skill

tempest

Tempest — a cross-harness software factory: opinionated build & QA workflows as portable Agent Skills for Claude Code and pi.

mtthsnc/tempestcommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/mtthsnc/tempest itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 22, 2026

Tempest

A software factory for AI coding agents — opinionated build workflows shipped as portable Agent Skills that run identically on Claude Code and pi.

A sprint loop — office-hours → plan → build → review → ship, plus browser-driven QA — encoded as skills your agent loads on demand. One source tree, two harnesses, no per-tool forks.

Status: v0.1. The full sprint-loop spine, the QA suite (qa/verify/bench), and the bcp brand bridge are implemented; the conformance gate and tests are green.

Why it's portable

Both Claude Code and pi implement the same Agent Skills SKILL.md standard (name + description frontmatter, progressive disclosure). pi reads ~/.agents/skills/ and ~/.claude/skills/ natively and exposes every skill as /skill:<name>; slash commands map to pi prompt templates through a shared argument subset. So a single plugins/tempest/skills/ tree installs to both with only thin manifest/symlink differences. See docs/ARCHITECTURE.md.

Install

Quick (both harnesses, symlink-based)

git clone https://github.com/mtthsnc/tempest.git
cd tempest
./install.sh            # or --no-claude / --no-pi to install one

This symlinks skills into ~/.agents/skills/ (read by pi) and ~/.claude/skills/, symlinks commands into ~/.claude/commands/, and registers the commands directory as a pi prompt-template path in ~/.pi/agent/settings.json. Re-run after git pull; it is idempotent. Remove with ./uninstall.sh.

Claude Code via the plugin marketplace

/plugin marketplace add mtthsnc/tempest
/plugin install tempest@tempest

pi as a package

pi install git:github.com/mtthsnc/tempest

Restart your agent sessions after installing so the skills and commands load.

Use

The sprint loop — office-hours → plan → build → review → ship (with reflect handled by the separate reflect project):

Workflow	Skill	Command	Status
Reframe a raw idea with forcing questions	`office-hours`	`/office-hours <idea>`	✅ implemented
Reviewed implementation plan	`plan`	`/plan <task>`	✅ implemented
Execute an approved plan increment	`build`	`/build <increment>`	✅ implemented
Staff-level diff review (optional auto-fix)	`review`	`/review <diff/PR>`	✅ implemented
Land a reviewed change (PR, CI, deploy)	`ship`	`/ship <what>`	✅ implemented

Each ships as a deliberately generic starter skill — a spine you customize with your own standards. In Claude Code, type /plan <task> or let the agent auto-invoke the skill. In pi, type /plan <task> or /skill:plan <task>.

Quality & QA (browser-driven)

Workflow	Skill	Command	What it does
Browser UI acceptance / E2E / smoke	`qa`	`/qa <flow/URL>`	Drives the real UI, asserts, captures screenshot evidence
Run-and-observe verification	`verify`	`/verify <change>`	Runs the app and confirms real behavior (hands UI off to `qa`)
Performance / Core Web Vitals	`bench`	`/bench <URL>`	Measures LCP/CLS/INP vs a budget or baseline

qa and bench are powered by agent-browser — a browser-automation CLI. The skills are harness-adaptive over the same command vocabulary: on pi they use the native agent_browser tool (pi-agent-browser-native), and on Claude Code they drive the agent-browser CLI directly via the shell.

Dependency: install the agent-browser CLI and keep it on PATH (see its install docs); on pi also run pi install npm:pi-agent-browser-native. Then ./scripts/doctor.sh verifies the setup (it's also run, non-fatally, at the end of ./install.sh).

Brand (BCP)

Workflow	Skill	Command	What it does
Brand Context Protocol	`bcp`	`/bcp [scaffold \| <task>]`	Scaffold/operate a brand: capture truth, produce on-brand output, score it

bcp bridges the dev stack to BCP — a git-native Brand Context Protocol where a brand is cited, dated truth that humans and agents build from. The skill scaffolds the BCP template into a project and routes to its own skills (brand-truth, landing-page, brand-check); the brand lives in its own owned repo, not locked inside this tooling.

Develop

The skills are the product; everything else is thin wiring. See AGENTS.md for the golden rules and docs/ARCHITECTURE.md for the design. Before committing:

./scripts/validate.sh   # conformance gate (the "Brand Check")
./tests/run.sh          # sandbox install + wiring assertions