gstack-extensions

Personal skills layered on top of gstack, discovered by Claude Code alongside gstack's own skills.

Install

git clone <this-repo> ~/dev/gstack-extensions
cd ~/dev/gstack-extensions
./install

Then restart your Claude Code session. Skills become invokable as /pr-watcher, /qa-quincey-manual-browser-testing, etc.

What's included

Standalone skills (skills/):

• /pr-watcher . Foreground watcher that pairs the main agent (dispatcher and fix-applier) with a passive polling subagent (sensor): the sensor blocks silently in one Agent call until CodeRabbit posts a settled round of feedback, then returns a single JSON blob; the main agent classifies, fixes, tests, commits, pushes, and replies on the PR before spawning the next sensor. Invoke manually after /ship.
• /coderabbit-config . Generates a tailored .coderabbit.yaml for the current repo. Detects languages, monorepo shape, generated/vendored dirs, and lifts conventions from CLAUDE.md/AGENTS.md into path_filters, path_instructions, and tools. Wraps the coderabbit CLI for optional live validation.
• /first-principles-thinking . Goal-first reframe coaching skill. Interrupts in-flight optimization-inside-an-inherited-frame with a seven-step walk: goal, success signal, hard constraints, assumed constraints, current path, constraint-class attacks, pressure-test. Light mode by default; escalates to Deep (Socratic, one-question-per-turn) on pushback or missing context. Modeled on the SpaceX rocket-floor and Neuralink no-surgeons reframes.
• /feature-spike . Pre-plan risk-discovery skill. Drives a four-phase loop to cheaply prove or disprove the riskiest unknown of a feature before committing to plan and build: lock a one-line outcome ("I will know X if Y"), prepare isolation (worktree or branch) and open SPIKE.md as a live ledger with a THROWAWAY SPIKE marker, write the leanest possible falsifier (karpathy-guidelines explicitly OFF), escalate to /second-opinion then user when blocked, complete the verdict (PROVEN / DISPROVEN / INCONCLUSIVE) in SPIKE.md. Sits before /plan-eng-review. Burn-after-reading: the branch and code are throwaway; only the SPIKE.md knowledge is meant to survive.

Bundles (top-level dirs with their own skills/ and shared/), skill groups that share common context files:

• PM Penny (pm-penny/): /pm-penny-feature, /pm-penny-bug, /pm-penny-next-issue. Product manager who turns ideas, bug reports, and "what should I work on next?" into well-structured GitHub issues. Shares identity, label conventions, scope/repro gates, and fast-mode logic across the three sub-skills via pm-penny/shared/*.md.
• Feature Frank (feature-frank/): /feature-frank-pr-feedback. Engineer who works through PR review comments, patches the code, and captures durable lessons.
• QA Quincey (qa-quincey/): /qa-quincey-manual-browser-testing, /qa-quincey-manual-headless-testing. Manual QA specialist who verifies one defined flow against the spec or mockup. Browser sub-skill drives the gstack browse daemon autonomously and AI-compares screenshots against Pencil mockups; headless sub-skill (successor to the prior /qa-headless) verifies backend features by capturing side effects. Shares persona, deviation vocabulary, plan/report storage, and reconcile loop via qa-quincey/shared/core.md.

How it works

./install symlinks every directory under skills/ AND every sub-skill under <bundle>/skills/ into ~/.claude/skills/. Claude Code scans that directory at session start and discovers any directory containing a SKILL.md. Bundle sub-skills resolve their shared files (e.g. shared/core.md) relative to the bundle root, which works through the symlink.

This repo lives outside ~/.claude/skills/gstack/, so gstack-upgrade never touches it. gstack and these extensions coexist as peers in the flat ~/.claude/skills/ namespace.

Updating

cd ~/dev/gstack-extensions
git pull
./install   # idempotent; refreshes links and cleans stale ones

Uninstall

./uninstall

Removes only symlinks that point into this repo. Leaves gstack and any other skills alone.

Adding a new extension

For a standalone skill:

1. Create skills/<name>/SKILL.md with valid frontmatter (name, description).
2. Run ./install.
3. Restart Claude Code.

For a bundle (a group of sub-skills that share shared/*.md context):

1. Create <bundle>/skills/<sub-skill>/SKILL.md for each sub-skill at the repo top level (sibling to skills/).
2. Put shared context in <bundle>/shared/; sub-skills reference it as "from the bundle root".
3. Run ./install. Each sub-skill is symlinked into ~/.claude/skills/ directly (no bundle-name prefix on the invocation).