Plan Init

Create Project Plan (v2)

Overview

Produce a structured plan document for the given task, marked as v2 so the /plan-phase and /plan-run skills recognize and operate on it. The document becomes the anchor for all future work breakdown and execution. Once written, it is never modified by subsequent skills — it is a stable reference.

This is the v2 variant of /plan-init-v1. Three differences change what the skill does; the rest of the divergence between the two files is wording:

1. The plan's Status table is replaced, not extended: Format: v2 and Suite are the only two rows, and v1's Phase / State / Blocker / Last updated are gone (Step 6 says why).
2. The plan is registered in the plans/README.md discovery index — when it lives under plans/. The index row is a relative link resolved from plans/README.md, so a plan the user asked for somewhere else has no row that would resolve (Step 7).
3. When UI is in scope, a visual-verification success criterion is added.

Everything else about the plan's content model is identical to v1. In particular, this skill stays breakdown-unaware: it writes no phases and no grouping — that is /plan-phase's job.

Step 1 — Understand the initial task

The user's argument (everything after /plan-init) is the task description. If no argument was given, ask: "What task do you want to plan?" and wait for the answer. If operating autonomously (no user available) and no argument was given, stop and report: "No task description provided — cannot proceed autonomously without input."

Parse the task description to extract what the user has already told you:

• The core objective
• Any explicit constraints
• Any explicit success criteria
• Any explicit non-goals
• Whether the work involves a web UI (a frontend, pages, components, or user-visible visual behavior) — this decides the UI success criterion in Step 4.

Step 2 — Interview the user

Before doing any codebase exploration, ask the user clarifying questions in a single message. Only ask about things genuinely unclear or missing from the task description — skip any question the user already answered.

Cover these areas (omit any that are already clear):

• Scope — Is there anything specific this should or should not touch? Any related systems or concerns that aren't obvious from the description?

• Done criteria — How will you know it's working? Any specific tests, benchmarks, or manual checks that must pass? Anything currently broken that this must fix?

• UI in scope — Does this change anything a user sees or interacts with? If so, which flows/routes matter most for verification?

• Hard constraints — Any technical decisions already made that can't change? (specific library, API shape, DB schema, backwards compatibility requirement)

• Non-goals — What should this plan explicitly not address, even if related?

• Output location — Propose a location: "I'll write the plan to plans/<slug>/plan.md in the current directory — does that work, or would you prefer a different path?"

Wait for the user's answers before proceeding. Do not explore the codebase yet.

If operating autonomously (no user available), proceed with reasonable assumptions based on the task description and note each assumption explicitly in the plan.

Step 3 — Explore the codebase

First work through the framework/library standards check in the subsection below — it takes precedence over any pattern you would infer from the code, so it has to precede exploration rather than follow it. Then explore the parts of the codebase relevant to the task. Identify:

• Files, modules, and components that will need to change
• Files that won't change but are important context (callers, dependencies, tests)
• Existing patterns and conventions to preserve
• Any existing tests that currently cover the affected area
• Whether a web UI / Playwright setup exists (informs the UI success criterion)

Search the codebase for relevant files and read them. Do not skip this step — the constraints and affected-areas sections of the plan depend on what you find here.

Always check for framework/library standards

Before exploring the main codebase, check for standards documents in any referenced framework or library:

1. Read AGENTS.md and CLAUDE.md in the current repo (if present).
2. Check package.json, pyproject.toml, or other project config for references to sibling repos or framework dependencies. For each sibling repo found:
   • Read its AGENTS.md — it may contain an explicit "Building on X" section
   • Read any technical-standards or conventions document it publishes
3. If the task involves building on a sibling framework, check for a canonical demo or example directory before designing any architecture.

These documents take precedence over any patterns you infer from reading code alone.

Check for project quality gates

If the project has validation or quality gate scripts, verify they are properly configured for the affected source tree:

1. Check whether scripts/validate.sh or an equivalent quality gate exists.
2. Check whether pre-commit hooks are configured (e.g., .husky/pre-commit, lint-staged in package.json, .pre-commit-config.yaml).
3. If quality gates exist, verify they cover the directories affected by this task. If not, note the gap as a constraint or prerequisite in the plan.

Step 4 — Derive the plan content

From the task description, user interview answers, and codebase exploration, determine:

Goal — One to three sentences. What is being done and why? What is the desired end state? Keep it concrete.

