auto-claude-skills

auto-claude-skills

A Claude Code plugin that enables agentic PDLC (agent-assisted product development lifecycle): it helps Claude move from discovery through learning with the right skills, guardrails, and tool context for each phase.

Instead of relying on memory and ad hoc prompting, the plugin gives Claude a stronger operating model across the AI-enabled PDLC. It routes the right skill, surfaces the next step, guards planning/review/verification, and — in spec-driven mode — commits design intent to versioned artifacts so work stays aligned across sessions and teammates.

Why This Exists

AI accelerates every phase of the lifecycle, but speed exposes three problems:

• Skill sprawl. Dozens of installed skills, and remembering which to invoke is the user's job. Useful skills go unused.
• Skipped discipline. Review, failing tests, ship verification, and as-built docs are easy to skip when the model feels confident. Small features ship without review; completion claims go unverified.
• Lost intent. Design decisions live in chat transcripts, not committed specs. When work resumes in a later session or by a teammate, the why is gone and AI-assisted work drifts from intent.

The plugin addresses these by routing skills on intent, gating the skipped transitions, and committing design intent to versioned specs when spec-driven mode is on.

What It Does

The plugin turns a loose collection of skills, hooks, and integrations into a phase-aware delivery system.

• Prompt routing by intent. Each prompt is scored against trigger patterns and matched to relevant skills. Prompts that match nothing produce no output — no noise on non-development work.

• Phase-aware PDLC orchestration. The plugin infers a likely phase and composes the relevant skills, sequences, and hints across the delivery lifecycle:

  Phase	Baseline path	Optional enhancers
  DISCOVER	Pulls Jira/Confluence context, synthesizes a discovery brief before design	Context-stack lookups for prior decisions and specs
  DESIGN	Activates brainstorming, explores requirements before coding	design-debate, feature-dev, frontend-design when installed and appropriate
  PLAN	Structures implementation into discrete tasks with dependency ordering	Context lookups for APIs, blast radius, and past decisions
  IMPLEMENT	Drives the execution-plan/TDD workflow; halts if implementation is written before a failing test	agent-team-execution for independent tasks, security guidance when installed
  REVIEW	Routes into requesting-code-review plus security scanning	agent-team-review, PR toolkit specialists, runtime validation, and drift checks when applicable
  SHIP	Routes through verification, openspec-ship, and branch completion	Commit/PR automation and memory consolidation when installed
  LEARN	Routes to outcome-review for metrics and follow-up work	PostHog and Jira follow-through when installed
  DEBUG	Overrides the current phase with structured debugging or incident investigation	incident-analysis, alert-hygiene, and observability context

• Guardrails that support agentic work. Phase-specific red flags, required sequences, and selective hard gates for high-risk cases like unverified pushes — most enforcement is in-session guidance, not blocking.

• Optional enrichment from companion tools. When integrations like Jira, GitHub, Context7, Serena, or GCP Observability are installed, the routing engine injects phase-appropriate context to use them.

• Versioned intent, not ephemeral chat. In spec-driven mode, design proposals and as-built specs live in openspec/changes/. Hypothesis artifacts, plan files, and team checkpoints are saved to disk so later skills, next sessions, and teammates pick up from the same intent.

Example Prompts

"design a secure frontend component" → DESIGN phase. Activates brainstorming to explore requirements before coding. When design-debate is installed and the prompt warrants multi-agent exploration, it can also enter structured debate; frontend-design adds UI-specific patterns when installed. Companion tools query library docs for relevant API constraints.

"debug this login bug" → DEBUG phase. Activates systematic-debugging (structured root-cause analysis). TDD is injected as a mandatory parallel — reproduce with a failing test before fixing. If GCP Observability is installed, context hints toward runtime logs.

"investigate this 502 spike" → DEBUG phase. Activates incident-analysis, which inventories available infrastructure, classifies the failure against 7 bundled playbooks (bad-release, dependency failure, infra failure, node exhaustion, and others), runs a structured investigation with completeness gates, and generates a canonical postmortem with impact quantification.