Skip to content

Latest commit

 

History

History
179 lines (125 loc) · 10.2 KB

File metadata and controls

179 lines (125 loc) · 10.2 KB
layout default
title Skill Catalog
parent Part VI — Reference
nav_order 4

Skill Catalog

Skills shipped by the install.sh installer, grouped by tier. The authoritative source for what each tier contains is best-practices/manifest.json — this chapter is a human-readable projection of that file.

Install a tier with:

# Core (3 skills, 8 rules, 1 hook)
curl -sL https://raw.githubusercontent.com/ytrofr/claude-code-guide/master/install.sh | bash

# Recommended (16 skills, 30 rules, 7 hooks)
./install.sh --recommended

# Full (44 skills, 55+ rules, 12 hooks, governance scaffolding)
./install.sh --full

Skills install to ~/.claude/skills/<name>/SKILL.md and are global across every project on the machine — per-user, not per-project. One install; every Claude Code session across every repo picks them up.


Core tier (3 skills)

Newcomer-friendly starter. Covers the two highest-ROI loops (verify, session-start) plus a routing tree for debugging.

Skill Description
verify Verify app after changes — auto-detects scope, runs health checks, tests, code quality. Use before commit or deploy.
session-start Initialize session following Anthropic best practices (checkpoint discovery, git status, memory context).
troubleshooting-decision-tree Route troubleshooting patterns by issue type (connection errors, integration failures, behavioral mismatches).

Recommended tier — adds 13 (total 16)

Working developer set. Adds debugging workflows, planning rigor, TDD discipline, memory hygiene, MCP usage patterns, and documentation.

Workflow (5)

Skill Description
tdd Enforce Test-Driven Development: RED (failing test) → GREEN (min code to pass) → REFACTOR.
plan-checklist Generate plan templates with 13 mandatory sections, KPI dashboard, pre-validation probe, modularity gates.
session-end End session per Anthropic best practices — invoke when approaching context limit or wrapping up.
retrospective Create a skill from session learnings (guided). Use after major milestones or when discovering a reusable pattern.
document Document session work plus 7-check pattern analysis at 3 levels (machine / project / branch). Suggests rules, skills, memory notes. Stops for user confirmation.

Memory (3)

Skill Description
memory-defrag Defragment and reorganize agent memory — split bloated files, merge duplicates, remove stale info. Run weekly.
memory-notes Write well-structured Basic Memory notes with frontmatter, observations, semantic categories, and wiki-links.
memory-search-patterns Progressive-disclosure Basic Memory search: 3-layer workflow (index → preview → fetch) for 10× token savings.

MCP / tools (3)

Skill Description
mcp-usage-patterns Select the right MCP tool per task (PostgreSQL, GitHub, Memory, Perplexity).
perplexity-workflow Perplexity API usage with monthly budget ceiling, cache-first memory enforcement, tool selection.
playwright-mcp Automate browsers with Playwright MCP — testing, screenshots, form filling, data extraction.

Deploy / ops (2)

Skill Description
canary Post-deploy canary checks — screenshot capture, console error detection, network error scanning, baseline comparison.
doctor-workflow Interpret and act on doctor script results. Use when doctor fails, before deploys, or when diagnosing project health.

Full tier — adds 28 (total 44)

Power-user set. Adds governance scaffolding, AI/LLM patterns (Gemini, multi-agent, resilience), architecture patterns (observability, auth, data cascade), and cross-project coordination.

Governance (8)

Context-budget management, self-telemetry, hygiene scanning, rollback.

Skill Description
context-audit Audit context budget health — measures char costs, detects bloat, flags duplicates, identifies optimization targets.
context-optimization Optimize full context stack — audit, classify, condense, relocate, verify. Use when budget exceeds targets.
context-governance Orchestrate context governance — audit state, decide action by status, dispatch to /context-audit, /context-optimization, or rollback.
context-governance-rollback Roll back governance changes (rules, skills, scanners) to a tagged baseline.
audit-stack Audit Claude Code stack freshness against latest official docs and Hub data. Use after CC upgrades.
skill-metrics Display skill activation metrics from pre-prompt logging. Use when checking which skills are being triggered.
weekly-review Generate a CC weekly self-telemetry review — prompts, tool calls, MCP latency, skills, subagent dispatches, session KPIs.
gitignore-anchor-audit Audit .gitignore files for unanchored top-level dirs that recursively shadow nested paths.

AI patterns (9)

Gemini/Claude/multi-model patterns, orchestration, resilience, evaluation.

