Claude Skills by AcKeskin

Own the goal-directed autonomy contract — effort / stopping / ask — that execute, checkpoint, orchestrate, and the rule-prime hook read at their decision points, plus dispatch/coordinate/wrap. Set as persistent default, per-task at kickoff, live mid-flight steer, or by short interview. Use on /autonomize, \"set how hard to push\", \"how autonomous should you be\". Mode A — never auto-fires.

Author a concrete blueprint for a slug — what we wanted (intent from vision + spec) and what we're building now (the mature concrete shape: classes, interfaces, dependencies, module relationships, build order) with Mermaid UML. Optional, after /draft-plan — the commit-point view an engineer reaches for before coding. Also re-blueprints existing code (--from-code). Use on /blueprint [slug] or \"blueprint this\" / \"show the concrete shape\". Never auto-fires.

Ideation organ upstream of /envision — talk a half-formed idea into shape, or crystallize a blurry one into a name, one-line description, edges (in/out), and end-goal. Converges (doesn't just chat); writes a light .claude/ideas/<slug>.md note that /envision picks up. Use on /brainstorm [slug], "help me shape this idea", "I have a vague idea for…", "crystallize this". Mode A — propose-confirm, never auto-fires.

Capture a memory (lesson / decision / rule / preference / feedback / project fact / reference) via propose-confirm-commit. Use on /capture, \"remember that\" / \"save this\", or skill invocation. Auto-proposes (never silent-writes; accept/edit/reject gates) only when high-stakes: a reversal of a shipped decision, a correction phrased with finality, or a warning-shaped \"X broke Z, hard to undo\". Otherwise never auto-fires.

Scope-dialed fit-and-intent audit — \"does this serve the original point + cohere with the whole + what did I learn?\" — at diff / module / corpus zoom (--scope overrides auto-detection). Use on /checkpoint, \"does this fit\", \"did I drift\". Never auto-fires.

Close out the scope chain on a shipped slug — leaves the canonical spec reconciled with what shipped, spent plan/blueprint artefacts archived, and one ship line recorded. The terminus after /execute meets done-criteria. Use on /close-out <slug>, \"close out this slug\", \"reconcile what shipped\", \"wrap up <slug>\". Mode A — never auto-fires; all writes behind propose-confirm.

Keep multiple live Claude sessions aligned via a shared, file-based session board (.claude/sessions-active.md, gitignored) — session ownership, hand-off notes, file/slug collision detection; a session can claim coordinator to assign work. No daemon or polling. Use on /coordinate [register|check|handoff|coordinator|done], \"what are other sessions doing\", \"hand off to the other session\". Mode A.

Render a selection of retrieved fragments into Claude's working context per the deliver format contract. Library-only — callable by other skills (discover, prep, review, retrospect-core) that need to present bodies, not a user-invoked skill. Stateless.

Load relevant stored memories and codemap entries for the current task. Use when the user types /discover, asks to "load context," or a task is starting with stored context unsurfaced. Also callable by other skills (prep, review).

Use when facing 2+ independent tasks (different test files, different subsystems, different bugs) that can be investigated or fixed in parallel without shared state or sequential dependencies.

Produce a versioned plan under `.claude/plans/<slug>/v<N>.md` from the active spec for that slug, pinned to a specific spec version and presented for review (accept / edit / reject) before it is written to disk. Use when the user types /draft-plan or asks to plan out the implementation of a specced feature. Mode A only — never auto-fires.

Interview-driven top-level project vision — intent, UX shape, module partition, boundaries, non-goals — into a versioned .claude/visions/<slug>/. Breadth-over-depth, mandatory module-map; the once-per-project doc upstream of /spec. Use on /envision or "sketch / partition / shape a new project". Not for one piece of an existing system (use /spec). Never auto-fires.

Run a plan step-by-step with verification gates, honoring the autonomy contract — never advances silently past a failure. Use when a drafted plan is ready to implement. User-invoked: /execute [slug | path] [--from N] [--strategy sequential|subtask|multi-agent]. Never auto-fires.

Observe a scope's dominant code conventions and write a project-tier conventions.md rule overriding universal defaults; conflicts with shipped rules are surfaced, never silently overridden. Use on /extract-conventions or \"codify the house style\". Mode A — never auto-fires.

