parcha-skills

Our collection of portable Agent Skills for Claude Code, Codex, and pi.

Skill	What it does	Cross-harness note
hands-free	Calls your phone when the coding agent needs an answer or approval.	Same Python/Vapi contract in all three harnesses.
parable	Plans implementation batches, routes work to cheaper executors, verifies, and reviews.	Claude/native subagents are used only when available; stock pi needs a configured CLI-backed executor.
cascade	Carries large projects through verifiable tasks and parallel work.	Keeps a living task graph; executes serially when parallel tools are unavailable.
recall	Indexed local search over prior Claude Code and Codex sessions.	Runs from pi, but does not index pi's own transcripts yet.
recap	Reconstructs everything observable that happened in one exact coding-agent session.	Uses Recall's Claude/Codex evidence; pi can run it but is not yet an indexed source.
tether	Keeps Slack threads attached to the exact agents that created them.	Codex and Claude Code resume natively; stock pi publishes as a headless run. Requires a separately installed private runtime.
desloppify	Turns whole-codebase slop into evidence-backed cleanup with Peter O'Malley's official engine.	One canonical workflow selects honest native or prepared-packet review routes per harness.
autoqa	Points at any repo, discovers how it runs, QAs the live app end-to-end with a witnessed report, and pairs its UI witnesses with base-instance captures as the PR's Before / After.	Uses whatever browser tooling the harness has; API/CLI checks work everywhere.
precap	Recaps a task before doing it: a grounded, past-tense account of the finished work that long-running agents check themselves against for drift.	Pure markdown plus a stdlib Python checker; identical in all three harnesses.
tdd	Red before green at pre-agreed seams, one vertical slice per cycle, regression-first bug fixes.	Pure markdown; identical in all three harnesses.
blast-radius	Finds what a diff breaks beyond the diff and proves the one safety fact by running real code.	Invocation only (disable-model-invocation: true); parallel reviewer subagents are used where the harness has them.
review-loop	Drives every reviewer thread on a GitHub PR (Greptile, Devin, humans) to zero unresolved in bounded iterations.	GitHub only, through the caller's authenticated gh; identical in all three harnesses.
verify-release	Confirms the intended candidate is live in the target environment and the changed path works there, with live evidence.	Uses whatever deployment tooling and observability the repo already has; calls autoqa for matrix coverage.
blast-radius	cursor/plugins, pstack/skills/blast-radius	Lauren Tan
tdd	cursor/plugins, pstack/skills/tdd, adapted from a bug-fix-only workflow into a general red-green reference	Lauren Tan
unslop	Cuts AI tells from anything a human will read.	Pure markdown; identical in all three harnesses.

The skill payloads are canonical skills/<name>/SKILL.md directories. Harness-specific manifests package those same files; there are no Claude/Codex/pi forks to drift apart.

Skill roles in the PR lifecycle

Each shared skill has one job in the lifecycle of a pull request. A repo-specific orchestrator calls the skill by name and never copies its body.

Stage	Skill	Job, in one line
Start	precap	Write the finished-task recap first; detect drift against it.
Build	tdd	Red-green at pre-agreed seams.
Prove	blast-radius	Find what the diff breaks beyond the diff; prove the one safety fact by running code.
Prove	autoqa	Does the running app behave? Witnessed feature-by-modality matrix, baseline plus diff cases.
Describe	autoqa	The same run's UI witnesses, paired with base-instance captures, become the PR's Before / After section.
Review	review-loop	Drive every reviewer thread (bots and humans) to zero, bounded.
Release	verify-release	Is the intended candidate live in the target environment and does the changed path work there?
Everywhere	unslop	Cut AI tells from anything a human reads.

evidence is a rule, not a skill. It stays the always-on snippets/evidence block, and autoqa owns the witness contract (no witness, no verdict; PASS, FAIL, UNTESTED with reason, SKIPPED, BLOCKED). blast-radius proves a single fact by running code; autoqa exercises the app through its entry points and, from the same browser session, produces the paired before/after captures the PR body shows. The two do not overlap: analysis and execution.

Sources and credits

Some skills in this collection are vendored or adapted from other people's work. Each such package keeps the upstream license file in its directory and a Provenance section in its README listing every change made here.

Skill	Source	Author	License
unslop	cursor/plugins, pstack/skills/unslop, by way of michaelshimeles/skills	Lauren Tan	MIT
review-loop	greptileai/skills (greploop, greploop-apps), by way of michaelshimeles/skills	Greptile AI	MIT

verify-release was written for this collection and carries the repository's MIT license.

