clean-delivery

name	clean-delivery
description	Pre-delivery hygiene for anything that leaves the house — strip invisible/exotic Unicode from generated text (it breaks Hebrew RTL, Word, WordPress and newsletters) and strip authoring metadata from DOCX/PDF/JPEG/PNG. Use before sending an article, proposal, report or document to a client. Never rewrites wording.

clean-delivery — the last step before a file leaves the house

When this is the right skill

Anything produced by a model or exported from a local tool and then handed to someone outside the team: an article, a proposal, a report, a PDF, a newsletter draft.

Two problems, both invisible until they embarrass you:

1. Text carries characters nobody typed. Models emit zero-width spaces, no-break spaces, word joiners, bidi marks and — increasingly — Unicode tag characters that can hide an entire payload inside ordinary-looking prose. In Hebrew this is not cosmetic: a stray RLM or NBSP breaks RTL layout, and pasting into WordPress or Word drags the junk along.
2. Files carry authoring metadata. Every DOCX and PDF you export names the author, the tool and sometimes the local filesystem path. A proposal sent to a client should not disclose who wrote it, on which machine, with which template.

What this skill deliberately does NOT do

It never rewrites wording to disguise that text was written by AI. Tools in this space usually bundle a "beat the detector" layer; we left it out on purpose. We sell AI content systems openly — hiding it is a reputational risk to the client, and the technique is unprovable anyway. This skill only removes what should not have been in the file in the first place.

How it's built

clean-delivery/
  SKILL.md
  engine/clean.py     Python 3.9+, zero dependencies

PDF cleaning shells out to qpdf when it is on PATH. Everything else — OOXML, JPEG, PNG, all text handling — is pure stdlib.

Commands

C=<SKILL_DIR>/engine/clean.py

# --- text ---------------------------------------------------------------
python3 $C text article.md --check          # report only; exit 1 if dirty
python3 $C text article.md                  # cleaned text to stdout
python3 $C text article.md --in-place       # rewrite the file
python3 $C text draft.md --out final.md
cat draft.md | python3 $C text -             # stdin → stdout

# --- files --------------------------------------------------------------
python3 $C file proposal.pdf                # → proposal.clean.pdf
python3 $C file report.docx cover.jpg --in-place

Text flags

Flag	Effect
--check	Report what was found, change nothing. Exit code 1 = dirty. Use in a pre-send gate.
--keep-bidi	Keep RLM/LRM and bidi isolates. Use only when a Hebrew document genuinely mixes directions and relies on explicit marks.
--ascii-punct	Also fold curly quotes, em dashes and ellipses to ASCII. Off by default — for a magazine article you usually want the real typography. Turn it on for plain-text email, CSV or code.
--no-collapse	Keep trailing spaces and runs of blank lines.

Default behaviour: remove invisible characters, remove bidi controls, turn exotic spaces into a plain space, drop Unicode tag characters and variation selectors, normalise to NFC, trim trailing whitespace, collapse 3+ blank lines to 2.

Supported file types

Type	What is removed
.docx .xlsx .pptx	docProps/core.xml, app.xml, custom.xml — blanked, not deleted (a package whose declared parts are missing will not open)
.pdf	Document info dictionary and the XMP packet, by rebuilding the file from its pages. Requires qpdf.
.jpg .jpeg	Every APP1–APP15 segment (EXIF, XMP, Photoshop IRB) and comments. APP0/JFIF is kept — it is structural.
.png	tEXt iTXt zTXt eXIf tIME iCCP dSIG chunks

Unsupported types are reported as skipped, never silently passed. Output is JSON, one entry per file.

Working with the user

1. Default to --check first on anything you did not generate yourself. Show what was found before changing a file the user cares about.
2. Never clean in place without saying so. For a client-facing document prefer the .clean. copy, so the original stays available.
3. Two flags need a decision, not a default: --keep-bidi (Hebrew documents with mixed direction) and --ascii-punct (destroys intentional typography). Ask when it matters.
4. Image EXIF is a privacy call, not a formatting one — a photo from a phone carries GPS coordinates. Mention it when the user attaches photos to something public.

Where it fits

• Content pipeline (article generation, interview → article): run text --check on every draft before it reaches the client, and text --in-place before publishing.
• Proposals and reports: run file on the exported PDF before it is sent.
• Privacy baseline: metadata stripping is the cheap end of "do not leak what you did not mean to send".