Skill Description
multi-agent-patterns Advanced multi-agent orchestration patterns: streaming parity, sticky modes, correction flow, transparent delegation, knowledge propagation.
gemini-patterns Gemini production patterns: JSON mode coercion, thinking+JSON conflict, empty-response nudge, non-determinism band, instruction position.
llm-resilience-patterns Circuit breaker, multi-provider fallback, per-call retry, instruction-corruption recovery, context overflow, budget enforcement.
multi-model-llm-routing Route between LLM models by task type, cost, and quality. Implements primary/fallback patterns.
mcp-tool-evaluation Evaluate whether new tools should be MCP, Skill, or Agent. Prevents wrong-tool-type choices.
prompt-optimization-methodology Compress LLM prompts systematically with accuracy validation. Reduce tokens without regressing quality.
llm-finish-reason-probe Log Gemini finish_reason and Claude stop_reason non-normal values for observability. Catches silent mid-phrase truncations.
a-b-c-variant-experiment Run a 3-variant A/B/C probe to determine whether a prompt change improves behavior or hits a ceiling. ~45 min.
anthropic-eval-best-practices Apply Anthropic's 6 eval best practices: pass^k consistency, transient-failure handling, flaky-test recovery.

Architecture patterns (9)

Observability, cron safety, liveness, entity handling, refactoring, fetching, auth, invariants.

Skill Description
observability-first-pattern Break the hypothesis-debate loop after 3+ failed fixes at the same layer. Build a per-request persistent tracer before the 4th attempt.
cron-lock-patterns pg_try_advisory_lock async context manager + single-instance file lock + key-selection rationale + jitter formula. Prevents duplicate cron fires.
service-liveness-monitoring Two-signal service liveness: indirect (work piling up?) + direct (loop beat recorded?). Detects stuck loops.
entity-resolution-pattern Resolve entity names to IDs in code, not LLM chains. Fixes multi-step lookup failures.
god-file-extraction-methodology Split god files (1500+ lines) safely into modular components with zero regressions.
parallel-rate-limited-fetcher Fetch from multiple APIs in parallel with rate-limit checking, circuit breakers, and graceful failure isolation.
signed-url-auth Signed-URL auth for session-less users (Telegram/WhatsApp/SMS/voice/CLI). JWT signing + verification + jti replay registry.
ast-contract-tests Encode structural invariants as AST-walking pytests — required kwargs, decorators, call patterns.
write-side-cascade Cascade pointer consistency at mutation time, not read time. Feature-toggled, idempotent backfill, observability.

Cross-project coordination (2)

For machines running multiple Claude Code projects that need to share knowledge or coordinate.

Skill Description
ai-dna Maintain shared AI knowledge across multi-project setups — rules, skills, memory notes, pattern promotion, staleness checks, monthly health audits.
inter-agent Coordinate with Claude in another project via shared bus. /talk-new, /talk-send, /talk-stream, /talk-resolve — no more terminal copy-paste.

Invoking skills

Once installed, skills are available in three ways:

1. Slash invocation (explicit)

/verify
/tdd
/plan-checklist

Always-available at the prompt. Works for every skill unless user-invocable: false is set in its frontmatter.

2. Auto-invocation (model-driven)

Claude reads each skill's description: field on session start and can invoke the skill on its own when the description matches the task at hand. Well-written descriptions use Use when <trigger> phrasing — this is what matches the incoming request.

Opt out per-skill with disable-model-invocation: true — keep it user-invocable, but prevent Claude from firing it unprompted. Useful for side-effect skills like deploy, commit, /canary.

3. Background knowledge (non-invocable)

Set user-invocable: false to hide a skill from the slash menu entirely. Claude loads its description and can use it as reference knowledge, but the user can't call it directly. Useful for "this is how we do X in this codebase" domain patterns.

Description-field budget

Skill description: fields compete for ~2 percent of the context window. Keep each under 250 chars, with a clear trigger clause. Claude Code 2.1.105 raised the cap from 250 → 1536 chars, but shorter is still better — every byte counts against the budget.

Run /skills to see the full list and (in 2.1.111+) toggle sort-by-token-cost.


See also

  • [Part III, Ch. 04 — Skills authoring]({{ site.baseurl }}/guide/part3-extension/04-skills-authoring/) — how to write your own skill, frontmatter reference
  • [Part III, Ch. 05 — Skills maintenance]({{ site.baseurl }}/guide/part3-extension/05-skills-maintenance/) — lifecycle, promotion, archival, SLAs
  • [Part I, Ch. 05 — Installation]({{ site.baseurl }}/guide/part1-foundations/05-installation/) — the install.sh tiers end-to-end
  • best-practices/manifest.json — the authoritative JSON this chapter projects