enough to need formal handoff → use
to-prd instead
- The user wants to break work into separately-grabbable tickets → use
to-issues instead
- The conversation has no design context yet →
grill-with-docs first, then to-plan
Process
1. Gather what you already know
Work from whatever is already in the conversation context — decisions from grilling, terminology from CONTEXT.md, ADRs, codebase exploration you've already done. Do not interview the user.
If you haven't explored the codebase yet, do so now. Use read, grep, find, or read-only bash commands — or dispatch a subagent for broader exploration. Read the project's domain glossary (look for CONTEXT.md or context.md — case-insensitive) and respect any ADRs in the area you're touching.
2. Identify scope boundaries
Decide — and propose — what is in scope and out of scope for this plan. Be explicit about what you're excluding. The user can adjust.
3. Draft the plan
Write the plan as numbered slices, ordered by dependency. Each slice is a vertical cut through all relevant layers — not a horizontal task list. Use the project's domain glossary vocabulary from CONTEXT.md throughout.
Each slice should specify:
- What: one behaviour or capability this slice delivers
- Test: what test confirms it works (behaviour-level, not implementation)
- Touches: which modules/files this slice modifies (approximate, not exhaustive)
- Depends on: which earlier slices (if any)
The first slice should be a tracer bullet — the thinnest end-to-end path that proves the approach works.
End the plan with:
- In scope: what this plan covers
- Out of scope: what it deliberately doesn't
- Open questions: anything you're unsure about that the user should decide
4. Review with the user
Present the plan. Ask:
- Does the scope feel right? (Anything missing or over-scoped?)
- Is the slice order correct? (Any dependency issues?)
- Are the open questions answerable now, or should they be resolved during implementation?
Iterate until the user approves the plan.
5. Write the plan file
Once approved, write the plan to plan-<slug>.md at the project root. Use a short kebab-case slug derived from the goal — e.g. plan-add-bcrypt-hashing.md. If the plan is the main or only plan in progress, plan.md is fine too.
If a plan file with the same slug already exists, read it first — it may contain progress from a previous session. Merge any new decisions rather than overwriting. See Plan format for the structure including progress tracking.
6. Execute
Start implementing the first incomplete slice. If the user wants TDD, switch to the tdd skill now. Otherwise, implement slices in order, running tests after each one.
After completing each slice, update the plan file — mark the slice status as done and add a brief note of what was implemented. This ensures the plan is a reliable resume point if the session closes.
7. Resume
If a plan-<slug>.md already exists when the user says "continue the plan" or "pick up where we left off":
- Read the plan file to see which slices are done, in-progress, or pending.
- Summarise progress to the user: what's done, what's next.
- Continue from the first incomplete slice.
If the context from the original session is lost, the plan file plus CONTEXT.md and any ADRs should give enough context to resume.
Plan format
Each slice has a status field that tracks progress. Update the file after completing each slice.
# Plan: <short title>
## Goal
[One sentence describing what we're building and why]
## Slices
1. **[Slice name]** — ✅ done
- What: [behaviour this delivers]
- Test: [how to confirm it works]
- Touches: [approximate module/file areas]
- Depends on: none / #N
- Notes: [brief note on what was implemented, added after completion]
2. **[Slice name]** — 🔄 in progress
- ...
3. **[Slice name]** — ⏳ pending
- ...
## Scope
- In scope: [list]
- Out of scope: [list]
## Open questions
- [list, or "None — ready to go"]
Status markers:
- ⏳ pending — not started
- 🔄 in progress — currently being worked on
- ✅ done — completed, with a notes line summarising what was done
Rules
- Synthesize from existing context. Do not interview the user — use what you already know.
- Use vertical slices ordered by dependency — each slice is a thin, complete path through all layers.
- The first slice is always a tracer bullet (thinnest end-to-end proof of concept).
- Use the project's domain glossary vocabulary from
CONTEXT.md throughout. Look for both CONTEXT.md and context.md (case-insensitive) when searching.
- Respect any ADRs in the area you're touching.
- Persist the plan as
plan-<slug>.md at the project root. Update it after completing each slice.
- If resuming, read the existing plan file first — it contains progress from the previous session.
- Do not include specific file paths or code snippets in the plan (they go stale fast). Exception: a notes line on a completed slice may mention key files that were created/modified.
- When exploring the codebase, use
read, grep, find, or read-only bash — or dispatch a subagent for broader exploration. Use read over shelling out to cat for file contents.
