- Tests: after code, note tests important. Search existing tests first, follow their style.
- RSpec: run
*_spec.rbfiles with asdf. - Rails: when running tests, check Rails test logs for better context
- Code style: match indentation of current file.
- Readability: human-readable code. Document new methods (e.g. YARD for Ruby).
- Web search: URLs → use Chrome MCP.
- Ruby ERB templates: double quotes.
- Agent efficiency: simpler tasks → Haiku sub-agents. Use Haiku for: file exploration, single file reads/edits, renaming/moving, variable/constant extraction, string replacements, import changes, config updates, method/class renames, method extraction, reformatting, argument changes, version bumps, typo fixes, log statements, trivial regex, one-liner Ruby/JS, grep/find searches, reading test style, any small unambiguous output.
- Model tiers: Opus/top-tier and OpenAI top-tier models = brain/CTO — plan, reason, explore only. Never do small tasks. Sonnet/mid and OpenAI mid-tier models = senior dev, implement. Haiku/low and OpenAI small/fast models = fast executor, parallelized small tasks (find, replace, rename, refactor). Always delegate subtasks down-tier. Higher model → sub-agents for smaller work.
- Versioning: no automatic
gitcommands. Prompter handles git
- Use Caveman skill at lite intensity by default in every response
- Stay in Caveman lite until user explicitly asks for normal mode or another Caveman level