Extract the project's ubiquitous language — term → definition → code symbol(s) → collision note — into a project-tier .claude/rules/glossary.md primed by the rule-prime hook and cited by the naming audit. Use on /glossary or \"write down the domain vocabulary\". Mode A — never auto-fires, never auto-writes.

Project an LLM-optimized planning artefact (a plan, blueprint, spec, or vision) into a human-readable approval view — the goal + the concrete decisions in plain prose, so you can approve it and check the discussions were pointing at the right thing. Callable by the draft-plan / blueprint review gates; also /human-view <artefact|slug> on demand. Extends two-doc-tracks to the working set. Mode A.

Detect, score, and rewrite AI-generated texture in user-facing prose — docs, professional email, PR/proposal/issue — preserving every argument; verdicts are advisory, never binary. Use on /humanize or \"make this sound human / does this read like AI\". Refuses the terse model corpus. Mode A — never auto-fires.

Improve any prompt — for LLMs/chat/agents or for generative image/video/audio models — so it produces more reliable, on-target output. Model-agnostic; uses techniques that hold across vendors. Interviews to fill real gaps, then returns a rewritten prompt plus a short rationale. Use when the user types /improve-prompt, pastes a prompt and asks to "make this better / sharpen / fix this prompt", or wants help phrasing a request to any AI model. Mode A only — never auto-fires.

Audit the project's memory tree for integrity drift + bloat — orphans, duplicates, broken relations, stale references, schema gaps, uncompressed bodies. Ten dimensions, propose-confirm-commit per finding, read-only by default. Use when the user types /memory-audit. Mode A only — never auto-fires.

Scaffold a new Claude Code subagent under agents/ — interview-driven, forces the user to articulate job-to-be-done, pre-flight questions, anti-patterns, and a debugging workflow rather than a generic persona. Use when the user types /new-agent or asks to create / scaffold / add a subagent. Mode A only — never auto-fire.

Generate a vendor-neutral AGENTS.md by projecting the Claude-Code instruction corpus — CLAUDE.md tree, rules, landmines, canonical commands — into a flat file other coding agents read, interviewing only for gaps the corpus can't know. Use on /new-agents-md or \"make the repo readable by non-Claude agents\". Mode A — never auto-fire.

Scaffold a new Claude Code hook end-to-end from a recipe — produces the hook file, payload fixtures, a Node-based runner, and the ~/.claude/settings.json registration (diff previewed). Use when the user types /new-hook or asks to create / scaffold / add a hook. Mode A only — never auto-fire.

Scaffold a new MCP server project under mcps/ — interview-driven, supports TypeScript and Python, simple tool servers and API wrappers; delivers a built, dependency-installed project registered in ~/.claude.json. Use when the user types /new-mcp or asks to create / scaffold / add an MCP server. Mode A only — never auto-fire.

Goal-directed orchestration and convergence for concurrent work — turns a single goal into parallel agent work and verifies it back into one result. Use when the user types /orchestrate or says \"split this across agents\", \"set up worktrees for these branches\", \"fan this goal out\". Mode A only — never auto-fires.

Draft a PR title and body — Summary + Test plan + Closes-line on top of any PR template. Auto-fires when about to run \"gh pr create\", after a first push with no open PR, or on \"open a PR\". Hands over a ready-to-run gh command; NEVER runs the GitHub write itself. Skip on the default branch. /pr-author.

Review a GitHub PR — structured findings on correctness / design / hygiene / security; no GitHub posting. User-invoked: /pr-review [number] or \"review PR 123\". Reviews an incoming external diff — not your local code (use review), not bug-hunting your own change (use code-review); does not auto-fire.

Triage a PR's unresolved review comments with you, deciding per comment whether to act on it, defer it, or record your reasoning for the reply YOU will write. Never drafts or posts replies; read-only GitHub access. Manual trigger only: /pr-triage [PR]. Skip on a single directed comment (\"just fix #3\").

Pre-flight checklist before a git push — commits-to-ship summary + branch-name / commit-hygiene / AI-attribution / staged-leftover / debug-artifact / hook-bypass / unresolved-secret checks; stops on any flag, never pushes past it. Auto-fires when about to run git push or on "push this" / "send it up". Skip on the scratch branch or "skip pre-push". /pre-push.

