Phase 2: optional repo skills and agents
The first pass adds shared markdown + Cursor rules + thin CLAUDE.md / GEMINI.md. If repeated agent mistakes appear, add narrow automation below.
Candidate Cursor / Codex skills (workflows)
- docs-sync — Trigger when
doctranslate/main.pyflags ordoctranslate/translator/*public behavior changes: checklist againstREADME.md,docs/configuration.md,docs/multi-translator.md,docs/local-translation.md,mkdocs.ymlnav. - translator-change-checklist — Router TOML schema (
NestedTranslatorConfig),validate_router_config,LiteLLMProviderExecutor, metrics flags; require tests undertests/test_translator_*. - pdf-pipeline-investigation — Symptom → stage mapping using
TRANSLATE_STAGESanddocs/ai/pdf-pipeline.md; point tohigh_level.pyand the relevantdocument_ilmidend/backend module. - release-readiness — Version in
pyproject.toml/doctranslate/__init__.py/main.pyper bumpver; changelog if the project uses one;uv run pytest, ruff, optionalmkdocs build --strict.
Codex custom agents (optional)
OpenAI Codex supports project-scoped agents under .codex/agents/ when you need parallel explorer vs worker flows. Only add TOML agents if the team uses Codex CLI regularly and wants standardized subagent prompts.
Claude Code
- Prefer
.claude/rules/withpathsfrontmatter ifCLAUDE.mdgrows past ~200 lines. - Skills: see Anthropic docs for progressive disclosure when a workflow is stable and repeated.
Keep any new artifact short and verifiable; long instruction files reduce adherence.