write-plan

SKILL.md

Write Plan

Overview

Produce a complete, self-contained implementation plan that can be executed by execute-plan with minimal ambiguity, even after /clear or by another agent.

This skill is for planning only:

Do not implement code

Do not modify production files (except plan artifacts)

Workflow

Step 1: Contextualize

Load only the project context relevant to the requested change:

If docs/SUMMARY.md exists, read it first.

Load only task-relevant detail docs.

Prioritize Code Standard docs for implementation conventions.

If docs conflict with code or user intent, use the available input/question tool before broad changes.

Then inspect only the code areas relevant to the requested change.

Capture:

User-visible purpose and expected outcome

Existing patterns to follow

Constraints and dependencies

Files, modules, commands, and docs that orient the executor

Risks, assumptions, and unknowns

Step 2: Initialize Plan Artifacts

Create: docs/plans/YYMMDD-HHmm-<plan-slug>/

Create:

SUMMARY.md

one phase file per implementation phase with naming convention phase-XX-<name>.md

Add research/ only if needed.

Rules:

Use timestamp commands from the shared General Principles for folder and document timestamps.

Step 3: Clarify Requirements

Ask clarifying questions to resolve any ambiguity in the request. Focus on:

Scope and boundaries

Success criteria

Constraints and non-goals

Priorities and trade-offs

Rules:

If requirements are already clear or come from the brainstorm context, no need the confirmation step.

Use input/question tool for gathering answers, context.

State assumptions explicitly in SUMMARY.md. If multiple interpretations of the request exist, list them and ask — never pick silently.

Step 4: Define Strategy and Phases

Design a phased strategy that is safe and verifiable.

Each phase should have:

A clear objective

The complexity and risk level appropriate to the phase with values: S, M, L, XL

Ordered tasks

Verification commands

Observable acceptance criteria and exit criteria

Granularity rule:

Tasks should be small, concrete, and typically 2-10 minutes each.

Prefer phases that can be resumed safely. Document idempotency, recovery notes, or rollback constraints for risky work.

Step 5: Research (Only if Needed)

Research is optional and should be proportional to uncertainty.

Preferred order:

Existing project docs and code

Existing skills and local references

External references (only if available in the current environment)

If external research capability is unavailable, proceed with local evidence and explicitly list assumptions and open questions.

Document findings in:

docs/plans/YYMMDD-HHmm-<plan-slug>/research/<topic>.md

Step 6: Write Plan Content

SUMMARY.md format

Follow the template inside references/summary-template.md

The summary must be a living plan, not a static proposal. Include empty sections for execution-time updates: progress, surprises/discoveries, decision log, and outcomes/retrospective. These sections give execute-plan a stable place to record what changed and why.

phase-XX-<name>.md format

Follow the template inside references/phase-template.md

Step 7: Review and Refine

Before presenting the plan, verify:

Paths are exact and consistent

Phase order is logical

Tasks are actionable (no vague steps)

Verification is defined for each phase

Acceptance criteria are observable

Risks/assumptions are explicit

Plan is executable without hidden context from the current chat

Then present for user review.

If multiple viable approaches exist, present options and ask for one of: (use input/question tool for selection)

Confirm: approve current plan for execution

Confirm and Visualize: approve current plan and create a source-adjacent visualization in the same session

Validate: refine via additional clarifying questions

If the user chooses Confirm and Visualize:

Use the current write-plan session context and the plan artifacts just created.

Do not restart project context loading or rediscover background that is already available in the session.

Follow the visualize skill output convention for plan folders:

docs/plans/YYMMDD-HHmm-<plan-slug>/visualize.html

docs/plans/YYMMDD-HHmm-<plan-slug>/visualize-assets/

Copy the fixed visualization theme into the adjacent assets folder.

Verify the visualization enough to confirm the HTML, local CSS link, Mermaid import, source metadata, and primary content blocks are present.

Then continue to the normal handoff.

Step 8: Handoff

End with:

Plan <relative_path_to_plan>/SUMMARY.md is ready.

Make new session and use execute-plan <relative_path_to_plan>/SUMMARY.md to execute it.

If visualization was created, also include:

Visualization <relative_path_to_plan>/visualize.html is ready.

Rules

Never automatically implement or execute the code change in the same session. Optional plan visualization is allowed only after user selection and only for the plan artifacts.

Prefer explicit file paths and concrete commands

Align with project standards and existing architecture

Keep plans self-contained, deterministic, and resumable. A fresh agent should be able to continue from the plan folder alone.

Plan the minimum viable change: No speculative phases, no "just in case" abstractions, no flexibility that wasn't requested. If a plan can be 3 phases instead of 6, make it 3. Every task should trace directly to a stated requirement.

If the write-plan request comes from a brainstorm session, we can skip many steps like gathering documents, clarifying requirements, and researching, because those should have been covered in the brainstorm session. In that case, we can directly start from Step 4: Define Strategy and Phases, using the information from the brainstorm session as context.

1d