Anti-Drift Planning

{
  "name": "anti-drift-planning",
  "description": "Drift-resistant multi-session planning methodology — 5 locks (binary specs, anti-silent-arbitrage, scope ledger, verification phase, mechanical lint) for shipping multi-phase projects without losing scope between sessions",
  "version": "0.4.0",
  "author": {
    "name": "bendevcat",

---
description: Bootstrap the anti-drift methodology + folder structure in the current project. Creates <base>/{specs,plans,handoffs}/ (default docs/anti-drift/) and writes the methodology spec.
allowed-tools: ["Read", "Write", "Bash", "Glob", "AskUserQuestion", "Skill"]
---

# `/anti-drift-planning:init` — bootstrap a project

Bootstraps this project to use the anti-drift planning methodology.

## Steps

1. **Load the `anti-drift-planning` skill** first to absorb the methodology context.
2. **Check current state** of the project — git status, existing `docs/` structure, CLAUDE.md if any. If a methodology spec already exists anywhere under `docs/*/specs/*-methodology.md`, abort with a clear message instead of overwriting (this command is idempotent).
3. **Ask the user** (via `AskUserQuestion`):
   - What is the project's high-level vision or current gap to close?
   - Where should specs/plans/handoffs live? Default: `docs/anti-drift/`

---
description: Check the mechanical invariants of the plan artifacts — spec/ledger coverage, deviation references, status vocabularies, required files, unfilled placeholders. Reports only; exits non-zero on any violation.
allowed-tools: ["Bash", "Read", "Glob"]
---

# `/anti-drift-planning:lint [N]` — check the mechanical invariants

Methodology lock 5: what is countable is counted by code, not by prose review. This
command runs that code. It never asks the user anything and never edits a file.

## Argument

- `[N]` — optional plan number. Omitted, every plan in the roster is checked.

## Steps

---
description: Write a new plan spec from the anti-drift template. Args - N (plan number) and topic (kebab-case). Prompts the user for goal, success criteria, and scope.
allowed-tools: ["Read", "Write", "Bash", "Glob", "AskUserQuestion", "Skill"]
---

# `/anti-drift-planning:new-plan <N> <topic>` — author a new plan spec

Generate a new plan spec following the anti-drift binary-criteria format.

## Arguments

- `<N>` — plan number (e.g., `5`)
- `<topic>` — kebab-case short label (e.g., `agents-and-chat`)

## Steps

---
description: Regenerate a resume prompt for a plan interrupted mid-execution (handoff). Reads the ledger to find where to continue. Outputs a ready-to-paste prompt.
allowed-tools: ["Read", "Bash", "Glob"]
---

# `/anti-drift-planning:resume <N>` — resume an interrupted plan

Produce the prompt to paste as the first message of a fresh session that CONTINUES plan N from where the last session stopped (handoff: low context, or an `expensive` deviation that blocked progress).

## Argument

- `<N>` — plan number to resume

## Steps

1. **Resolve the base dir:** glob `docs/*/specs/*-methodology.md`. If no match, instruct the user to run `/anti-drift-planning:init` first and stop.

---
description: Generate the bootstrap prompt for a fresh execution session of plan N. Outputs a ready-to-paste prompt with paths and substitutions filled in.
allowed-tools: ["Read", "Bash", "Glob"]
---

# `/anti-drift-planning:start-session <N>` — prepare a fresh execution session

Produce the bootstrap prompt to paste as the first message of a new Claude Code session for plan N.

## Argument

- `<N>` — plan number to execute

## Steps

1. **Resolve the base dir:** glob `docs/*/specs/*-methodology.md`. If no match, instruct the user to run `/anti-drift-planning:init` first and stop.

---
description: Cross-plan dashboard — per-plan ledger status, deviations, and milestone tags. Announces project completion when every plan has shipped clean.
allowed-tools: ["Read", "Bash", "Glob"]
---

# `/anti-drift-planning:status` — cross-plan dashboard & project-end audit

Show the status of every plan in the project, and announce the project complete when all plans have shipped through Phase Z with zero pending decisions.

## Steps

1. **Resolve the base dir:** glob `docs/*/specs/*-methodology.md`. If no match, instruct the user to run `/anti-drift-planning:init` first and stop.
2. **Read the decomposition roster:** parse §4 of the methodology spec (the plan table). This is the authoritative list of planned plans. If §4 is empty, report "no plans defined yet — run `/anti-drift-planning:new-plan 1 <topic>`" and stop.
3. **For each plan N in the roster:**
   - Locate the spec: glob `<base>/specs/*-plan-<N>-*.md`. If missing → status `not started (no spec)`.
   - Locate the ledger: `<base>/handoffs/plan-<N>-ledger.md`. If missing → status `not started`.

---
description: Run the Phase Z verification audit for plan N. Runs the mechanical lint first, re-reads spec, marks every requirement Done/Deferred/Cut, checks deviations log has zero pending-user, runs the test suite, prompts for user-story walkthrough.
allowed-tools: ["Read", "Edit", "Write", "Bash", "Glob", "AskUserQuestion", "Skill"]
---

# `/anti-drift-planning:verify <N>` — run the Phase Z audit

Execute the verification phase of plan N. **This command is the canonical implementation of Phase Z** (methodology lock 4) — the five audit steps live here and nowhere else. Non-skippable; the only path to the release script and the milestone tag.

## Argument

- `<N>` — plan number to verify

## Steps

1. **Load the `anti-drift-planning` skill** for context.

---
name: anti-drift-planning
description: Use when a project must ship multiple features across separate sessions and the user wants to avoid silent scope cuts, invisible work-in-progress, or work going missing between sessions. Triggers on "roadmap", "split into plans", "multi-phase feature", "avoid drift", "session per plan", "ne rien perdre entre les sessions", "decompose into specs", or when a previously shipped milestone silently missed planned scope.
---

# Anti-Drift Planning

A methodology for shipping multi-phase projects across separate Claude Code sessions without losing scope, silently cutting features, or shipping incomplete work.

## When to use this skill

Activate when:

- The project's vision is significantly larger than one session can ship (multiple features, multiple subsystems)
- The user is recovering from a previous cycle where work was lost silently
- The user will run 2+ plan execution sessions sequentially and each must know what came before