Skip to content

CLAUDE.md

Latest commit

 

History

History
44 lines (35 loc) · 2.44 KB

CLAUDE.md

File metadata and controls

44 lines (35 loc) · 2.44 KB
Raw

WRITING AGENT

A self-correcting, autonomous writing system (books, articles, and more). Interactive TUI (writing-agent / python -m writingagent) + one-shot CLI, on any OpenAI-compatible model host (OpenRouter, OpenAI, Anthropic, DeepSeek, Gemini, … — no blessed default). Cross-platform: Linux · macOS · Windows (CI runs the suite on all three × Python 3.10–3.13). Full spec in docs/plan.md; how-to-run in README.md.

Start every session here

  1. Read docs/dev/resume.md - the session log: what happened last time and the next step. (maintainer-local, not in the repo; skip if absent)
  2. Read docs/plan.md - the authoritative spec (architecture, memory, nodes, state machine, learning, production, model routing, CLI).

End every session

  • Prepend a dated entry to docs/dev/resume.md (what changed, decisions, concrete next step).
  • Put durable decisions in docs/plan.md; keep docs/dev/resume.md as the journal.

Map of the repo (each thing has ONE home)

  • src/writingagent/ — all code. Entry points: the writing-agent console script and python -m writingagent. There is no launcher script at the repo root.
  • src/writingagent/resources/ — ALL shipped data (models.yaml default routing, seed skills, gold corpus, personas). The single source; no repo-root copies exist.
  • Runtime state (brain/, .index/, config/*.yaml, .env keys) lives in the agent home: $WRITINGAGENT_HOME, else the OS user-data dir (see src/writingagent/paths.py). The app never writes into the repo.
  • docs/ — all project docs: plan.md (spec), design.md (UI system), prd.md, learning.md, roadmap.md, proposals/, dev/ (maintainer journals).
  • demo/ — the Gradio Hugging Face Space demo (NOT the package's web dashboard; that is src/writingagent/webui/).
  • examples/ — real output, including examples/sample-run/ (a full book run).
  • website/ — the Astro docs site: its OWN nested git repo (remote vikast908/writingagentdocs, deployed on Vercel), deliberately untracked here.
  • tests/, benchmarks/, scripts/, assets/ — the usual.

Conventions

  • docs/plan.md = single source of truth. docs/dev/resume.md = running log (with docs/dev/test.md, the verification log). Don't duplicate between them.
  • All numeric thresholds in docs/plan.md are tunable config, not hard-coded.
  • Borrow patterns from Hermes/GBrain, not their surface area. This is a narrow writing pipeline.