ClaudeCode Cheat Sheet

Personal skill index — auto-updated when new skills are created. Last updated: 2026-05-04. File: ~/.claude/skills.html

Daily drivers

Skills you'll reach for frequently — strategic thinking, debate, learning. Type the slash command or speak any trigger phrase.

Skill Description Auto-triggers Where it works
/multi-model-chat Cross-provider debate: Claude+GPT+Gemini agents with different personas debate a problem over multiple rounds; Claude (orchestrator) synthesizes the final report. "multi-model chat" "cross-provider debate" "Claude vs GPT vs Gemini" "have all models discuss" Any project
/agent-chatrooms Spawns N Claude sub-agents into a shared conversation that debate, disagree, and converge via shared chat.json. Useful for architecture/design trade-offs. "chatroom" "agent debate" "have agents discuss" "let them debate" Any project
/video-to-action Uses Gemini as a video passthrough to extract step-by-step actionable instructions from YouTube tutorials, demos, and walkthroughs. "watch this video" "extract steps from video" "learn from this video" "analyze this YouTube" Any project
/reverse-prompting Before starting, asks Manuel up to 5 clarifying questions that would most change the approach. Skips questions already answered by CLAUDE.md / HANDOFF.md / seed/ / contracts/. COO-style only (no data-hygiene questions per GLB29). "reverse prompt" "ask me questions first" "clarify before starting" "surface assumptions" "ask before building" Any project
/reverse-prompting-contract Two-phase: ask up to 5 questions → turn answers into a 4-section delivery contract (GOAL / CONSTRAINTS / FORMAT / FAILURE) → implement with self-verification. Auto-folds GLB1/18/19/20/29/30 FAILURE clauses where relevant. Offers Tier B storage if recurring. "reverse prompt with contract" "questions then contract" "ask me first then write a contract" "spec this with questions first" Any project
/delivery-contracts Defines success, constraints, output format, AND explicit failure conditions in a single structured contract. Treated as spec, not suggestion. Auto-applies on stored contracts in <project>/contracts/. "delivery contract" "define success criteria" "failure conditions" "spec this out" "what does done look like" Any project
/multi-agent-chrome Spins up 1-8 stealth Chrome browsers (default 4, max 8 — hard cap on M1 Max + 64 GB) on the fingerprint-chromium binary. N=1 = single-browser stealth for big solo scrapes. N≥2 = parallel research, comparison shopping, form-filling. Includes a built-in skeptic-review loop (Gemini 2.5 Pro reviews findings; orchestrator can re-run flagged agents up to 2× for max 3 iterations). Asks Manuel for N at runtime; workspace under active project per LR13. "multi-agent chrome" "parallel browsers" "spin up N browsers" "research X across many sites in parallel" "bulk browser task" Any project

Situational tools

Skills you'll use occasionally — when starting a new project, scaffolding a write pipeline, polling agents, or pinning down a spec. Not daily, but powerful when the moment fits.

Skill Description Auto-triggers Where it works
/init-project Scaffolds a new project with the canonical folder layout (apps/, scripts/, docs/, data/, debates/, logs/, secrets/, .claude/) plus CLAUDE.md, HANDOFF.md, .gitignore templates per LR31. "new project" "let's start a [name] project" "init project" empty dir under /ClaudeCode/ with no CLAUDE.md New project init
/rwd-scaffold Generates the Reversible-Writes Discipline pipeline (audit_log + before/after snapshots + undo CLI + drift detector + dry-run defaults) per LR18. Mirrored from wynwd_atlas/backfill. "set up audit logging for [project]" "new write pipeline to [system]" "add RWD to [project]" Projects writing to CRM/ATS/SaaS
/init-intelligence-project Creates the persistent intelligence architecture: seed/manuel_explanations/ (Tier 0) + intelligence/ + auto-load @-imports in CLAUDE.md per LR30. Mirrored from wynwd_wiki. "set up knowledge system for [project]" "init intelligence layer" starting a data-ingestion project Knowledge projects (ingesting Manuel's business data)
/stochastic-multi-agent-consensus N agents with framing variations independently analyze a problem; aggregates results into consensus / divergence / outliers. Filters hallucinations, surfaces creative ideas. "consensus" "poll agents" "what do 10 agents think" "get multiple opinions" Any project
/create-project-helper Build a project-specific HTML helper at <project-root>/PROJECT_HELPER.html capturing custom subagents, Tier B contracts, Makefile targets, project rules, reference docs, critical artifacts, and gotchas. Designed for the cold-start session after weeks/months away from a project. Mirrors this skills.html visual style but project-scoped. Manual invocation only — does NOT auto-fire. Canonical reference: wynwd_atlas/PROJECT_HELPER.html. "create project helper" "build a project helper" "make a helper page for this project" /create-project-helper manual invocation only — opt-in per project Any project
/create-googlecloud-project Create a new Google Cloud project fully configured for Cloud Run / Build / Artifact Registry / Vertex AI deploys, with the claude-code-agent SA granted appropriate roles + Compute SA grants + APIs enabled. Captures the full setup sequence learned 2026-05-04 during the wynwood-atlas + Ideas project setups so future projects don't require re-discovery of the role chain. Per LR49 asks for budget cap before activation. "create a GCP project called X" "new GCP project" "set up a Google Cloud project for [name]" /create-googlecloud-project Any project

Background helpers

Skills you almost never invoke manually — Claude detects the right moment (corrections, end-of-task, complex multi-step starts) and fires them automatically. Listed here for transparency.

Skill Description Auto-triggers Where it works
/handoff-refresh Updates the active project's HANDOFF.md (Status / Active Rules / Next Steps / Recent Decisions / Open Questions) for cross-session continuity per LR14. proactively at end of long task chunks "update handoff" "refresh handoff" "compact session" "before /compact" Any project with HANDOFF.md
/capture-correction Auto-captures user corrections to project CLAUDE.md (and global if cross-project) without asking. Implements LR17 / BRES Layer 2. implicit, when Manuel corrects Claude "you forgot X" "I told you Y" "from now on always Z" Any project
/check-rules Lists ALL active rules for the current session — global rules from ~/.claude/CLAUDE.md + project-specific rules from the active project's CLAUDE.md. proactively at start of complex multi-step tasks "what rules apply" "show me your rules" "are you following my instructions" Any project
/lower-model-work At start of any multi-step task (≥2 subtasks), analyzes whether parts can be safely delegated to cheaper models (Haiku 4.5 ~15–20× cheaper, Sonnet 4.6 ~3–5× cheaper) via Agent calls with model: override. Orchestrator (latest Opus) handles synthesis. 95% confidence floor — when in doubt, Opus. Plan shown to Manuel before executing. Per LR42 / GLB42. auto-fires at start of multi-step tasks "lower model work" "delegate to cheaper" "what can Haiku do here" "split this by model" Any project