Bundled skills: one description over 1024 characters, three reserved-word names, fifteen guideline-level findings

[kasparovabi]

I went through the 22 skills under .claude/skills/ against the Agent Skills best-practices page and found four things that conflict with a documented limit, plus fifteen that are guideline-level. I opened each file and counted by hand before writing this, so the numbers below are measured, not inferred.

Conflicts with a documented limit

description past the 1024-character maximum, one file.

.claude/skills/claude-api/SKILL.md has a description of 1068 characters. The documented ceiling is 1024. I have not reproduced a hard load failure, so I am describing this as past the documented limit rather than broken.

Reserved word in the skill name, three files.

The naming rules say a skill name should not contain "anthropic" or "claude". These three do:

File	name:
.claude/skills/claude-api/SKILL.md	claude-api
.claude/skills/claude-md-improver/SKILL.md	claude-md-improver
.claude/skills/claude-automation-recommender/SKILL.md	claude-automation-recommender

Two caveats worth stating before you act on this. Renaming a skill changes the name it is invoked by, so it is a breaking change for anyone who has already wired these into a workflow. And claude-automation-recommender matches a skill Anthropic itself distributes under that exact name, which suggests the restriction is not enforced in practice today. Your call entirely, and leaving them alone is a defensible answer.

Guideline-level

SKILL.md body over the 500-line guideline, five files.

Everything after the frontmatter loads into context the moment the skill triggers, which is the reason for the guideline.

File	Body lines
.claude/skills/django-patterns/SKILL.md	729
.claude/skills/django-tdd/SKILL.md	724
.claude/skills/django-security/SKILL.md	588
.claude/skills/claude-api/SKILL.md	540
.claude/skills/api-design/SKILL.md	518

The usual fix is to move the long reference material into references/ and leave the procedure in the body.

Reference file over 100 lines with no table of contents, ten files.

The docs ask for a table of contents past 100 lines so a partial read still shows the full scope. The longest ones:

File	Lines
.claude/skills/mcp-builder/reference/node_mcp_server.md	970
.claude/skills/mcp-builder/reference/python_mcp_server.md	719
.claude/skills/mcp-builder/reference/evaluation.md	602
.claude/skills/claude-automation-recommender/references/skills-reference.md	408
.claude/skills/claude-automation-recommender/references/mcp-servers.md	263

The remaining five are mcp_best_practices.md (249), claude-md-improver/references/templates.md (253), claude-automation-recommender/references/hooks-patterns.md (226) and subagent-templates.md (181), and claude-md-improver/references/quality-criteria.md (109).

Several of these skills look vendored from upstream collections, mcp-builder and the document skills in particular. Where that is true the fix belongs upstream and tracking it here may be all that makes sense.

One unrelated thing while I was in there

.claude/skills/mcp-builder/scripts/evaluation.py:223 defaults to claude-3-7-sonnet-20250219, which reached its retirement date on 19 February 2026 per the deprecation page. Anyone running that script without passing --model gets the retired id. I have not run it against the API to see what comes back, so the claim rests on the published retirement date and nothing more.

---

I read every file named above before writing this, and I am happy to be wrong about any of it. What brought me here was a checker I maintain that validates published skill packages against the Anthropic guidance, so I am mentioning it up front rather than leaving you to wonder. If you would rather not get this kind of report, say so and I will not send another.