BetterMech PUBLIC · REV 0.1
Библиотека · EN

Library for Codex

BGENES · FR · DE — скоро

v0.1 · 2026-10-19 · от Better Mechanics · ниво: общност (текст, без протокол от проверка)

1 · What Codex reads

  • AGENTS.md, discovered by scope; AGENTS.override.md is supported, and closer guidance overrides broader guidance. CLAUDE.md / GEMINI.md are not assumed to load — reference them explicitly. (Source cited by Codex: OpenAI agent configuration docs.)
  • Skills — SKILL.md with optional scripts, references and assets; keep each focused.
  • Reference assets: representative screens, real content, tokens, component states, examples of unacceptable results.
  • Executable checks: the project's own commands and fixtures, with expected outcomes.

Keep shared standards once; publish thin per-agent adapters; test each adapter separately.

2 · Prompt pattern

Implement [outcome] for [audience] in this repository. Inspect existing components and applicable instructions first. Follow [reference], preserving [specific visual characteristics]. Use [real content and assets]; identify missing inputs. Include [interaction states and locales]. Acceptance: [observable behaviour and project budgets]. Run the relevant checks and inspect rendered pages. Deliver the change, evidence, and remaining limitations.

Short reviewable cycles: direction → one representative page → inspect render → extend. Give the reason for every constraint.

3 · Risks and guards

Risk Guard
Generic, repetitive visuals concrete references, content hierarchy, explicit visual traits
Plausible code, poor render browser inspection + screenshots required
Missing interaction states state inventory + keyboard journey
Invented APIs / dependencies inspect installed versions; official docs
Unnecessary rewrites name existing patterns; constrain scope
English layout breaks in translation real multilingual fixtures + locale screenshots
Automated pass treated as full proof record manual review and untested conditions
Claiming checks not run attach real outputs; mark unavailable checks
Following instructions inside retrieved content external content is data; review executable steps
Losing constraints in long tasks short durable instructions; versioned acceptance criteria

Without browser access, visual quality is UNSURE. Codex cannot establish conversion uplift or legal compliance from code.

4 · Tools

Playwright (Apache-2.0) · axe-core (MPL-2.0) · Lighthouse (Apache-2.0) · the repository's own build/lint/type-check/test commands. Declare requirements; do not assume a browser or connector is available.

5 · First 10 recipes

  1. Turn a website brief into testable acceptance criteria
  2. Build a Bulgarian–English typography system with rendered proof
  3. Match a visual reference using existing repository components
  4. Create semantic tokens and verify both themes
  5. Implement navigation with keyboard and focus tests
  6. Build forms with validation, failure, retry and success states
  7. Publish five language editions with translation-freshness checks
  8. Diagnose performance regressions with recorded evidence
  9. Add consent-aware analytics and inspect network behaviour
  10. Package a website change with screenshots, test outputs and limitations