forward-deployed-engineer-toolkit

KO | EN

gitlite — search

#api #kotlin #vue #ai #list #html #agent #generative-ai #web #linux

forward-deployed-engineer-toolkit

(★ 9)

opinionated tools for shipping AI into real-world domains

파일 탐색기

ZIP 다운로드

- - braintrust-evals.yml
  - ci.yml
- - panel-review-pr-1.md
- - north-star.md
  - panel-review.md
- append-trace.ts
- braintrust-smoke-test.ts
- flush-traces.ts
- promote-to-dataset.ts
- validate-skills.mjs
- - README.md
  - SKILL.md
- - package.json
  - README.md
  - schema.test.ts
  - schema.ts
  - SKILL.md
- - package.json
  - README.md
  - schema.test.ts
  - schema.ts
  - SKILL.md
  - traces.test.ts
  - traces.ts
.gitignore
CLAUDE.md
LICENSE
package-lock.json
package.json
README.md
tsconfig.json
vitest.config.ts

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

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`](./skills/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. |
| [`mvp-plan`](./skills/mvp-plan/) | v0.1 (draft) | Turns a locked north star into a scoped, gated MVP plan. Checklist-spined artifact from idea to launch; founder-owned success metrics and kill criteria; DRAFT until the founder flips it; cross-model fan-out at lock. |
| [`panel-review`](./skills/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):

```bash
# 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](./LICENSE).

forward-deployed-engineer-toolkit

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README