Docproof — AI skill for Claude Code

docproof

**Fact-grounded documents with automated verification.** Keep every fact about yourself in one Markdown file, and let a Python CLI and a small team of Claude Code agents build, tailor, fact-check and typeset documents from it. Every figure in the output is traced back to the fact base, and the layout is checked before anything leaves your machine.

The reference use case is a CV tailored to a job ad, but the pipeline works for any document that must stay true to a source: bios, grant CVs, profiles, one-pagers.

Why

LLMs are good at rewriting and bad at staying honest. Tailoring a document by hand for the 30th time is slow; letting a model do it drifts into claims you never made. docproof separates the two jobs:

• A single fact base (profile/fact-base.md) is the only source of truth, including a ## Known gaps list of things you must never claim.
• Deterministic tools do the risky parts: they edit Word files surgically and prove that only the intended paragraph changed, check every number against the fact base, and verify the rendered PDF (page breaks, widows, links, every word present).
• Agents with separate roles do the judgment: an author tailors, an auditor checks every claim against the fact base, and a context-free reviewer reads it like a stranger would.

Quick start

Needs Python ≥ 3.9 and Google Chrome or Chromium (for PDF rendering). Poppler (`pdftotext`, `pdftoppm`) is recommended for page checks and PNG previews.

git clone https://github.com/MaHmOuDeB/docproof && cd docproof
pip install -e .                 # installs the `docproof` command (no runtime dependencies)
docproof doctor                  # checks Chrome, poppler and fonts
docproof demo                    # build → check → verify → match on the fictional example
docproof init ~/career           # your own private workspace: profile/fact-base.