doc-metadata

Files (2)

• SKILL.md
• set_doc_metadata.py · scripts/

SKILL.md (first lines)

# Doc metadata: deliverables carry the user's chosen identity, never the tool's

Libraries leave fingerprints in file properties that anyone can see in File > Info:
python-docx writes author `python-docx` and comment `generated by python-docx`; openpyxl
writes `openpyxl`; reportlab and PyMuPDF write themselves as producer. These are always
removed. Which name and company go IN their place is the user's call, because not every
document belongs to the same organisation (work, personal projects, customer templates).

## Step 1: decide the identity (ask once per project)

1. Look for a recorded choice: the project's newest handoff or `CLAUDE.md`, or a project
   memory named `doc-identity`. If found, use it and do not ask again.
2. If none is recorded, ask the user before writing metadata, one question:
   - **Work identity** (their name, their company)
   - **Personal / no company** (their name, company blank)
   - **Other** (they name the author and company)
   - **Leave as-is** (only strip tool and AI fingerprints)
3. Record the answer in the project handoff and a `doc-identity` project memory, so later
   files in the same project are stamped without asking. A different answer for one file
   ("this one goes out under the customer's name") applies to that file only.

## Step 2: stamp the file

```bash
"$(command -v python3 || command -v python)" "$HOME/.claude/skills/doc-metadata/scripts/set_doc_metadata.py" <file-or-folder> \
  --author "<Name>" --company "<Company>" --title "<document title>"
# no company:   --company ""
# strip only:   --strip-only
```

The script refuses to run without `--company` (or `--strip-only`), so the choice is never
made silently. Then confirm with `--check`; it must print `fingerprints: none` and `xml: ok`
(it exits 1 on a broken docProps part; the stamp itself refuses to write one).
(Installed somewhere other than `~/.claude/skills`? Adjust the path.)

## Rules

- Tool and AI fingerprints are always removed, whatever identity is chosen.
- No "Claude", "AI", "Anthropic", "generated by", or tool names in titles, subjects,
  comments, keywords, headers, footers or body text of deliverables, unless the user asks
  for disclosure.
- No references to AI-only companion files (for example `*-AI.md`) inside a human document.
- If a generator script builds the file, set the same properties inside the script too
  (python-docx: `doc.core_properties.author = ...`, `comments = ""`), so a rebuild never
  regresses.
- Markdown or text working files the agent keeps for itself are exempt.

## Notes

- Office files are rewritten in place; only `docProps/core.xml` and `docProps/app.xml`
  change. Close the file in Word first (a locked file raises PermissionError).
- PDF needs PyMuPDF (`pip install pymupdf`); it sets author, creator and producer.

From mhshaon98/agent-skills, MIT.

More from mhshaon98/agent-skills