Ai Skills

ai-skills — Claude Code Plugin

A reusable team-of-agents plugin for the full software development lifecycle. Drives feature design, implementation, bugfixing, environment analysis, and refactoring through coordinated subagents, with built-in iteration loops (RALF), layered memory, and systematic eval.

Project-agnostic by design. Operations live in this plugin; project-specific context (brand, conventions, terminology) lives in the target repo's CLAUDE.md / AGENTS.md / marketing/MARKETING.md and is read at runtime.

Status — v0.3.10

Current release. The plugin is the canonical Claude Code delivery format for this repository (the legacy .claude/ mirror was retired in v0.2.0). Codex and Windsurf packages remain alongside for those runtimes; parity is enforced between Codex and Windsurf only.

For existing ~/.claude/ install users from before v0.2.0, switch to:

claude --plugin-dir /path/to/ai-skills/plugin

After cleaning up the old install: rm -rf ~/.claude/agents ~/.claude/skills ~/.claude/rules ~/.claude/hooks ~/.claude/settings.json (preserve any of your own personal ~/.claude/ content first).

The local validator (python3 plugin/dev/validate.py) passes 23 checks against the live tree.

Component	Count	Notes
Hooks	18	Across 13 lifecycle events; includes ralph-iter-meter (v0.1.6) and subagent-depth-guard (v0.1.7)
Agents	26	22 normalized roles + 4 orchestrators
Rules	12	Security, memory discipline, RALF budget, untrusted-content wrapping, etc.
Skills	73	36 user-invocable (32 with context: fork + 4 main-thread orchestrators); rest are background knowledge loaded as context
JSON schemas	2	G7 spawn payload + return contract
Memory templates	6	In memory/templates/ (pii-patterns.txt lives at hooks/scripts/pii-patterns.txt so _lib.py can load it)
Eval rubrics	45	17 base + 4 meta-tools + 24 per-skill workflow rubrics
Calibration samples	270	6 per rubric (3 good + 3 bad) × 45
Output styles	2	concise-pr, design-pack
User docs	15	1 getting-started + 10 workflows + 4 concepts
Workflows (/<command>)	10	Primary user-invocable workflows
Companion skills	9	Named user-invocable companions
userConfig knobs	13	Declarative configuration

Distribution: local install via claude --plugin-dir <path> (Anthropic-canonical for development). Marketplace install via /plugin marketplace add <repo-url> once published.

Install

Local development (recommended)

Per official Anthropic docs, the canonical way to use a plugin from a local path is the --plugin-dir flag:

# Start Claude Code with the plugin loaded for this session
claude --plugin-dir /path/to/ai-skills/plugin

# After editing plugin files, reload without restarting:
/reload-plugins

All 32 user-invocable skills appear in /help under the ai-skills: namespace, e.g. /ai-skills:feature-design, /ai-skills:develop, /ai-skills:plugin-doctor. Plugin namespacing is automatic per Anthropic spec — prevents conflicts when multiple plugins ship same-named skills.

Future: marketplace distribution (after GitHub publish)

Once the plugin is pushed to GitHub, install via the official marketplace flow:

/plugin marketplace add alex-voloshin-dev/ai-skills
/plugin install ai-skills@ai-skills

The repo's .claude-plugin/marketplace.json already declares the registry. Local marketplace install (/plugin marketplace add <local-path>) is supported by Claude Code in principle but currently brittle for same-host development on v2.1.x — use --plugin-dir instead.

