claude-skills

Use when a thumbs-upped design spec is ready and you want the whole spec-to-PR pipeline to run autonomously — hardens the spec, writes the plan, executes it via dex, opens a PR, runs PR review, and closes with an architect completeness review — in one invocation. Triggers on: ship it, ship-it, run the pipeline, spec to PR, autonomous feature pipeline, conductor.

Use when CLAUDE.md has grown bloated and is taxing every session's context — the inverse of claude-md-improver. Triages each block into session-invariant rules (stay always-on) vs feature/subsystem detail (page out to docs/ behind a plain-path pointer), moving — never deleting — the detail so nothing is lost and the always-on token bill drops. Triggers on: slim claude.md, minimal claude.md, claude.md is bloated, shrink claude.md, split claude.md, make claude.md minimal, page out claude.md, claude.md too long, reduce project memory, point claude.md to docs, trim claude md.

Use to run a whole-system adversarial audit of the current repo across three lenses at once — codebase (defects, incoherences, affordance gaps), docs (drift, structure, coverage), and process (end-to-end workflow dead ends). Dispatches three fresh Claude subagents in parallel, each writing a ranked findings report to docs/audits/, then offers an optional fix pass on CRITICAL findings. Triggers on: review codebase, audit codebase, audit this repo, adversarial audit, codebase audit, docs audit, process audit, fable audit, three reviews, full audit, whole-system review, find everything wrong.

Use after a feature is built — e.g. the final step of a ship-it run, right after review-pr — to check it is actually complete and wired, not just line-correct. Dispatches a fresh-context Claude subagent that traces the whole repo to hunt completeness and integration gaps in the branch's changes, then reports ranked, evidence-backed findings and fixes nothing. Triggers on: architect review, architecture review, completeness review, wiring review, is this wired, did we finish this, find the gaps, incomplete feature, created but not wired, dead code check, deep pass on what we built.

Use BEFORE starting work on a task to surface the unknown-unknowns you're about to walk into but can't see. Dispatches a fresh-context Claude subagent that hunts what you don't know you don't know about the codebase, the domain, and the decisions ahead, calibrated to your stated experience, then reports ranked, evidence-backed findings and fixes nothing. Triggers on: blind spot pass, blindspot, unknown unknowns, what don't I know, what am I missing, help me prompt better, I know nothing about, teach me my unknowns, before I start, de-risk this task, surface my blind spots.

Use when running an already-hardened Superpowers implementation plan through the dex orchestrator (codex backend). Triggers on: convert plan to dex, plan to dex, dex from plan, run plan with dex, plan-to-dex.

Use to raise answer quality on a hard prompt by running it through a blind multi-model panel, then having Opus judge and synthesize one grounded answer. Panelists: 2x Opus subagents, GPT-5.5 via codex, and a local LMStudio model via pi — all over subscription, no API keys. Triggers on: fusion, run through fusion, fuse models, multi-model panel, panel of models, second opinion from multiple models.

Use when a Docker Compose project should stop colliding on host ports across git worktrees or separate projects, by switching containerized services to OrbStack routable domains instead of published host ports. Edits docker-compose and local env files directly, then verifies live. Requires OrbStack as the active Docker engine (aborts otherwise). Triggers on: make orbstack compatible, orbstack migrate, fix docker port collisions, eliminate docker port collisions across worktrees, orbstack ports, switch compose services to orbstack domains, make project orbstack-compatible.

Use when a feature should be implemented as parallel user stories by a swarm: Claude orchestrates, Codex workers implement in isolated worktrees via the Workflow (ultracode) tool. Takes a plain-language feature request or a plan/spec file directly — no prd.json or tasks/ files needed. Triggers on: swarm execute, parallel execute, run stories parallel, swarm this.

Use when a brainstorming design spec has been written and needs adversarial review before implementation planning, using OpenAI Codex as the independent reviewer. Requires the codex CLI. Triggers on: spec review codex, codex spec review, review spec with codex, codex review, review spec, spec review.

Use when a brainstorming design spec has been written and needs adversarial review before implementation planning, using a local model served by LMStudio (via the pi CLI) as the independent reviewer. Works offline — no hosted API needed. Triggers on: spec review local, local spec review, review spec locally, lmstudio review, review spec, spec review.