Tier-2 Review: humanize-ppt

What we attempted

We pulled humanize-ppt (cluster: document-generation, source: nexu-io / open-design / plugins-community) and ran it through the standard four-probe harness: install, minimal-roundtrip, edge-corner, and a fourth probe that was skipped before execution. The goal was to confirm that the skill could be installed, invoked end-to-end on a small PPTX, and exercised against an edge case (embedded media handling).

What failed

Zero tests passed. Two probes ran to partial completion, zero outright failed, and one was skipped. The single blocking issue is structural, not incidental:

SKILL.md documents only an outline/brief/checkup workflow with no install command, file format, or rendering capability, so install and roundtrip cannot be fully validated.

Concretely:

• install — partial. The SKILL.md contains no install command, no package manifest, no declared system dependencies. There is no documented entrypoint to invoke. In the harness this resolved to a no-op; on a clean machine it would likely fail outright because there is nothing to install from.
• minimal-roundtrip — partial. The skill is explicit that it "never renders slides itself." Its output is an AST outline plus a production brief for a downstream renderer. So a roundtrip does not produce a transformed PPTX — it produces an outline artifact. No file format or write-back path is specified, which means "roundtrip" in the usual sense (in → transform → out, byte-comparable) is not a thing this skill does. We could only confirm the outline stage produced something; we could not confirm it was well-formed, because no schema is published.
• edge-corner — skipped. The skill mentions per-page visual-enhancement decisions (image / SVG diagram / video) at the outline stage. But because it does not parse or render PPTX, embedded-media handling in a rendered deck cannot be exercised. The probe has nothing to attach to.

What we observed

The SKILL.md is well-written prose. Trigger language is crisp ("给这份 deck 做演讲体检", "PPT 渲染质检"). The three-stage model — outline → brief → capped 3-round checkup — is coherent and the "never renders slides itself" boundary is stated plainly rather than buried. That honesty is a virtue and is almost certainly why D1–D3 score well.

But the same honesty is the failure mode. The skill is a specification for a workflow, not a runnable artifact. It hands off to "HTML-PPT skills or native PPTX" without naming a concrete downstream, without a brief schema, and without a file-format contract. In a harness — and, more importantly, in an agent pipeline — that handoff is where the work actually happens, and it is undefined here.

Rating caveat

The composite 4.2 / 5.0 is theoretical. It reflects the quality of the written specification, not verified behavior. Until a physical re-run resolves the install entrypoint and the brief/output schema, the D4 (self-containment, 4.0) and D5 (reusability, 3.0) scores in particular should be treated as provisional — D5 in particular looks generous given that nothing downstream can consume the output without additional glue.

Is it still valuable in principle?

Yes. The idea — separate what the talk must do to the audience (AST) from how the slide looks, then run a bounded QA pass on the rendered result — is a real gap in most PPT-generation skills, which jump straight from prompt to pixels. A skill that forces an outline and a checkup before/after rendering is doing useful discipline work.

The fix is small and concrete: publish (1) an install/entrypoint command, (2) a JSON schema for the AST outline and the production brief, and (3) a named downstream renderer contract. With those three, this becomes testable and genuinely reusable. Without them, it remains a thoughtful prompt wearing a skill's clothing.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-09-12.

Overall: broken. 0 tests passed, 2 partial, 0 failed, 1 skipped; key blocker: SKILL.md documents only an outline/brief/checkup workflow with no install command, file format, or rendering capability, so install and roundtrip cannot be fully validated.

Test	Status	Notes
install	partial	SKILL.md describes an outline/AST + brief + checkup workflow but does not document an install command, package manifest, or system dependencies; install cannot be verified from the provided text and would likely be a no-op or fail without a documented entrypoint.
minimal-roundtrip	partial	The skill explicitly 'never renders slides itself' and only produces an AST outline plus a production brief for a downstream renderer, so a roundtrip would yield an outline/brief artifact rather than a transformed PPTX; no file format or write-back path is specified in the SKILL.md.
edge-corner	skipped	SKILL.md mentions per-page visual-enhancement decisions (image / SVG diagram / video) at the outline stage, but since the skill does not render or parse PPTX itself, embedded-media handling in a rendered deck cannot be exercised here.