Workflows (10 user-invocable)

Slash command	What it does
/feature-design	Multi-agent design pack from a 1-3 sentence idea (PRD + ARD + UX + impl plan)
/develop	Full SDLC implementation from a design pack (DEVELOP → REVIEW → QA pipeline)
/bugfix	Triage, diagnose, fix, ship — with reproduction-test RALF loop
/env-analyze	Standalone diagnostic for Docker / K8s / CI environments
/refactor	Plan and execute refactor across N files with safety nets
/migrate	Schema/library/version migration with rollback plan
/spike	Time-boxed exploration with go/no-go writeup
/security-audit	OWASP Web Top 10 + GenAI Top 10 audit + remediation plan
/docs-pack	Generate user-facing docs (README, API ref, runbook)
/ai-skills-init	Bootstrap a target repo to be ai-skills-aware

Companion skills (9)

/ralph (power-user RALF entry) · /eval (skill/agent evaluator) · /plugin-doctor (self-diagnostic) · /memory-init · /memory-recall · /learnings-write · /context-load · /subagent-spawn · /plugin-skill-create

What's inside today (v0.3.10)

• 73 skills covering the full SDLC plus marketing and content. 36 are user-invocable: 32 with context: fork plus 4 main-thread orchestrators (develop, feature-design, bugfix, team-bugfix) that retain Agent-spawn capability. The remaining 37 are knowledge skills (disable-model-invocation: true) loaded as context by workflows or agents — Single-Responsibility split per the May 2026 /plugin-skill-audit refactor (split fat workflows, extract per-stack tooling tables, thin engineer agents via the prompt-engineer pattern).
• 26 specialized agents — cloud architect, security engineer, all major language engineers, content/marketing roles, 4 orchestrators.
• 18 hooks across 13 lifecycle events — security guardrails, untrusted-content wrapping, session memory flush, RALF loop control, per-iteration token meter (v0.1.6), subagent depth guard (v0.1.7), .committed/ allowlist enforcement.
• 12 rules — security, memory discipline, RALF budget, untrusted-content wrapping, etc.
• 45 eval rubrics + 270 calibration samples + Tier 1 linter + Tier 2 judge-calibration smoke for systematic regression detection. Tier 2 added in v0.1.4 and requires ANTHROPIC_API_KEY. Tier 3 is planned but not yet shipped (runner returns error code 3 if invoked).
• G1/G2 attack-surface validation (v0.1.5) — 6 indirect-prompt-injection fixtures plus a structural runner that confirms the <untrusted_content> envelope wraps attacker-planted instructions across 5 attack vectors (poisoned CLAUDE.md, malicious env logs, poisoned learnings, bash role-switch, poisoned PRD). 1 fixture documents the sub-200-token wrap-skip design tradeoff. Optional behavioral mode round-trips wrapped payloads through Haiku to verify no compliance escape.
• Per-iteration RALF token measurement (v0.1.6) — ralph-iter-meter.py PostToolUse hook estimates tokens per tool call (chars/4) and accumulates while a RALF run is active. ralph-stop.py consumes the per-iter accumulator on each Stop intercept, persists iter-NNN/tokens.json, and fires a runaway warning when a single iteration exceeds 3× fair share (token_budget / max_iterations). Closes the v0.1 gap where session-aggregate token caps only fired inside Tier 3 eval runs.
• Defensive subagent depth-guard (v0.1.7) — subagent-depth-guard.py SubagentStart hook walks the parent_trace_id chain in G7 spawn payloads, computes spawn depth, and blocks at depth > userConfig.subagent_max_depth (default 3). Logs every start/stop/rejected event to .ai-skills-memory/sessions/<sid>/spawn-chain.jsonl. Anthropic's runtime enforces depth=1 max; this is the defensive backstop in case orchestration accidentally bypasses or future versions add Task to additional agents.
• 2 output styles — concise-pr for PR descriptions, design-pack for /feature-design artefacts.
• 13-knob declarative config + 0 external dependencies (fully standalone).

Learn more

• docs/getting-started.md — 30-min tutorial (install + first feature design)
• docs/workflows/ — one user doc per slash command (10 files)
• docs/concepts/memory.md — 6-layer memory model summary
• docs/concepts/eval.md — eval framework summary
• docs/concepts/ralf.md — RALF iteration loop summary
• docs/concepts/skill-frontmatter-extensions.md — the 4 Claude Code-specific frontmatter fields and how they relate to the agentskills.io spec
• CHANGELOG.md — version history per batch

License

MIT