ab-method
A workflow system for Claude Code and Codex. It grills a problem into a domain-grounded plan, then either drives it through test-driven missions you review one at a time, or hands it to an autonomous /goal loop that runs to a verifiable stop condition.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Check it deeper on your machine
It has changed since its scan on 11 Nov 2026.
As a Claude Code plugin, install it and then ask Claude Code:
/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/ayoubben18/ab-method with scanskillsafety
Or from a terminal, clone it and run it:
git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/ayoubben18/ab-method'
It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.
Badge for your README
Markdown
[](/listings/ayoubben18--ab-method)HTML
<a href="/listings/ayoubben18--ab-method"><img src="/api/listings/ayoubben18--ab-method/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
npx ab-method
The installer detects your environment and installs accordingly:
| Detected | Result | |-----------------|---------------------------------------------------------| | `.claude/` only | Slash commands + skills under `.claude/` | | `.agents/` only | Skills under `.agents/skills/` (Codex layout) | | Both | Both targets receive the right files | | Neither | Asks which to install (default: both) |
It also installs in every case:
- `.ab-method/` — workflow definitions and the structure index - `docs/architecture/` and `docs/tasks/` — output scaffolding - Helper skills: `grill-with-docs`, `grill-me`, `tdd`, `domain-model`, `codebase-design`, `ubiquitous-language`, `critique-plan`, `reconcile-roadmap`, `review-implementation`, `sync-architecture`, `improve-codebase-architecture`, `request-refactor-plan`, `to-issues`, `to-prd`, `write-a-skill` - Workflow skills: `ab-create-task`, `ab-create-goal`, `ab-analyze-project`, and one per workflow - Slash commands (Claude only): `/ab-master` plus one per workflow - `AGENTS.md` (Codex only): orients Codex and lists the workflow skills - Optional built-in subagents (Claude only, prompted)
Skills are copied as real files, not symlinks — portable across OS and CI.
As a plugin (Claude Code and Codex)
This repository is also a plugin marketplace for both harnesses. Nothing is copied into your project: the workflows and skills live in the plugin and update with it.
Claude Code:
/plugin marketplace add ayoubben18/ab-method /plugin install ab-method@ab-method
Codex:
codex plugin marketplace add ayoubben18/ab-method codex plugin add ab-method@ab-method
Every skill is namespaced the same way on both sides — `ab-method:<name>`. Run a workflow with `/ab-method:ab-create-task` (Claude Code) or `$ab-method:ab-create-task` (Codex), or describe what you want and the matching skill fires.
| | `npx ab-method` | Plugin | |---|---|---| | Where it lives | Copied into the project (`.ab-method/`, `.claude/`, `.agents/`) | In the harness's plugin cache, shared by every project | | Updates | Re-run the installer | With the plugin | | Trigger (Claude Code) | `/create-task`, `/ab-master`, ... | `/ab-method:ab-create-task`, ... | | Trigger (Codex) | `$ab-create-task` | `$ab-method:ab-create-task` | | Custom paths | Edit `.ab-method/structure/index.yaml` | Create that same file in the project — it overrides the bundled default | | Built-in subagents | Optional, prompted | Not included |
The plugin ships the skills only. The slash commands under `.claude/commands/` are thin pointers at the same workflows the `ab-*` skills already run, so shipping both would list every workflow twice; the eight built-in subagents are stack-specific and stay an `npx` opt-in.
**How a plugin install finds `.ab-method/`.** Workflows and skills name `.ab-method/core/<workflow>.md` and `.ab-method/structure/index.yaml`. Each is looked up in the project first and, when the project has none, in the copy bundled with the plugin. So a project-level `index.yaml` always wins, and a project with no `.ab-method/` at all just works. Paths inside the index (`docs/...`, `CONTEXT.md`) are always relative to the project, never to the plugin.
To set it up for a whole team, commit this to the repo's `.claude/settings.json` — the marketplace registers when a teammate trusts the folder:
{
"extraKnownMarketplaces": {
"ab-method": {
"source": { "source": "github", "repo": "ayoubben18/ab-method" }
}
},
"enabledPlugins": {
"ab-method@ab-method": true
}
}Don't mix the two installs in one project on Claude Code — you would get each workflow twice (`ab-create-task` from `.claude/skills/` and `ab-method:ab-create-task` from the plugin).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 17 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.
- {
- "name": "ab-method",
- "owner": {
- "name": "Ayoub"
- },
- "description": "The AB Method plugin for Claude Code.",
- "plugins": [
- {
- "name": "ab-method",
- "source": "./",
- "description": "Incremental task management for coding agents: grill the problem, plan test-driven missions, bracket every implementation with critics.",
- "author": {
- "name": "Ayoub"
- }
- }
- ]
- }
README.md · README · 97 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.
- # AB Method
- 
- A workflow system for Claude Code and Codex. It grills a problem into a domain-grounded plan, then either drives it through test-driven missions you review one at a time, or hands it to an autonomous `/goal` loop that runs to a verifiable stop condition.
- ## Install
- ```bash
- npx ab-method
- ```
- The installer detects `.claude/` and/or `.agents/` and installs slash commands and skills for Claude Code, Codex, or both (it asks if neither exists). It also adds `.ab-method/` (workflow definitions and the path index), `docs/` scaffolding, `AGENTS.md` for Codex, and optional subagents for Claude Code.
- ### Or as a plugin
- Nothing is copied into your project; the plugin updates on its own.
