franklin CLAUDE.md

Franklin

Turns technical books (EPUB) into Claude Code plugins via a staged pipeline: ingest → map → plan → reduce → assemble.

Stack

• Python 3.12, managed with uv
• Typer CLI (franklin), Pydantic v2 models, Rich output
• Anthropic SDK for LLM calls, tenacity for retries
• ruff (line-length 99), mypy strict, pytest

Commands

uv sync                      # install
uv run franklin ingest BOOK  # run ingest stage
uv run pytest                # tests
uv run ruff check .          # lint
uv run ruff format .         # format
uv run mypy                  # type check

Run directories land in ./runs/<slug>/ — book.json plus raw/chNN.json per chapter. Ingest is deterministic, no LLM calls.

Layout

• src/franklin/cli.py — Typer entrypoint; per-stage commands are thin shells over _do_<stage>_stage helpers.
• src/franklin/services/ — stage services (IngestService, MapService, PlanService, ReduceService, AssembleService). Each takes a Pydantic input, emits ProgressEvents via a callback, returns a Pydantic result. Independent of Typer/Rich — usable from non-CLI callers.
• src/franklin/ingest/, mapper/, planner/, reducer/, assembler/ — stage-pure primitives (parsers, LLM callers, validators). Services orchestrate; these do the work.
• src/franklin/llm/ — Anthropic client wrapper, prompt loader, prompts as markdown.
• src/franklin/schema.py — Pydantic schemas (BookManifest, ChapterSidecar, PlanManifest, Artifact, etc.) — the contract between stages.
• tests/ — pytest. test_golden_path.py is the end-to-end regression oracle that threads all five stages with a scripted fake Anthropic client.

Working preferences

• Use uv run for anything Python — never bare python or pip.
• Keep mypy strict-clean and ruff-clean before declaring a task done.
• Prefer editing existing files; don't create new modules speculatively.
• Prompts live as markdown under src/franklin/llm/prompts/ and are loaded by the prompt loader — add new ones there, don't inline strings.
• Pydantic models are the contract between stages. When changing a stage's output, update the model first.
• Don't commit unless asked. When asked, write terse commit messages matching the existing log style (Add X, Classify Y, imperative, no body unless needed).