oh-my-claude

Maximum-autonomy professional work mode. Use for non-trivial coding, writing, research, planning, or mixed tasks that should be carried through the right specialist workflow with minimal hand-holding.

Deactivate ultrawork mode for the addressed session. Clears its workflow and enforcement state without affecting concurrent Claude sessions.

Adaptive project evaluation that maps the task's actual risks, assembles the smallest useful specialist team, and fills material coverage gaps before synthesis.

Guided oh-my-claude configuration via multi-choice prompts — first-time setup walkthrough, post-update review, or ad-hoc inspect/change. Auto-detects mode from `~/.claude/oh-my-claude.conf`. Use when the user asks to install, set up, configure, update, walk through settings, change flags, turn the watchdog or bias-defense on/off, switch model tier, or any phrase like "help me install/update/configure oh-my-claude", "show my settings", "what flags do I have on".

List all available oh-my-claude skills with descriptions and when to use each one.

Audit whether a project's tests still earn their runtime and maintenance cost. Maps behavior contracts to owners, identifies redundant or obsolete proof, and recommends KEEP, EXTEND, MERGE, REPLACE, DELETE, or ADD with evidence. Use --apply to implement proven portfolio changes.

Render a markdown digest of recent harness activity — sessions, gate fires, top reviewers, classifier misfires, Serendipity catches, and finding/wave outcomes — over a chosen time window. Use to answer "is the harness actually helping me?". Pass `--share` for a privacy-safe shareable card (numbers + distributions only; no prompt text or free-text reasons).

Show the current ULW session state — workflow mode, domain, intent, counters, and flags. Use for debugging or inspecting what the hooks decided. Pass `summary` for a compact end-of-session recap, `classifier` for intent-classifier telemetry, or `explain` for a per-flag rationale.

Show where this prompt's (or session's) wall time and tokens went — polished end-of-turn card with a stacked-segment top bar (`█` agents · `▒` tools · `░` idle), per-bucket rows, input/output/cache and agent-token share, and a closing insight. No argument (or `current`) is the active session; pass `last-prompt` to slice the most recently finalized prompt, `last` for the most recent session, `week` / `month` / `all` for cross-session rollups. Use to answer "why did that prompt take so long?" and "which agents are the most expensive in my workflow?". The same card renders automatically at the end of every Stop hook above the 5s noise floor (`time_card_min_seconds`, env `OMC_TIME_CARD_MIN_SECONDS`) — invoke this skill when you want it on demand or want to slice a different window.

Create distinctive, production-grade frontend interfaces with high design quality. Use when building web components, pages, or applications where visual craft matters. Avoids generic AI aesthetics by establishing visual direction before code.

Guided onboarding demo that shows oh-my-claude's quality gates in action. Use after first install to see how stop-guard blocking, review loops, and verification work.

Canonical user-facing name for the maximum-autonomy professional workflow (v1.31.0 naming consolidation). Use when you want Claude Code to auto-detect the task domain and run the right specialist path. Legacy aliases `/autowork`, `/ultrawork`, `sisyphus` remain wired for muscle memory.

Scientific data-analysis workflow — explore, fit with honest uncertainties, and produce publication-quality figures with full provenance. Use for experimental/measurement data (spectra, traces, curves, instrument output) where the output must survive peer review, not just run.

Academic manuscript workflow — structure, verified citations, drafting, and rigor critique for papers, theses, proposals, and referee responses. Sequences evidence before prose so every citation is registry-verified before it is written.

Verified academic literature review — finds papers and returns registry-verified sources (Crossref/OpenAlex/Semantic Scholar/arXiv) with claim-support anchors and honestly-marked gaps. Never cites from model memory.

Set a persistent, user-declared GOAL that the harness drives toward relentlessly across turns — re-anchoring it and blocking premature Stops until the goal is verifiably achieved (fresh completeness audit + attestation) or it hits a no-progress wall it cannot pass alone. The user-facing port of Codex CLI's /goal. Lifecycle subcommands cover set / status / pause / resume / clear / done. Use when the user wants the agent to keep working toward one durable objective instead of stopping after a normal turn.

One-shot oh-my-claude install health check — verifies the installed tree at ~/.claude (core files, CLAUDE.md @-includes, hook wiring in settings.json, state-root writability) from any directory, no source clone needed. Use when the harness feels off, gates stopped firing, or after an update.

