Agentic quality gates
Skill g-shevchenko/agentic-quality-skills/skills/agentic-quality-gates
Production-grade quality skills for AI coding agents: red-first TDD, quality gates, and golden benchmark uplift loops.
npx -y skills add g-shevchenko/agentic-quality-skills --skill agentic-quality-gatesAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use before non-trivial agentic, coding, workflow, tool, automation, release, or public-surface work, and when the user says quality gate, acceptance gate, red flags, risk gate, architecture gate, decision handoff, side-effecting workflow, contracts first prompts last, prompt-only fix, human gate, or release gate. Classifies risk, defines deterministic acceptance gates, catches red flags, and decides whether to proceed autonomously or ask a human.
SKILL.md
1.9 KB, as published. Nobody here has run it
Agentic Quality Gates
Use this skill to avoid "sounds done" work. It turns vague agentic work into explicit ownership, gates, and proof.
Quick Triage
Before implementation, classify:
Task class: not_agentic | llm_boundary | agent_workflow | side_effecting_workflow | runtime_infra | eval_only
Reason: <specific reason>
Protocol path: skip_full_gate | full_architecture_gate
Use full_architecture_gate when the task changes LLM behavior, prompts with behavioral effect, tools, workflows, automation, evals, side effects, or recurring bug handling.
Read references/gates-and-red-flags.md for the full checklist.
Contracts First, Prompts Last
The model may classify, draft, extract, rank, or judge inside a boundary. Deterministic code owns execution, state, validation, retries, rollback, side effects, and acceptance.
Prompt-only fixes for recurring bugs are incomplete unless paired with a validator, schema, test, golden case, or proof gate.
Decision Handoff
Proceed autonomously when the decision is reversible, low risk, evidence-backed, and testable.
Ask a human when the decision is destructive, public, credential-related, expensive, legally sensitive, security-sensitive, or under-evidenced.
When asking, provide one recommendation, one real alternative, evidence, risk, and the exact answer needed.
Done Criteria
Report:
- task class;
- acceptance gate;
- red flags checked;
- proof command or artifact;
- human decision required, if any.