Buildwright

Spec-driven autonomous development. Humans approve intent; agents handle everything else.

What this skill does

When activated, Buildwright directs the agent to:

1. 1. Read your codebase and steering documents
2. Write a one-page spec (docs/specs/[feature]/spec.md)
3. Stop for human approval — unless INLINECODE1
4. Implement the feature with TDD
5. Run quality gates: typecheck, lint, test, build
6. Run optional security scans (if semgrep / gitleaks / trufflehog are installed)
7. Run a Staff Engineer prompt-based code review
8. Commit, push, and open a PR via INLINECODE2

Requirements

Credentials (required)

The token must have repo scope to push branches and create pull requests. For minimal privilege, use a fine-grained personal access token scoped to a single repository with "Contents: Read and write" and "Pull requests: Read and write" permissions.

Alternatively, if you use SSH for git push, the GITHUB_TOKEN is still needed for gh pr create. You can use gh auth login to authenticate the GitHub CLI separately.

Binaries (required)

Optional tools

Agent Personas (prompt-based, no binaries)

Staff Engineer and Security Engineer are prompt-engineering personas — instructions loaded from .buildwright/agents/ files in the workspace. They are not external tools or binaries. The agent adopts these personas to review specs and code using defined criteria and confidence thresholds. These files contain only prompt instructions and review checklists — no secrets or credentials.

Configuration

BUILDWRIGHTAUTOAPPROVE (optional, not a credential)

This is an optional boolean flag that controls whether the agent waits for human approval at the spec stage. It is not a secret and not declared in requires.env because it is not required to run the skill.

Recommendation for first use: Leave BUILDWRIGHT_AUTO_APPROVE unset until you have reviewed a few specs and are comfortable with the workflow. Start with interactive mode in a sandbox repository to observe behavior before enabling autonomous commits and PRs.

Commands

/bw-new-feature \

Full pipeline for new features. Auto-detects greenfield vs existing projects.

CODEBLOCK0

Flow: Detect (greenfield or existing?) → Research → Spec → Staff Engineer validates → Human approves → TDD build → Verify → Security scan → Code review → PR

Artifacts produced:

• - docs/specs/[feature]/research.md — what the agent found in your codebase
• INLINECODE23 — implementation plan with approaches considered

/bw-claw \

Multi-agent pipeline using the Claw Architecture. Architect decomposes the feature into domain-specific claw tasks (UI, API, DB), defines interface contracts, and coordinates execution.

CODEBLOCK1

Flow: Architect analyzes → Decomposes into claw tasks → Defines interface contract → Claws execute per domain (TDD) → Architect integrates → Buildwright quality gates → PR

Best for: Features that cross domain boundaries (e.g., need DB schema + API endpoint + UI component).

Artifacts produced:

• - docs/specs/[feature]/claw-plan.md — decomposition plan with interface contracts
• INLINECODE25 — per-claw execution report

/bw-quick \

Fast path for bug fixes and small tasks (<2 hrs). No spec, no approval step. Runs security scan and code review on the changed diff before committing.

Flow: Understand → Research → TDD → Verify → Security scan → Code review → Commit

CODEBLOCK2

/bw-ship \[message\]

Quality pipeline for existing work: verify → security → review → PR.

CODEBLOCK3

/bw-verify

Quick checks only: typecheck → lint → test → build.

/bw-analyse

Analyse an existing codebase and write structured docs to .buildwright/codebase/. Creates docs from scratch if missing; auto-refreshes existing docs when BUILDWRIGHT_AUTO_APPROVE=true (only asks in interactive mode). Creates tech.md from template if it doesn't exist, then populates it with the discovered stack and commands. Run this first on any brownfield project to give every subsequent session real context.

CODEBLOCK4

Produces: STACK.md, ARCHITECTURE.md, CONVENTIONS.md, CONCERNS.md under .buildwright/codebase/.

/bw-plan \

Research a question or topic and produce a written deliverable — no implementation, no commits.
Use when someone asks a question or needs an analysis, plan, or report before (or instead of) writing code.

CODEBLOCK5

Flow: Understand question/task → Clarify if needed → Research (read code + run read-only tools) → Synthesize findings → Write deliverable → Summarize

Accepts two invocation styles:

• - Inline question — describe the question or topic directly; the agent infers scope and writes to INLINECODE34
• Task file — a structured .md file with Inputs, Rules, Research Areas, and Outputs blocks; the agent parses and executes it exactly

Use /bw-plan when you want: a performance review, architecture decision record, migration plan, technology evaluation, static analysis report, or any "research this and give me a written output" task.

Use /bw-new-feature instead when you want the plan executed (research + spec + implement + ship).

Artifacts produced: whatever the task specifies; at minimum plan.md in the output directory.

Hard constraints: never modifies source files, never commits or creates PRs, every finding must cite evidence.

/bw-help

Show all available commands.

Failure Behavior

If any gate fails after retries, the agent commits completed work, pushes, and opens a PR with a structured failure report. It does not leave orphaned branches or silent failures.

Retry Policy

Security Considerations

This skill performs autonomous code changes, commits, and pull requests. Understand what it does before enabling it on repositories with sensitive or production code.

What the skill reads: Your repository source code, .buildwright/agents/ persona files (prompt instructions only, no secrets), and .buildwright/steering/ context files.

What the skill writes: Spec files under docs/specs/, source code changes, git commits on feature branches, and pull requests via gh.

What the skill does NOT do: It does not modify .env files, access secrets stores, run destructive git operations (force push, reset), or merge PRs. All changes go to feature branches with PRs for human review.

Recommended setup for first use:

1. 1. Start with a fork or sandbox repository, not production code
2. Leave BUILDWRIGHT_AUTO_APPROVE unset (interactive mode) to review specs before builds
3. Use a fine-grained GitHub token scoped to a single repository with minimal permissions
4. Rotate tokens regularly and revoke when no longer needed
5. Review generated PRs before merging — the skill creates PRs, it does not merge them

More Information

Full documentation, source code, and setup instructions: https://github.com/raunakkathuria/buildwright