Success criteria — A checklist of observable, testable conditions that prove the work is done. Each item should be falsifiable: a specific test command, a specific behavior to verify, or a specific check to run. Examples:

• All tests pass with no regressions (<test command>)
• Feature X behaves like Y when Z happens
• No build errors (<build command>)
• Manual check:

If the work involves a web UI, add a visual-verification success criterion, branching on whether a Playwright setup exists (the Step 3 finding):

• With Playwright: UI verified: the target flow's spec is green N× (default 3) across the mandatory viewport matrix, with frames/screenshots inspected (not merely produced) — checked with /web-verify at the phase gate.
• Without Playwright: UI verified by manual inspection: /web-verify's bundled manual UI-verification checklist is completed across the mandatory viewports (its degraded mode) — no spec is required.

Technical constraints — Things the implementation must respect. Include:

• Architectural rules from project agent instructions (e.g., AGENTS.md/CLAUDE.md) (if relevant)
• Patterns already in use that should be followed
• Dependencies or APIs that cannot change
• Performance or compatibility requirements
• Anything that would block a PR if violated

Non-goals — What this plan explicitly does NOT address. Include anything the user stated plus anything implied by the scope. At least one entry required.

Affected areas — A list of real file paths that will change or must stay consistent. Use actual paths from the codebase. Group by:

• "Will change" — files that get edited or created
• "Must stay consistent" — callers/consumers that must still work
• "Tests" — test files that need new or changed tests. For new behaviour, prefer TDD: the failing test is written before the implementation that makes it pass.

Step 5 — Derive the plan slug and output path

Use the output location the user confirmed in Step 2. If they accepted the default, generate a short kebab-case slug from the task description (3–5 words max) and use plans/<slug>/plan.md. Examples: "auth-refactor", "vocab-export-endpoint". No date prefix.

Step 6 — Write the plan document

Before writing, check whether the path you are about to write already exists — the one resolved just above, which is plans/<slug>/plan.md only when the user accepted the default. Checking that literal instead would look for a file the skill is not writing, find nothing, and overwrite whatever the user actually named; on that branch there is no <slug> to substitute either, because it is generated only on the default path. If it does, do not overwrite it — pick a more specific slug (noting the change) or stop and ask the user whether to resume or replace the existing plan.

Create the directory and write the plan file using the plan-document skeleton in references/plan-template.md. The Status table is exactly two rows — Format and Suite. Write no phases and no grouping.

## Assumptions is the section that autonomous mode's "note each assumption explicitly in the plan" writes into. Include it whenever an interview question went unanswered — the reader needs to know which parts of the plan rest on a guess, and which would change if the guess is wrong. Omit the section entirely when every question was answered; an empty heading says less than no heading.

plan.md is written once and never edited again: /plan-phase and /plan-run both read it and leave it alone. So it carries no State, Phase, Blocker or Last updated cell (v1 plans do) — a status field in an immutable document can only ever be stale. Live status lives in execution.md and the phase documents, as checkboxes.

The | Format | v2 | row is the load-bearing marker: /plan-phase and /plan-run act only on plans that carry it, and the v1 skills never do. Emit it exactly as the two-cell table row shown in the template.

Step 7 — Register the plan in the discovery index

Only when the plan lives under plans/ — that is, the user accepted the default in Step 2 and Step 5 generated a <slug>. Off that path there is no slug to link and no index to belong to: skip this step, and never create plans/ to hold a row. The row snippet resolves relative to plans/README.md, so a plan written to docs/proposal.md would be indexed as plans/<slug>/ — a link to nothing, in a directory invented to hold it. /plan-phase and /plan-run both refuse to invent that directory; this must too.

On the default path, maintain plans/README.md as the discovery index (it renders on GitHub/GitLab), using the create/append snippets in references/plan-template.md:

• If plans/README.md does not exist, create it with the index title and table header from the template.
• Append one row for this plan using the row snippet from the template: the linked slug and one sentence. Do not rewrite existing rows for other plans.

The index carries no status column — a plan's own checkboxes are the progress record, and a column here would be a second copy that goes stale.

Step 8 — Report to the user

Print a brief summary:

• The plan file path
• The goal (one sentence)
• How many success criteria were identified (note if a UI-verification criterion was added)
• How many files are in "will change"
• That the plan is stamped Format: v2, and that it is indexed in plans/README.md — or, off the default path, that it was not indexed and why
• Next step: "Run /plan-phase <path> to break this into executable phases and generate the execution.md tracker."

References

• references/plan-template.md — the plan-document skeleton (Step 6) and the discovery-index create/append snippets (Step 7).