Skills

A collection of agent skills for Claude Code. Each skill is a folder containing a SKILL.md with frontmatter (name, description) and instructions that Claude loads on demand when the task matches.

Available Skills

before-and-after

Captures before/after screenshots of web pages or elements and outputs a PR-ready markdown comparison table, driving the @vercel/before-and-after CLI.

Use it when:

• A PR needs visual proof that a UI change does what it claims
• You want a | Before | After | table generated and uploaded in one step
• Comparing two URLs, two existing images, or a mix of both

Vendored from vercel-labs/before-and-after (PolyForm Shield 1.0.0 — license included in the folder). Install the CLI with npm i -g @vercel/before-and-after agent-browser.

code-structure

Service layer architecture guidance. Enforces a two-layer separation where actions orchestrate domain rules (the "why/when") and a service layer centralizes reusable operational mechanics (the "how").

Use it when:

• Multiple workflows duplicate the same operational logic
• You're deciding what belongs in actions vs. shared services
• A bug fix in one flow doesn't propagate to others doing the same thing
• Adding a feature that shares mechanics with existing ones

Includes a migration checklist for extracting shared logic safely and a table of anti-patterns to avoid (god services, leaky services, over-abstraction).

evidence-driven-testing

Records visual proof while testing UI behavior — the agent drives the app live via computer use (or cua-driver when the harness has no computer-use tools) while the bundled recorder captures the session — then posts the video and a results summary to the PR and tracker issue. The recorder (scripts/evidence.py, Python 3 + FFmpeg) runs on Linux, macOS, and Windows and has doctor, start, annotate, and stop commands: annotations are timestamped as the agent tests, burned into evidence.mp4 on stop, and summarized in a generated report.md and manifest.json. Headless environments swap the recorder for scripted screenshots and Playwright captures; non-UI changes still get evidence (measured numbers, output pairs, transcript excerpts).

Use it whenever a change needs verifiable evidence that it works, instead of prose claims.

The recorder needs ffmpeg/ffprobe built with libx264 and the ass filter, plus a screen-capture source: X11 (DISPLAY) or wlroots Wayland (wf-recorder; GNOME/KDE are not supported) on Linux, Screen Recording permission on macOS, any standard ffmpeg on Windows. python3 scripts/evidence.py doctor reports both. The raw capture is MPEG-TS, so a crashed or hard-killed recorder still yields usable evidence. The headless path needs only a running app and a scriptable browser (Playwright via npx). Posting evidence requires the gh CLI (or equivalent). tests/test_evidence.py smoke-tests the recorder end to end with a synthetic video source (python3 -m pytest tests/ -q).

greploop

Iteratively fixes a PR (GitHub), MR (GitLab), or shelved changelist (Perforce) until Greptile gives a perfect review: 5/5 confidence with zero unresolved comments. Triggers the review, fixes actionable comments, resolves threads, pushes, and repeats (max 5 iterations).

Use it when a PR should be fully optimized against Greptile's code review standards before merge.

Vendored from greptileai/skills (MIT — license included in the folder). Requires Greptile installed on the repo and an authenticated gh/glab/p4 CLI.

greploop-apps

Identical loop to greploop, but triggers reviews by tagging @greptile-apps, which bypasses Greptile's file-count limit on huge PRs that the plain @greptile mention refuses to review — including a fallback that polls Greptile's edited summary comment when no check run appears.

Use it when greploop's trigger gets "Too many files changed for review".

Local variant derived from greptileai's greploop (MIT — license included in the folder); no separate upstream.

new-feature

Starts every new task in an isolated Git worktree branched from origin/main — unique task naming, a scope check against open PRs, fresh dependency installs, and cleanup after merge — so multiple agents can work on the same repo in parallel without conflicts.

Use it when:

• Starting any new feature, fix, or task, before writing code
• Multiple agents (or sessions) work the same repository concurrently
• You need a consistent branch-per-task convention with safe cleanup

Includes harness deltas for Claude Code and Cursor, which manage worktrees themselves.

Workflow

AGENTS.md ties the skills together into a four-beat workflow — isolate (new-feature) → build (code-structure) → prove (evidence-driven-testing) → ship (before-and-after + greploop). Drop it into a repo alongside the skills and fill in the repo-specific callouts (checks, invariants, environment).

Installation

Clone the repo and copy (or symlink) a skill folder into your skills directory:

# Available in all projects
cp -r code-structure ~/.claude/skills/

# Or scoped to a single project
cp -r code-structure /path/to/project/.claude/skills/

Claude Code picks up the skill automatically and invokes it when a task matches the skill's description. You can also invoke one explicitly with /code-structure or /evidence-driven-testing.

Adding a New Skill

1. Create a folder named after the skill (kebab-case).
2. Add a SKILL.md with name and description frontmatter — the description is what Claude uses to decide when the skill applies, so make it trigger-focused ("Use when...").
3. Keep instructions concise and actionable; link out to reference files in the folder if they get long.