Skip to content

CLAUDE.md

Latest commit

 

History

History
65 lines (43 loc) · 3.61 KB

CLAUDE.md

File metadata and controls

65 lines (43 loc) · 3.61 KB

3d_resources — project rules

Controlled vocabulary (entry tags)

Source of truth: schema/vocab.yml + schema/entry.schema.json.

Rule: When adding/editing entries, only use values that already exist in the vocab for these tag groups:

  • license (closed enum)
  • entry_type (closed enum)
  • tags.workflow (closed enum)
  • tags.output (closed enum)
  • tags.platform (closed enum)
  • tags.skill (closed enum)

tags.tech is freeform but curated — prefer existing values; new ones produce warnings, not errors.

Never invent new values for closed enums. If an entry seems to need one (e.g. product-design, print, visionos, tracking, rotoscoping), do NOT add it silently to vocab.yml or entry.schema.json. Instead:

  1. Map to the closest existing value (e.g. product-design → product-viz, tracking/rotoscoping → compositing), OR drop the tag if no good fit.
  2. Ask the user if a genuinely new category is needed before extending the vocab.

CI runs node scripts/validate.js — closed-enum violations exit 1 and block merge.

Validation

Run node scripts/validate.js after any data edit. Must show ✓ Validation passed. before commit.

Writing style (user-facing text)

Applies to: data/**/*.yml (descriptions, titles, readme_tags), template strings in scripts/render.js, scripts/build-html.js, scripts/build-section-pages.js, scripts/build-feed.js, scripts/build-llms-txt.js, scripts/build-og-images.js, and any other surface a site visitor reads.

Rules:

  1. No em-dashes (—). Ever. In titles, use : . In prose, use . (period + capitalize next word) or , depending on context. The OG-image splitter (build-og-images.js) keys on : for two-line layout — don't reintroduce em-dashes.
  2. No AI tells. Banned words/phrases: comprehensive, robust, powerful, seamless(ly), leverage(s/d) (use "use"), utilize(s/d/ing) (use "use"), cutting-edge, state-of-the-art, the ultimate, game-changing, unleash. Banned trailing phrases: and more, and beyond, and so much more.
  3. No marketing adjective stacks ("fast, free, and powerful"). One concrete attribute beats three vague ones.
  4. Caveman lite for descriptions. Terse, factual, 1 sentence preferred. State what it is and what it does. Drop hype.
  5. Internal docs/memory/CLAUDE.md/comments are exempt. Em-dashes and informal hedging are fine there.

Project memory system (/memory)

Local, project-scoped memory lives in ./memory/. Distinct from global ~/.claude memory.

Files:

  • goals.md — active + done goals
  • tech_stack.md — frameworks, langs, build tools in use
  • decisions.md — dated architectural/structural decisions w/ rationale
  • preferences.md — user collab style + project rules
  • tools.md — CLIs / utilities in active use
  • agents.md — sub-agents used + outcomes
  • plugins.md — skills + MCP extensions in use
  • user-prompts.md — verbatim log of every user prompt + 1-line answer

Protocol (MANDATORY)

Session start: Read all 8 files in ./memory/ before responding to first prompt. Treat them as authoritative project context.

After every user prompt:

  1. Append the verbatim prompt to user-prompts.md with timestamp + 1-line answer summary.
  2. Update any other file whose content changed (new decision → decisions.md; new tool used → tools.md; new preference learned → preferences.md; etc.).

Session end: Final sweep — ensure all files reflect latest state. Add new decisions, move completed goals to Done, log any agents spawned.

Rule: keep entries terse (caveman lite). No filler. Date everything (YYYY-MM-DD).