- ```bash
- # Claude Code
- /plugin marketplace add ayoubben18/ab-method
- /plugin install ab-method@ab-method
- # Codex
- codex plugin marketplace add ayoubben18/ab-method
- codex plugin add ab-method@ab-method
- ```
- Skills are namespaced `ab-method:<name>`: run `/ab-method:ab-create-task` (Claude Code) or `$ab-method:ab-create-task` (Codex), or describe what you want. A project's own `.ab-method/structure/index.yaml` overrides the bundled one. The plugin ships no slash commands or subagents. Don't combine it with `npx ab-method` in one project, or every workflow shows up twice.
- ## Commands
- On Claude Code, run a workflow as a slash command. Codex has no shared slash commands, so use the matching skill (`$ab-create-task`) or describe your intent.
- | Command | Purpose |
- |---|---|
- | `/mastermind` | Entry point: routes your intent to the right workflow |
- | `/create-task` | Grill a problem into a holistic task: deepen what it touches, design new seams, composed missions |
- | `/resume-task` | Continue a task from its progress tracker |
- | `/start-task` | Run a task autonomously to a reviewed PR, one commit per mission (`inline` skips mission subagents) |
- | `/extend-task` | Add missions to an existing task |
- | `/create-task-from-handoff` | Turn a tangent parked mid-grill into a task |
- | `/create-goal` | Write a prompt for an autonomous `/goal` loop |
- | `/extend-goal` | Extend a goal from what its run built |
- | `/create-roadmap` | Split a bigger idea into a dependency-ordered graph of tasks |
- | `/start-roadmap` | Run a planned roadmap in dependency order |
- | `/analyze-project` | Document domain, tech stack and patterns (`/analyze-frontend`, `/analyze-backend` for one side) |
- | `/update-architecture` | Refresh architecture and domain docs |
- | `/ab-cleanup` | Compact docs to budget and remove task leftovers; also the 3.x → 4.0 migration |
- | `/ab-master [name]` | List workflows, or run one by name |
- **Task, goal or roadmap?** A task keeps you in the loop mission by mission. A goal is one objective with a checkable stop condition, run unattended. A roadmap is several ordered tasks; it can leave open decisions and uncharted areas for later.
- ## How it works
- 1. **Always grill.** Every task and goal starts with `grill-with-docs`, which checks your terms against `UBIQUITOUS_LANGUAGE.md` and `CONTEXT.md` and records decisions as ADRs.
- 2. **Plan before code, holistically.** Each task first zooms out: if the feature would bend around the current design, it plans the re-architecture as part of the task instead of a workaround, because agents make rewriting cheap. Then it scans the modules it will touch and plans the refactors that deepen them first. Every capability gets a placement (reuse, extend, extract, adopt or introduce); a large or already-solved capability is checked against open-source libraries first, and one is adopted only if it is actively maintained and fits your exact stack; a new seam is designed several ways before any code, and new UI gets a throwaway HTML prototype to pick from. Missions are composed slices, not one-layer steps.
- 3. **Seams are documented.** `docs/seams/` records where the app's seams live and what callers must never do. Seven defaults come from *A Philosophy of Software Design*; `analyze-project` maps them onto your code.
- 4. **Test what must not break.** After coding, a protect step tests through the seams and proves each test can fail. Bug fixes stay test-first (`verify: repro`).
- 5. **Critics on both sides.** `critique-plan` checks the plan against the domain model and the seams, and flags workarounds a reshape would remove, before coding; `review-implementation` and `sync-architecture` check the diff and the docs after; `review-pr` reviews the open PR against the seams and the coding rules. They stay silent unless there is a real problem.
- 6. **Measure the blast radius.** `change-map` predicts which modules a task will touch, then compares that with the real diff.
- 7. **Park, don't guess.** A question nobody can answer yet goes in `unresolved-questions.md` with a placeholder behind one seam (`TODO(UQ-n)`); a tangent becomes a handoff.
- 8. **Small docs.** Domain and architecture docs have word budgets, hold only current state, and are checked by `npx ab-method check-docs`.
- 9. **Parallel only when you agree.** Missions and roadmap tasks run sequentially unless you allow tagging them for parallel subagents. `/start-task <task> inline` runs every mission in the conversation instead of in subagents.
- ## File layout
- ```
- .ab-method/
- core/ Workflow definitions
- rules/ Communication, code and docs rules for every workflow
- structure/index.yaml Paths, outputs and doc budgets (all configurable)
- UBIQUITOUS_LANGUAGE.md Domain glossary
- CONTEXT.md Bounded-context overview
- docs/
- architecture/ tech-stack.md, frontend/ and backend/ pattern topics
- adr/ Decision records
- seams/ Where the app's seams live (living docs)
- tasks/<task>/ progress-tracker.md, change-map.md, review.md, ...
- goals/<goal>/ goal.md + progress-tracker.md
- handoffs/ Tangents awaiting their own task
- ```
- ## Upgrading to 5.0
- Existing trackers keep working: legacy `verify: test` missions and `## Test Plan` sections still run through `tdd`.
- - **Map your seams.** Run `/analyze-project` and pick `[5] Seams only` to write `docs/seams/` from the code.
- - **`/start-task` now ends in a pull request.** It pushes the branch, opens a draft PR, reviews it with `review-pr`, then marks it ready. It needs a git remote and an authenticated `gh`; without them it writes the PR body to `docs/tasks/<task>/pr-body.md` and reviews the branch locally.
- - **`/test-mission` is gone.** Tests now come from the protect step every task runs after coding. The npx installer doesn't delete files, so remove `.ab-method/core/test-mission.md` and `.claude/commands/test-mission.md` from repos installed before 5.0.
- - Coming from 3.x: run `/ab-cleanup` once to migrate `index.yaml` and split pattern docs.
- ## Contributing
- See [CONTRIBUTING.md](CONTRIBUTING.md) for developing the plugin.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.