Reviews and guides game development code for Unity (C#), Godot (GDScript/C#), and web engines (Phaser, Babylon, PixiJS, Three.js) — engine-idiomatic patterns, frame-budget performance, and a run-capture-evaluate-fix verification loop. Use when reading, writing, or reviewing game code, game logic, shaders, ECS systems, scene trees, update loops, or game-engine projects.

Comprehensively reviews SwiftUI code for best practices on modern APIs, maintainability, and performance. Use when reading, writing, or reviewing SwiftUI projects.

Signal an operational pause for a missing input or hard external blocker — NOT for technical-judgment decisions (under ULW v1.40.0 the agent owns those). Use when credentials are required, a rate limit hit, infra is dead, or the next step needs user-supplied input the agent cannot fabricate. Distinct from /ulw-skip (gate bypass) and /mark-deferred (legacy soft-defer, disabled under ULW execution).

Correct a misclassified ULW turn (wrong intent or domain). Records the correction to cross-session telemetry so future tuning can learn from human-labeled misfires, and updates the active session's intent/domain when a corrected value can be parsed from the reason. Use when the classifier got it wrong and you want to redirect rather than rephrase.

Mark all pending discovered-scope findings as deferred with a one-line reason that names a concrete WHY. Use when the discovered-scope gate flags advisory-specialist findings you have consciously decided cannot ship in this session — typically because they require a separate specialist, a database migration, a stakeholder decision, or context not yet loaded. Each deferred row is timestamped and the reason is recorded so the gate stops blocking and `/ulw-report` can audit deferral patterns later. **Prefer wave-append over deferral when the finding is same-surface to your active work** — see *When NOT to use* below.

Show CHANGELOG entries between your installed version and the source repo's HEAD. Use after `git pull` (or to see what's new in an unreleased install) to discover features that landed since your last install. Runs the bundled `show-whats-new.sh` script which reads `installed_version` from `~/.claude/oh-my-claude.conf`, walks `${repo_path}/CHANGELOG.md`, and renders the delta in a readable format.

Bootstrap or refresh concise repo-specific instruction files for the current project so future Claude sessions follow the codebase's real conventions. Also handles deep-refresh audits of skill / agent definition files (`SKILL.md`, `agents/*.md`) when their bodies have drifted from the backing scripts and manual one-off patches no longer scale.

Generate 3-5 alternative framings of a problem BEFORE committing to one. Use upstream of `quality-planner` when the task admits multiple credible approaches and choosing the wrong shape would cost rework. Distinct from prometheus (interview to clarify WHAT to build), metis (stress-test a draft plan), oracle (debug second opinion), and abstraction-critic (critique an existing artifact's paradigm fit). Divergent-framer expands the option space; the others narrow it.

Resume a /ulw task that was killed by a Claude Code rate-limit StopFailure. Atomically claims the most relevant unclaimed `resume_request.json` for the current cwd (or matching project_key) and surfaces the prior session's objective + last user prompt so /ulw can replay the work as-if-uninterrupted. Pairs with the SessionStart resume-hint hook (Wave 1) and the headless watchdog (Wave 3) — all three consume the same artifact via the canonical `claim-resume-request.sh` helper. Use when the SessionStart hint surfaces a pending resume, when a continuation prompt arrives in a session with a claimable artifact, or when the user explicitly asks to resume the prior /ulw work.

Audit the user-scope auto-memory directory for the current project — classify each MEMORY.md entry as load-bearing, archival, superseded, or drifted, and propose rollup moves. Use when MEMORY.md feels noisy, when version-snapshot files have accumulated (e.g. multiple `project_v*_shipped.md`), or when the auto-memory drift hint has fired at session start. Read-only — never deletes or moves files.

Skip the current quality gate block once with a logged reason. Use when a gate is blocking but you're confident the work is complete. The skip is recorded for threshold tuning.

Official-docs and reference-implementation research mode. Use when the task depends on external APIs, frameworks, library conventions, or authoritative examples.

Stress-test a draft plan or current approach before execution. Use when the work is risky, ambiguous, or easy to get subtly wrong.

Deep debugging and architecture second-opinion mode. Use when you are stuck, uncertain, or deciding between multiple plausible technical paths.

Build a decision-complete implementation plan before coding. Use when the task is non-trivial, risky, or ambiguous.

Interview-first planning for broad, ambiguous, or product-shaped work. Use this when you want Claude to extract intent intelligently before implementation.

Gather targeted repository or API context before implementation when the correct approach is not yet obvious.

Perform a findings-first review of the current worktree or a focused area before finalizing work.