# Phase Plan > Create a phased implementation plan for a task, saved to docs-claude/plans/. Use when the user runs /phase-plan or asks to plan a task in phases. Runs a mandatory interactive sanity-questions loop before writing the plan, then registers per-phase execution commands. - Skill: `shaybush/phase-plan` (Agent Skill) - Install (CLI): `npx skillmds@latest add shaybush/phase-plan` - Raw SKILL.md: https://api.skillmd.com/api/skills/shaybush/phase-plan/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: Shaybush (https://skillmd.com/u/shaybush) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/shaybush/phase-plan --- **Task:** $ARGUMENTS --- Create a phased implementation plan in `docs-claude/plans/plan.md` for the task above. ## Plan Structure Requirements ### 1. Phases with Todo Checkboxes Break down into logical phases, each with checkboxes so devs can mark progress: ```markdown ### Phase X: [Phase Name] **Assigned to**: [subagent name(s)] **Date Started**: **Status**: [ ] Not Started | [ ] In Progress | [ ] Completed - [ ] Task 1 - [ ] Task 2 - [ ] Task 3 ``` ### 2. Implementation Questionnaire (at end of each phase) Devs fill this in so project managers can track what was done: ```markdown #### Phase X Completion Report | Question | Response | | ---------------------------------------- | -------- | | What was implemented? | | | Were there any deviations from the plan? | | | Issues/blockers encountered? | | | How were issues resolved? | | | Any technical debt introduced? | | | Recommendations for next phase? | | **Completed by**: [subagent name] **Date Completed**: ``` ### 3. Optional Info for Other Phases Include relevant info that other phases may need: ```markdown #### Notes for Future Phases - **Config changes**: [env vars, settings] - **New dependencies**: [packages, services] - **API changes**: [endpoints added/modified] - **Database changes**: [migrations, schema] - **Documentation updates needed**: ``` ## Important: - ask sanity questions to deep understanding what the job are. ## Sanity Questions Loop (MANDATORY before writing the plan) The sanity questions are an **interactive loop**, not a one-shot. Do NOT write the plan until the loop ends. 1. Ask a numbered list of sanity questions (1, 2, 3, ...). 2. Wait for the user's answers. 3. Parse each numbered answer. Each answer may be one of: - A direct choice/answer (e.g. `1 - B`) - A choice with a modification (e.g. `2 - C but add alignment to right support`) - A counter-question from the user (e.g. `3 - what does it mean?`) - A mix of the above in the same response 4. For each counter-question from the user: - Answer it clearly and concisely. - If your answer surfaces new ambiguity, add new sanity questions to the next round. 5. Restate the **remaining open questions** (the ones the user asked back about, plus any new ones you generated). Re-number them starting from 1 for the new round to keep things clean. 6. Repeat steps 2–5 until the user has answered everything with no counter-questions and you have no new questions. 7. Only after the loop is fully resolved, summarize all decisions and write the plan to `docs-claude/plans/.plan.md`, then register the execution commands (see **After Writing the Plan: Register Execution Commands** below). ### Example loop turn User answers: ``` 1 - B 2 - C but add alignment to right support 3 - what does it mean? ``` You should: - Record `1 = B` and `2 = C + right-alignment support` as resolved. - Answer question 3 (explain what it means). - Ask a new round only with the still-open items, e.g.: ``` Still need to confirm: 1. (was 3) Given the explanation above, which option do you prefer — A or B? 2. (new) Should X also apply to Y? ``` Never proceed to plan writing while any question is unresolved. ## After Writing the Plan: Register Execution Commands After the plan `docs-claude/plans/.plan.md` is written, register one execution command per phase in `docs-claude/plans/commands-execution.md` so another agent can run them in order. - **File**: `docs-claude/plans/commands-execution.md` - **If it does not exist**: create it with this header as the first line: ``` TODO command | STATUS ``` - **If it already exists**: only append the new lines at the end. Never overwrite, reorder, or change existing entries. - **One line per phase, in phase order.** `` is the plan filename stem (without `.plan.md`) — the same value passed as the optional plan argument to `/phase-execute`. Line format: ``` claude --dangerously-skip-permissions --effort max "/phase-execute " | NOT STARTED ``` - Sequential phases → one `/phase-execute ` line each. - Phases that are independent and may run together → group into a single `/phase-execute-parallel ` line. - New entries always start with status `NOT STARTED`. The executing agent updates STATUS (`NOT STARTED` → `IN-PROGRESS` → `DONE`). ### Example (plan `i18n-multilang`, 4 phases where 3 & 4 run in parallel) ``` TODO command | STATUS claude --dangerously-skip-permissions --effort max "/phase-execute 1 i18n-multilang" | NOT STARTED claude --dangerously-skip-permissions --effort max "/phase-execute 2 i18n-multilang" | NOT STARTED claude --dangerously-skip-permissions --effort max "/phase-execute-parallel 3,4 i18n-multilang" | NOT STARTED ``` ## Usage ``` /phase-plan create refresh token logic in the backend side /phase-plan add user authentication with JWT /phase-plan implement payment gateway integration ```