Instruction snippets

snippets/ contains harness-neutral blocks for always-loaded agent instruction files (AGENTS.md, CLAUDE.md). Skills load on demand through discovery or invocation; a snippet applies to every session unconditionally. Use a snippet when a rule must hold even in sessions where no skill is loaded.

Snippet	Purpose
effective-comms	Makes responses answer-first, concrete, structured, and brief by default, and bans marketing rhetoric in technical writing, while preserving safety and completeness exceptions.
evidence	Ties claims to witnessed evidence and keeps sensitive evidence out of public repositories.
control	Preserves user control over consequential actions and gives errors a bounded recovery path.
systems-thinking	Prevents local fixes from making the larger system worse without blocking bounded work.

The install commands below cover skills only. To install a snippet, paste its AGENTS.md block near the top of your root instruction file and keep the <!-- name:start/end --> markers so tooling can update the block in place.

Install with skills.sh

Browse all fifteen skills at skills.sh/Parcha-ai/parcha-skills, or install interactively:

npx skills add Parcha-ai/parcha-skills

Install one directly with --skill:

npx skills add Parcha-ai/parcha-skills --skill hands-free
npx skills add Parcha-ai/parcha-skills --skill parable
npx skills add Parcha-ai/parcha-skills --skill cascade
npx skills add Parcha-ai/parcha-skills --skill recall
npx skills add Parcha-ai/parcha-skills --skill recap
npx skills add Parcha-ai/parcha-skills --skill tether
npx skills add Parcha-ai/parcha-skills --skill desloppify
npx skills add Parcha-ai/parcha-skills --skill autoqa
npx skills add Parcha-ai/parcha-skills --skill precap
npx skills add Parcha-ai/parcha-skills --skill tdd
npx skills add Parcha-ai/parcha-skills --skill blast-radius
npx skills add Parcha-ai/parcha-skills --skill review-loop
npx skills add Parcha-ai/parcha-skills --skill verify-release
npx skills add Parcha-ai/parcha-skills --skill unslop

Add --global for a user-level install or --agent claude-code, --agent codex, or --agent pi to choose a destination explicitly. The skills.sh CLI discovers the same canonical payloads used by the native installs below; npm publication of the individual packages is not required.

Recall and Tether are instruction-only packages here. Their engines, runtimes, installers, and deployment configuration are maintained separately. Install the systems through your organization's private distribution before using the skills.

Install for Claude Code

claude plugin marketplace add Parcha-ai/parcha-skills
claude plugin install hands-free@unc-skills
claude plugin install parable@unc-skills
claude plugin install cascade@unc-skills
claude plugin install recall@unc-skills
claude plugin install recap@unc-skills
claude plugin install tether@unc-skills
claude plugin install desloppify@unc-skills
claude plugin install autoqa@unc-skills
claude plugin install precap@unc-skills
claude plugin install tdd@unc-skills
claude plugin install blast-radius@unc-skills
claude plugin install review-loop@unc-skills
claude plugin install verify-release@unc-skills
claude plugin install unslop@unc-skills

Install only the skills you want. Start a new session after installation.

Install for Codex

codex plugin marketplace add Parcha-ai/parcha-skills
codex plugin add hands-free@unc-skills
codex plugin add parable@unc-skills
codex plugin add cascade@unc-skills
codex plugin add recall@unc-skills
codex plugin add recap@unc-skills
codex plugin add tether@unc-skills
codex plugin add desloppify@unc-skills
codex plugin add autoqa@unc-skills
codex plugin add precap@unc-skills
codex plugin add tdd@unc-skills
codex plugin add blast-radius@unc-skills
codex plugin add review-loop@unc-skills
codex plugin add verify-release@unc-skills
codex plugin add unslop@unc-skills

Codex uses the native .agents/plugins/marketplace.json and package .codex-plugin/plugin.json manifests. Start a new session after installation.

Install for pi

pi install git:github.com/Parcha-ai/parcha-skills

The repository is one pi package that exposes all fifteen skills. In pi, invoke one explicitly with /skill:<name>, for example /skill:hands-free, /skill:autoqa, or /skill:review-loop.

Compatibility checks

The local gate runs all fifteen skills across the three harnesses for native installation, discovery, and credential-free smoke checks. Raw test output stays local and is not committed.

Run the local gate with:

npm test
for package in hands-free parable cascade recall recap tether desloppify autoqa precap tdd blast-radius review-loop verify-release unslop; do (cd "$package" && npm test); done
python3 scripts/prove_portability.py --output /tmp/unc-skills-portability