claude-proj-blueprint

Legacy code archaeology — multi-phase discovery that reads the codebase, git history, and optionally Jira to reconstruct WHY code exists. Produces discovery notes, ADR candidates, and PRD candidates for human review. Activated when user says 'descobrir módulo', 'arqueologia de código', 'entender legado', 'contexto do módulo', 'por que esse código', 'legacy context', '/discover', or when starting work on an unfamiliar legacy module.

Legacy code archaeology and context bootstrap. Activated when the user mentions "legacy code", "código legado", "sistema antigo", "código sem documentação", "discover module", "entender módulo", "arqueologia de código", "código que ninguém entende", or runs /discover. Populates the empty upper layers of tiered lookup (memory and docs) before context-engineering techniques can be effective. Read-only for source code: outputs discovery notes, ADR candidates, and suggested intent markers. Remote discovery supports GitHub (via gh CLI) and Azure DevOps (via az CLI plus REST).

Context engineering — proactive context window management. Activated when working on large tasks, long sessions, multi-file changes, or when context is getting bloated. Keywords: "context", "too much code", "split this task", "focus", "scope down", "token budget", "keep it small". Complements the reactive context-guard hook with proactive techniques.

Reconcile a Claude Design handoff bundle (PROMPT.md) with the project's PRD and conventions. Activated when the PRD references a Claude Design handoff, when a PROMPT.md exists in docs/design/, or when the user mentions "Claude Design", "design handoff", or "PROMPT.md".

Pattern learner. Activated when the user says "learn from this", "what patterns", "improve skills", "retrospective", "session review", or wants to extract reusable patterns from recent work to improve the project's skills and conventions.

Long-term memory via vector DB. Activated when the user asks about past decisions, previous solutions, project history, "how did we do X", "last time", "we already solved", or when the agent needs historical context before planning an implementation.

PRD authoring and refinement. Activated when the user says "write PRD", "draft PRD", "break down feature", "scope this", "requirements for", "define feature", or wants to turn a rough idea into a structured product document.

Source-driven development. Activated when the user mentions a framework, library, or SDK decision, or when implementing features with external dependencies. Keywords: "official docs", "documentation", "how does X work", "what API", "which method", "framework", "library", "SDK". Ensures every framework-specific decision is backed by official documentation, not training data.

Code review following project conventions. Activated when the user asks for a review, CR, or code revision.

Systematic debugging workflow. Activated when the user says "debug", "why is this failing", "root cause", "broken", "investigate", "trace error", "bug", "not working", or wants to systematically investigate a problem.

Implement a feature from a PRD. Activated when the user asks to implement, build, or construct from a product document.

Persistence mode (Ralph). Activated when the user says "don't stop", "keep going", "Ralph mode", "persistence mode", "until it works", "until all tests pass", or wants iterative implementation until all acceptance criteria pass. The boulder never stops.

AI slop cleaner. Activated when the user says "clean up", "remove slop", "polish code", "review for slop", "AI artifacts", "deslop", or wants to clean LLM-generated code patterns. Also auto-activated after /implement as a final pass.

Recognize and apply intent markers — inline code annotations (:PERF:, :UNSAFE:, :SCHEMA:, :SECURITY:, :HACK:, :FLAKY:) that flag areas needing special attention during review. Activated when writing or reviewing performance-sensitive, security-critical, or schema-dependent code.

API design and contract definition. Activated when the user says "design API", "new endpoint", "API contract", "REST", "GraphQL", "OpenAPI", "swagger", "API spec", or wants to define API endpoints before implementation.

Database migration workflow. Activated when the user says "migration", "schema change", "add column", "alter table", "database change", "migrate", "create table", or wants to make database schema changes safely.

Safe refactoring workflow. Activated when the user says "refactor", "extract", "split module", "rename", "move", "restructure", "decouple", "simplify architecture", or wants to change code structure without changing behavior.

Technical debt identification and tracking. Activated when the user says "tech debt", "technical debt", "cleanup backlog", "what needs fixing", "code health", "debt audit", "code quality scan", or wants to systematically find and prioritize tech debt.

Test generation and coverage analysis. Activated when the user says "create tests", "write tests", "test this", "check coverage", "add test", "missing tests", "TDD", or wants to create or improve tests.

Generate UI components from PRD + design tokens when no Figma design is provided. Activated when implementing frontend features without a Figma link.

[SPEC] Clear description of when this skill should be activated. Include keywords that Claude uses to decide whether to auto-invoke. The description is CRITICAL for auto-activation — be specific.

Create Architecture Decision Records. Activated when the user mentions ADR, architectural decision, or technical trade-off.