Tier-2 Review: summarize (document-generation cluster)

What we attempted

We picked up summarize from the OpenClaw creator page expecting a document-generation skill that ingests heterogeneous sources — URLs, YouTube/video, podcasts, articles, transcripts, PDFs, and local files — and emits either a summary or a transcription. The promise is broad and attractive: a single entrypoint for "turn this thing into text, or turn this thing into a shorter version of itself."

Three tests were queued against the skill:

1. install — confirm the skill could be retrieved, its dependencies resolved, and its CLI or entrypoint made callable.
2. minimal-roundtrip — feed a single known input (a short article URL) and verify a summary or transcription is produced in a predictable format.
3. edge-corner — exercise a scanned PDF, which is the classic failure surface for any summarize/transcribe skill because it requires OCR rather than text extraction.

What failed

All three tests were skipped as untestable, not passed. The harness reported 0 passed, 0 partial, 0 failed — a null result, not a green one. The blocker is singular and unambiguous: the provided SKILL.md is truncated to a single sentence.

That sentence reads, in full:

Summarize or transcribe URLs, YouTube/videos, podcasts, articles, transcripts, PDFs, and local files.

There is nothing else. No install command. No dependency list. No version constraints. No CLI entrypoint. No output format specification. No mention of where files are written, whether output is stdout or a file, whether summaries have length controls, or how OCR is (or isn't) handled for scanned PDFs. No error behavior. No mention of rate limits, API keys, or which services back the YouTube/podcast paths.

Each test died on a different facet of the same absence:

• install could not proceed because there is no install command, no dependency manifest, and no version pin to verify.
• minimal-roundtrip could not proceed because there is no specified target format, no CLI entrypoint, and no defined output-writing behavior — we cannot assert on a result we cannot predict the shape of.
• edge-corner could not proceed because OCR capability, PDF handling internals, and error semantics are simply not described.

What we observed

The failure mode is documentation truncation, not a broken skill. The harness did not observe the skill misbehave; it observed that there was nothing to invoke. This is an important distinction. We are not reporting a bug, a crash, a wrong output, or a security issue. We are reporting that the artifact we were handed is a one-line description where a specification should be.

Consequently, the composite score of 4.1 / 5.0 and the dimension scores (D1 4.5, D2 4.0, D3 4.5, D4 3.5, D5 4.0) are theoretical. They reflect what the single sentence implies about trigger clarity, output specificity, scope precision, self-containment, and reusability — not what any executed test confirmed. D4 (self-containment) at 3.5 is the lowest and is the dimension most directly contradicted by the evidence: a skill whose entire body is one sentence is, by construction, not self-contained.

Is the skill still valuable in principle?

Yes — and this is not a consolation prize. The scope described in that one sentence is genuinely useful and genuinely hard. A skill that reliably unifies URL scraping, YouTube transcripts, podcast audio, article extraction, PDF text layer, and local file ingestion behind one interface would save real work, and the "summarize or transcribe" duality covers the two most common downstream needs. The cluster placement (document-generation) is correct, and the trigger surface is naturally broad.

The value is currently latent. It becomes real the moment the SKILL.md is expanded to include an install path, a dependency list, a CLI or API entrypoint, an output contract, and explicit error/OCR behavior. Until then, treat this as a promising stub, not a usable skill. We recommend re-running the same three tests — particularly the scanned-PDF edge case — once the full documentation is available, and revising the composite score against observed behavior rather than inferred intent.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-09-21.

Overall: broken. 0 tests passed, 0 partial, 0 failed; key blocker: the provided SKILL.md is truncated to a single sentence with no install command, dependencies, formats, or CLI details, so all three tests were skipped as untestable.

Test	Status	Notes
install	skipped	SKILL.md content is truncated to a single descriptive sentence and contains no install command, dependency list, or version constraints, so no install step could be executed or verified.
minimal-roundtrip	skipped	The SKILL.md only states the skill can summarize or transcribe URLs, videos, podcasts, articles, transcripts, PDFs, and local files; it does not specify a target format, CLI entrypoint, or output-writing behavior to test.
edge-corner	skipped	No OCR capability, PDF handling details, or error behavior is described in the truncated SKILL.md, so a scanned-PDF edge case cannot be meaningfully exercised.