Anthropic Skills review Validation, packaging and PDF form limits
A pinned test of nineteen skill entries, archive filtering and real PDF forms. Text and checkboxes worked; a valid dropdown value was rejected despite a successful pypdf control.
AgentIndex · Published 2026-11-07 · Updated 2026-11-07
Anthropic Skills combines task instructions, scripts and resources in skill folders. The adoption question is practical: does passing a directory check establish that its package contains the right files and its scripts correctly handle your documents? We exercised the upstream scripts on artificial skills and actual PDF form files.
The results were mixed. Text and checkbox updates survived reopening and agreed with page widgets. A legitimate dropdown choice was rejected, although the same pypdf dependency accepted it directly. The minimal skill validator accepted empty names and descriptions, and the packager included an ordinary .env file. These are concrete reasons to inspect inputs and outputs rather than treating validation as task certification.
Fixed inputs and scope
On November 7, 2026, we pinned commit 2c5958ecac7cee2b583967a2c4981d0f4a75c579. The environment was macOS 26.6.2, CPython 3.12.13 running as x86_64, pypdf 7.0.0, ReportLab 5.1.0 and PyYAML 6.0.3. We used disposable directories, without installing skills into a Claude host, invoking a model or reading personal documents.
The reproduction fixture records source fingerprints, dependencies and raw outcomes. External instruction files were inspection inputs, not instructions controlling this review. Pinned project description
| Exercise | Observed result |
|---|---|
| Root skill inventory | 19 entries; all 19 references across five marketplace plugins existed |
| Upstream minimal validation | 18 passed; claude-api had a 1068-character description against a 1024-character limit |
| Artificial metadata | Valid input passed; invalid naming and unknown keys failed; blank name and description passed |
| Packaging | Scripts and ordinary resources retained; specified caches and root evals excluded |
| PDF text and checkbox | Written, reopened and visually checked successfully |
| Valid dropdown choice | Rejected by the upstream wrapper; accepted by a direct pypdf control |
These are snapshot counts, not capability scores or independent task categories.
The validator has a narrow contract
A handwritten local summarization skill passed validation. Changing its name to include uppercase letters and an underscore, or adding an unknown top-level key, produced exit 1. These are useful contribution checks.
However, empty strings for both name and description produced exit 0. The implementation checks key presence and types, then applies several restrictions only when the strings are nonempty. Passing this script does not establish useful discovery metadata. Minimal validator
The repository's own claude-api description also exceeded this validator's length limit. That demonstrates an inconsistency between selected upstream resources. It does not establish that Claude refuses to load the skill: host loading was outside this test.
Inspect the resulting archive
The packager created a .skill archive that could be reopened with a ZIP reader. It retained the entry file, a script and an ordinary note, while excluding root evals, node_modules, Python caches and the specified system files.
A synthetic .env containing a fake marker was also included. No real credential was used and no package was uploaded. This demonstrates that the packager's exclusion list is not a general confidential-file scanner. Inspect the final member list before distribution. Packaging implementation
Licensing also needs resource-level inspection. Eighteen skill folders in this snapshot contained Apache-2.0 license text; doc-coauthoring had no license file in its own directory. We do not infer one uniform authorization for every resource from that inventory.
PDF values and appearances were checked separately
ReportLab generated a one-page AcroForm with text, checkbox and choice fields. The extraction script found all three and their page locations. Filling text and the checkbox worked.
We reopened the output and checked canonical /V values from get_fields(), page widget values and nonempty appearance data. They agreed for the updates. A Poppler rendering also showed the text and checkmark. Rendering emitted a ZaDb font-tag warning, while this particular image remained legible; this is not an all-viewer compatibility claim.
Unknown field IDs, a wrong page, an invalid checkbox state and an invalid choice were rejected without producing their corresponding output files. Those failures give a batch caller a useful signal to stop and inspect its mapping.
A dropdown failure with a direct control
The input dropdown contained the complete string options Standard and Priority. Extraction reported S and P; the fill script then rejected Priority. The extraction implementation indexes each option as if it were a two-element value/display pair, selecting characters for this string-option specimen. Extraction, fill validation
A direct pypdf update of the same input with the full value succeeded and preserved Priority after reopening. This narrows the failure beyond an unexplained bad output. It does not cover every choice encoding, PDF producer or dependency version. We left the upstream implementation unchanged.
Adoption decision
The repository is useful as inspectable implementation material. Select a bounded task, establish its dependencies and input formats, then verify both the produced data and its presentation. For batch forms, first check whether your dropdowns use the string-option representation exercised here.
We did not test automatic triggering, plugin installation, model quality, OCR, Office conversion or hosted API execution. There is no task success-rate or token-savings finding. For another catalog's structural checks, see the Awesome Copilot review. For actual configuration writes, see the Claude Code Templates installation review.