Forward Deployed Engineer Toolkit

A growing collection of opinionated tools for shipping AI into real-world domains. Each tool ships in two shapes: a Claude Code skill, and an importable schema or library, so the same artifacts can be produced by an agent and consumed by an app.

Contents

Skills

Skill	Status	Purpose
north-star	v0.2 (draft)	Establishes a project's mission and gut-check before brand or implementation work. Conversation-driven; produces a 3-5 sentence paragraph artifact with typed YAML frontmatter.
panel-review	v0.1 (draft)	Multi-agent code review of a PR or current-branch diff. Three-model fan-out (Opus, Sonnet, Haiku), discrete severity, Sonnet confidence recheck plus comment polish in one pass, dev approval gate before posting. Worktree-isolated. Build log committed alongside the PR. Optional braintrust trace logging.

Design principles

• Dual-shaped tools. Every entry exposes both an agent-facing surface (a skill) and an app-facing surface (a schema or library). The same artifact can be produced inside Claude Code and consumed by a downstream application.
• Opinionated, not configurable. Each tool encodes a specific point of view and refuses to be a do-it-all framework.
• Markdown plus YAML frontmatter as the lingua franca. Artifacts are human-readable markdown with typed YAML frontmatter, parseable by any consumer.

Contributing

Changes go through pull requests. Direct pushes to main are not allowed.

Workflow:

1. Cut a branch from main.
2. Make changes; run npm test and npm run validate:skills locally before pushing.
3. Open a PR.
4. Run a Claude Code review on the PR before merging. Use the /review skill (or equivalent automated reviewer) to surface drift, missing tests, design issues, and style violations before a human reviewer looks at it.
5. Wait for CI to pass: skill-shape validator, unit tests, and braintrust integration must all be green.
6. Address review findings.
7. Merge.

Branch protection on main enforces these rules at the GitHub layer (PR required, CI must pass, no force-pushes, no direct commits).

Braintrust integration

The panel-review skill ships with optional braintrust trace logging. CI's Braintrust integration job smoke-tests the SDK on every PR, so the secret must be set before merging anything that touches the trace surface.

Setup (one time):

# Local: in ~/.bashrc or your shell rc
export BRAINTRUST_API_KEY=<your key from braintrust.dev>

# Repo secret: required for CI
gh secret set BRAINTRUST_API_KEY --repo mollyretter/forward-deployed-engineer-toolkit

The project name fdet-panel-review is hardcoded for v0.1. Create the project in braintrust.dev before the first CI run; the SDK will not auto-create it.

Skill-shape validation

scripts/validate-skills.mjs runs in CI on every PR. For each directory under skills/ it checks:

• A SKILL.md exists and starts with valid YAML frontmatter
• Required frontmatter fields are present and non-empty: name, description
• name matches the directory name
• description contains no em dashes (style preference; em dashes interrupt scannability and are hard to read aloud)
• A sibling README.md exists

Run it locally with npm run validate:skills.

Unit tests

npm test runs vitest on every *.test.ts file. Each skill that ships a TypeScript schema (schema.ts) should also ship schema.test.ts with at minimum: validator accepts a fully-shaped object, validator rejects each missing or wrong-typed field, parser behaves as documented (throws if it's a stub, parses correctly otherwise).

Tests are deliberately less intense than evaluators. They catch structural and type-level drift, not the semantic quality of a skill's outputs. Semantic quality is checked by the cross-model review that's built into each skill's lock phase, plus eventual full evals once a skill produces enough artifacts to graded against a dataset.

License

MIT. See LICENSE.