agent-utilities
kylerberry/agent-utilities
Public utilities to supercharge your Claude Code
Instalação
npx convez-ai kylerberry/agent-utilities1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Agent Utilities
A skill-based toolkit for building software with AI coding agents — installable slash commands and primary agents that enforce disciplined, repeatable workflows.
Skills
skills/
├── craft/ # Autonomous phase-gate execution (C → R → A → F → T → S)
└── craft-hitl/ # Same flow with mandatory human-in-the-loop gating at TODO(human) seams
/craft — Autonomous CRAFTS Workflow
Invoke for every non-trivial task that can be completed without human judgment at a critical seam.
CRAFTS is a sequential phase-gate workflow: finish the current phase before moving to the next. Do not run phases in parallel.
Full flow (business logic, multi-file work, domain boundaries):
| Phase | Subagent | What happens |
|---|---|---|
| Conceptualize | craft-planner | Scope, test cases, implementation plan, risks before coding |
| Render | craft-builder | TDD strictly: Red → Green → Refactor |
| Assess | craft-evaluator | Review diff for quality, reuse, efficiency, type correctness |
| Fix | craft-builder | Address blocking issues from Assess |
| Tighten | craft-security | Security-hardening review of the diff |
| Sharpen | craft-sharpener | Update docs with lessons learned; commit |
Lite flow (config, scaffolding, single-file fixes): Render → Sharpen.
Start lite, escalate to full if the task grows.
Builder / Evaluator model diversity: When exact per-spawn model selection is available, the R/F craft-builder and A craft-evaluator must run on different but equal-capability models. If only tier aliases are supported, both remain at medium and the phase report notes that exact diversity could not be enforced.
/craft-hitl — Human-in-the-Loop CRAFTS Workflow
Invoke for issue slices labeled HITL implementation or HITL design/review, or any task where the PRD reserves a critical decision for human judgment.
Identical to /craft, but R — Render contains a mandatory pause. The agent scaffolds to the seam, leaves exactly one TODO(human) marker with specific context, then pauses. The human fills the critical logic; the agent resumes verification and completion.
When to pause:
- Issue labeled HITL implementation — agent scaffolds/tests, human owns critical logic
- Issue labeled HITL design/review — requires human taste/content/design review
- Subjective choice (naming, UX copy, algorithmic trade-off) explicitly reserved for human judgment
When NOT to pause: routine refactoring, clear-cut implementation, or decisions inferable from the PRD.
If a task starts as HITL but the human defers the decision back to the agent, switch to /craft and complete autonomously.
Agents
agents/
├── craft-planner/ # C phase — scope, tests, risks, plan
├── craft-builder/ # R/F phases — test-driven implementation and fixes
├── craft-evaluator/ # A phase — diff review, simplification, verification
├── craft-security/ # T phase — security and trust-boundary review
└── craft-sharpener/ # S phase — durable docs, standards, learnings
These are primary agents with priority fields. They appear in the agent picker and can be chatted with directly. They are also spawned on-demand by the /craft and /craft-hitl skills via AgentSpawn.
Installation
Copy skills and agents into your project's .agents/ directory (or your agent's global path):
cp -r skills/craft /path/to/your-project/.agents/skills/
cp -r skills/craft-hitl /path/to/your-project/.agents/skills/
cp -r agents/* /path/to/your-project/.agents/agents/
Then invoke with /craft and /craft-hitl from your agent interface.
Philosophy
Planning beats replanning. Wrong assumptions caught before code exists cost minutes. Caught after implementation they cost hours. CRAFTS front-loads the thinking in Conceptualize.
Self-review is impossible. The builder can't see their own blind spots. The Assess and Tighten phases are fresh-context reviews because the agent that wrote the code is the worst reviewer of it.
Security is a gate, not an afterthought. Tighten runs before every commit. Vulnerabilities found pre-commit cost minutes; found post-deploy they cost far more.
Knowledge compounds. Sharpen updates domain docs with lessons learned after every task. Week 1 those files are stubs. Week 8 they're handbooks that make every subsequent task faster and less error-prone.
The right human at the right seam. HITL gating isn't about distrusting the agent — it's about reserving human judgment for the decisions that actually need it, while the agent handles everything else autonomously.
License
MIT