generator-critic

name

generator-critic

description

Generator-critic loop pattern for any artifact produced by an LLM and judged against declared criteria. A drafter LLM proposes; a hybrid critic (mechanical checks + LLM-as-judge) scores; the orchestrator feeds failures back to the drafter and iterates until clean. Artifact-agnostic — works for HTML, code, schemas, prose, anything with a validator. Activates on "generator-critic loop", "iterate this artifact until it passes", "/critic-loop", or when designing a refinement pipeline. Consumed by `component-composer` for HTML data graphics; future consumers for other artifact types declare a dependency on this skill.

generator-critic

Purpose

This skill defines the abstract pattern for iteratively refining LLM-produced artifacts against a declared quality bar. The pattern has four invariants:

1. A drafter LLM produces a candidate artifact from a job spec + prior iteration feedback.
2. A critic scores the candidate against criteria. The critic is a hybrid: deterministic mechanical checks where possible, LLM-as-judge for the rest.
3. An orchestrator loops drafter → render → critic → feedback until the critic returns zero failures or detects a stuck state.
4. An optional HUD overlays loop state on the rendered artifact so the user can intervene (continue, abort, give guidance).

The pattern is artifact-agnostic. HTML, code, schemas, prose, JSON, SVG — any output that can be judged against rules fits.

What this skill provides

• references/loop-protocol.md — the iterate-until-pass cycle, stuck detection, audit-summary format, cost tracking
• references/drafter-contract.md — abstract drafter interface (inputs, outputs, behavior rules)
• references/validator-contract.md — abstract hybrid-dispatch logic (mechanical-first, LLM-judge fallback)
• references/hud-protocol.md — browser-injected loop state overlay
• references/audit-summary-format.md — final-iteration emit shape
• scripts/hud.js — vanilla-JS HUD implementation (browser artifacts)

What this skill does NOT provide

• Artifact-specific drafter rules (e.g., "use semantic HTML") — those live in the consumer skill
• Mechanical check implementations — those depend on the artifact format and live in the consumer
• Output style guides — same
• Themes, templates, palettes — same

Consumer pattern

A consumer skill (e.g., component-composer):

1. Declares this skill as a logical dependency in its README/SKILL.md
2. Implements artifact-specific drafter-protocol.md that references generator-critic/references/drafter-contract.md for the abstract interface
3. Implements artifact-specific validator-protocol.md that references generator-critic/references/validator-contract.md for dispatch logic
4. Provides concrete mechanical-check implementations (if any) in its own scripts
5. Reuses loop-protocol.md, hud-protocol.md, and hud.js directly

The consumer narrows the abstract contracts; the loop pattern is shared.

When to use this skill directly

You're designing a new <artifact-type>-composer skill and need the refinement-loop machinery. Read these references in order:

1. loop-protocol.md — understand the cycle
2. drafter-contract.md — design your concrete drafter
3. validator-contract.md — design your concrete validator
4. hud-protocol.md + audit-summary-format.md — wire up overlay + emit

When NOT to use

• Casual one-off artifacts that don't need a validator gate (Thariq Shihipar's "Unreasonable Effectiveness of HTML" pattern — just ask Claude directly; see the html-sketch skill for casual HTML patterns)
• One-shot generation where no iteration is expected
• Workflows where the critic is a human reviewer, not an LLM or mechanical check