Inspect repair

Skill smartdatabrokers/slideforge-mcp/skills/inspect-repair

Run a free deterministic quality report on ANY PowerPoint file (yours or generated elsewhere) and optionally auto-repair it. Use when the user asks to check/lint/QA/review a pptx, find broken slides, fix overflowing or hidden text, clean up a deck before sending, or verify a generated deck. Uses the SlideForge REST API (works without the MCP server); repair never alters words.From its SKILL.md

Install

npx -y skills add smartdatabrokers/slideforge-mcp --skill inspect-repair

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

3.5 KB, as published. Nobody here has run it

Inspect & repair any deck (Deck Doctor)

POST /v1/inspect returns a Deck Quality Report for any .pptx — decks SlideForge never rendered included. Free, deterministic, no LLM: the file is analyzed with real font metrics and geometry, never sent to a model. Repair is $0.02 per repaired slide, and its dry_run returns the exact fix plan (and therefore the exact price) free.

Inspect

curl -X POST https://api.slideforge.dev/v1/inspect \
  -H "Authorization: Bearer sf_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"pptx_base64": "'"$(base64 -w0 deck.pptx)"'"}'

Sources (exactly one): pptx_base64 (≤25 MB), pptx_url (https), or job_id (a SlideForge render). First 50 slides; 30 requests/hour.

What it finds — each with a machine code, severity, message, remedy, and the offending shape_id: text overflow / clipping (measured with real font metrics, autofit-aware), content hidden behind opaque shapes (z-order), off-canvas leftovers (parked notes leak content!), WCAG contrast, sub-9pt fonts, image-only slides, empty slides, hidden slides (marked; they never fail the deck), density.

Read it honestly: status is clean | review | failed. summary.font_size_coverage / contrast_coverage state how much of the deck the checks could actually resolve — inherited theme values it can't resolve are skipped, never guessed. Report findings to the user by slide number with the remedy; don't paraphrase severities up or down.

Repair (deterministic, never your words)

# free exact quote first
curl -X POST https://api.slideforge.dev/v1/repair -H "Authorization: Bearer sf_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"pptx_url": "https://…/deck.pptx", "dry_run": true}'
# then apply
curl -X POST https://api.slideforge.dev/v1/repair -H "Authorization: Bearer sf_live_YOUR_KEY" \
  -H "Content-Type: application/json" -d '{"pptx_url": "https://…/deck.pptx"}'

Actions (subset via actions[], default all): shrink_to_fit (measured, floor 9pt), raise_min_font (only when it still fits), remove_off_canvas (fully-invisible junk only), fix_contrast (WCAG AA, hue-preserving). No characters are ever altered — the response carries a computed visible_text_untouched proof, a before/after report diff, and flagged_not_fixed[] for anything ambiguous (with remedies). A run that fixes nothing bills nothing; the repaired deck persists as a job — download it with your API key via GET /v1/jobs/{id}/pptx.

Workflow patterns

Pre-send QA: inspect → show the user errors/warnings by slide → offer repair dry-run quote → apply on confirmation → return the repaired file + the before/after diff.
Verify generated decks: after any tool (SlideForge or otherwise) produces a pptx, inspect it before delivering — it's free and takes ~1s per 30 slides.
Hygiene sweep: off-canvas findings often reveal parked working notes (names, comments) that ship inside the file — flag these as a content-leak risk, not just a visual issue.