Built-in Skills Catalog
SkillForge ships with 16 production-grade, zero-slop skills. These skills directly counter known failure patterns observed in repository-level agent evaluations (SWE-bench).
1. Code Quality & Architectural Discipline
blast-radius-containment
- Purpose: Analyze upstream callers and enforce strict minimal-diff boundaries before modifying shared code or public APIs.
- Trigger: Use when modifying core utilities, database models, public API signatures, or widely imported types.
- Key Invariants: Rule of Least Mutation, upstream call-graph audit, maximum 3-file touch budget.
antislop-code
- Purpose: Enforce comment hygiene, eliminate generic AI-generated narrative filler, and stop speculative over-engineering.
- Trigger: Use when reviewing code, refactoring modules, or auditing pull requests.
- Key Invariants: Deletes trivial echo comments, preserves domain invariants, blocks premature abstractions (YAGNI).
spec-driven-development
- Purpose: Design systems contract-first using typed schemas, OpenAPI/JSON-Schema models, and failing test harnesses before writing application logic.
- Trigger: Use when architecting new modules, APIs, or data models.
- Key Invariants: Red-green verification loop, typed boundary definitions.
zero-downtime-migration
- Purpose: Plan and execute non-breaking database schema changes and API payload evolutions using the Expand-Contract pattern.
- Trigger: Use when modifying SQL tables, ORM models, or public API endpoints in production systems.
- Key Invariants: Four phases (Expand dual-write → Async batch backfill → Switch reads → Contract cleanup). No in-place column renames.
2. Autonomous Loop Stability & Debugging
systematic-debugging
- Purpose: Execute a disciplined four-phase root-cause investigation loop before modifying code.
- Trigger: Use when diagnosing failing tests, runtime crashes, race conditions, or performance regressions.
- Key Invariants: Phase 1 Reproduce → Phase 2 Isolate → Phase 3 Falsifiable Hypotheses → Phase 4 Surgical Patch.
cognitive-loop-breaker
- Purpose: Detect and interrupt trial-and-error death spirals where repetitive edits fail tests across consecutive turns.
- Trigger: Use when stuck in recurring test failures or repetitive debugging loops.
- Key Invariants: Three-Strikes Circuit Breaker: forces
git restore ., invalidates failed hypotheses, resets to first principles.
flaky-test-isolation
- Purpose: Diagnose and eliminate non-deterministic test failures caused by race conditions, unseeded randoms, state leakage, or clock skew.
- Trigger: Use when tests fail intermittently in CI or pass inconsistently.
- Key Invariants: 50x stress reproduction loop, inspection of the 4 culprits (state leak, clock skew, arbitrary sleep, unordered sets).
runtime-api-verification
- Purpose: Introspect installed packages, verify method signatures, and inspect type definitions before invoking unfamiliar third-party library APIs.
- Trigger: Use when integrating third-party SDKs, upgrading dependencies, or resolving import errors.
- Key Invariants: Probing
inspect.signature(), compiler.d.tschecks, running a 3-line probe script before application code.
3. Context Efficiency & Multi-Agent Operations
caveman
- Purpose: Ultra-terse, high-density technical communication mode that reduces output tokens while preserving technical accuracy, exact commands, and safety warnings.
- Trigger: Use when concise responses or lower token consumption is desired.
context-compactor
- Purpose: Distill extensive agent session transcripts into dense knowledge states, active hypothesis ledgers, and verified file citation indices.
- Trigger: Use when context windows approach limits or during multi-agent handoffs.
mcp-tool-designer
- Purpose: Architect high-performance, type-safe Model Context Protocol (MCP) servers and tools with strict JSON schemas, token bounds, and idempotent annotations.
- Trigger: Use when designing MCP tools or building AI agent integrations.
agent-observability
- Purpose: Instrument agent tool calls, prompt latency, token expenditure, and error boundaries with structured OpenTelemetry or JSON-L logging.
- Trigger: Use when debugging agent loops or optimizing multi-agent tracing.
human-in-the-loop
- Purpose: Enforce explicit human approval boundaries before executing irreversible actions, deploying code, modifying schemas, or making external financial/data mutations.
- Trigger: Use when building safe autonomous agent workflows.
summarize
- Purpose: Distill complex engineering documents, logs, and pull request diffs into structured executive summaries with citations and key architectural decisions.
- Trigger: Use when summarizing technical repositories or code reviews.
tavily-search
- Purpose: Execute targeted web research across high-trust primary sources, official documentation, and technical repositories.
- Trigger: Use when researching cutting-edge library APIs, SDKs, or current industry standards.
self-improving-agent
- Purpose: Extract concrete lessons and operational heuristics from runtime failures, user corrections, and test cycles into durable project memory.
- Trigger: Use when capturing recurring debugging patterns or user preferences.