Prime the session with architectural rules for the current task — universal + language + domain + project tiers — the deep pass on top of the rule-prime hook's floor. Auto-fires on the first substantive task, after /clear, and on topic shift; asks before re-priming on drift. Manual trigger /prep. Not for trivial questions.

Write a session recap — an episodic record of what a work session was about, what was investigated, learned, completed, and what's next; Learned items can be promoted to memories via /capture. Use on /recap, \"wrap up\" / \"write session notes\" / \"recap this session\", or session-closure signals (\"done for today\", PR merged, deliberate pause). Mode A — never auto-fires.

Shared engine for the meta-review organs — orientation pass, NEW/CARRIED/RESOLVED baseline diffing, propose-confirm-commit-with-routing, and review-output-contract report rendering. Library-only — callable by checkpoint (and any future meta-review skill), not user-invoked. There is no /retrospect-core command. Stateless per call except for the report artefacts it persists.

Audit code in scope against your architectural rules; report drift (dead code, SoC, monolith, missing-pattern, principle, comment-drift, naming-quality) with propose-confirm per fix; route misses to /capture. Triggers: /review [path | --since <ref> | --vault], \"review this\", \"audit the architecture\", \"check for drift\". User-invoked only. Not a linter, not a security audit (use security-reviewer), not a PR review (use pr-review).

Manage the architectural-rule overlay — list / disable / enable / edit / sync / where across the shipped, company, user, and project tiers. Use when the user types /rules, says "override that rule", "disable this rule", "add a company rule", "show my active rules", "why is this rule applying", or wants to point at a company rules repo. Mode A only — never auto-fire.

Interview the user in detail using AskUserQuestion, then write a versioned spec under `.claude/specs/<slug>/`. Specs evolve via versioned files (v1, v2, …); INDEX.md tracks the active version per slug. Use when the user types /spec, or asks to spec out / write requirements for a non-trivial feature before building. Mode A only — never auto-fires.

Use when encountering a bug, test failure, or unexpected behavior that does not have an obvious one-line fix, before proposing fixes. Random fixes waste time and create new bugs — find root cause first.

Use when implementing a non-trivial feature or bugfix where regressions matter, before writing implementation code. Skip for throwaway prototypes, generated code, pure config changes, or one-line fixes — but only with the user's permission.

Append one dated line per shipped unit to the canonical CHANGELOG.md via propose-confirm — never auto-writes. Use on /update-changelog, \"log this ship\", or when a caller organ offers it on shipped work or a significant planning-artifact change. One line per unit, not per commit. Mode A — never auto-fires.

Regenerate the project codemap at .claude/codemap.md — produce an LLM-facing architecture document with prose overview, per-module roles, detected conventions, hub files, and structured file/class graphs for downstream visualization. Use when the user types /update-codemap or asks to refresh / create / update the codemap for the current project.

Use when starting feature work that needs isolation from the current workspace, or before executing a non-trivial implementation plan that risks polluting the working tree. Creates an isolated git worktree with directory-selection priority and gitignore safety verification.

Render .claude/codemap.md into a UML-heavy document — layered module map, per-module class/file graphs, cross-module relations, call graphs, sequence diagrams — written to .claude/codemap.diagrams.md and the vault. Consumes the existing codemap, never rescans. Use on /visualise-codemap or \"diagram the codemap\". Mode A — never auto-fires.

Report a slug's position in the scope chain (envision→spec→draft-plan→blueprint→execute→archive) as deterministic structured state — a per-stage missing/present/stale/done table, a single next-action line, and the load-bearing stale-spec-pin check. Read-only, mostly filesystem-derived (no model call for the structural state). Use when the user types /work-state <slug> or /status <slug>, asks "where is <slug> in the chain", "what's the next step for <slug>", "is <slug>'s plan stale". Mode A — ...

Session-terminus driver — one /wrap runs the full closing ceremony by driving the existing organs, from coordination teardown through changelog and build progress. Auto-runs only the reversible spine; every canonical write gated accept/edit/reject. Use on /wrap or \"wrap up the session\". Mode A — never auto-fires.

Author a quality test suite for EXISTING code — convention-matching, test plan confirmed before any writing, pins current behavior but flags suspicious-as-bug. Use on /write-tests or \"write / add tests for existing code\". Mode A — never auto-fires.