Validate Skill Spec

Validate Skill Spec

Validates a skill directory against the Agent Skills specification (agentskills.io/specification). Checks file reference consistency, frontmatter compliance, progressive disclosure, and structural correctness.

This skill provides a manual validation workflow that complements automated validation scripts. It covers checks that scripts may not fully enforce — particularly file reference consistency per the spec's file-references rules.

When to Use

• After adding, renaming, or removing files in a skill directory
• After editing file references in SKILL.md or any referenced file
• After modifying frontmatter fields
• Before a release or distribution
• When unsure if a structural change broke something

Step 1: Identify the Skill Directory

Determine the skill directory to validate. The skill root is the directory containing SKILL.md. Cross-file references resolve file-relative (standard markdown semantics) under the redefined path-resolution rule — every link resolves from the directory containing the file the link lives in. The skill root and each capability root own their own subgraph; capability bodies reach the shared skill root via the explicit ../../<dir>/<file> form. The full rule lives in the foundry skill's reference document at skill-system-foundry/references/path-resolution.md.

If validation scripts are available (e.g., validate_skill.py from this repository), run them from within the skill-system-foundry/ directory:

cd skill-system-foundry
python scripts/validate_skill.py <skill-path> --verbose

If the skill uses nested references intentionally, add --allow-nested-references. If the scripts report failures, fix them before proceeding to the manual checks below.

Step 2: Verify Frontmatter Compliance

Read the SKILL.md and check its YAML frontmatter against the spec:

Check each required field against its constraints. For optional fields, verify they meet their limits if present.

Step 3: Verify File Reference Consistency

The spec defines file reference rules at agentskills.io/specification#file-references. Every file reference in the skill must comply.

3a: Extract all file references

Read SKILL.md and extract every reference to a file:

• Markdown links — [text](path) where path is not a URL or anchor
• Backtick-quoted paths — paths in backticks that contain / and point to skill-internal files

Filter out:

• URLs (starting with http://, https://, or protocol-relative //)
• Anchor-only links (#section)
• Template placeholders (containing < or >)

3b: Check that every referenced file exists

For each extracted reference, resolve it file-relative (from the directory containing the file the link lives in) per the redefined path-resolution rule. Strip any fragment (#section), query string, or title annotation before resolving.

Verify that each resolved path:

• Points to an existing file (not a directory)
• Is readable (no encoding or permission errors)

Report any reference that does not resolve to an existing file.

3c: Check that every bundled file is referenced

List all files in the skill's subdirectories (references/, scripts/, assets/, and any other directories). Compare against the references extracted from SKILL.md.

An unreferenced file means agents do not know it exists and cannot load it on demand. Every bundled file should be linked from SKILL.md with a description of what it contains and when to use it.

Report any file that exists in the skill directory but is not referenced from SKILL.md.

3d: Verify path format

All file references must comply with these rules:

• File-relative resolution — paths must be relative, resolved from the directory containing the file the link lives in (standard markdown semantics)
• No absolute paths — any path starting with / or a drive letter (e.g., C:\) is a violation
• .. segments are legal — they are how a capability reaches the shared skill root (../../<dir>/<file>). Paths whose .. chain escapes the skill root entirely are out of scope and surfaced as INFO, not failures
• Forward slashes only — all paths must use /, not \. This is a cross-platform portability requirement

3e: Verify reference depth

The spec says: "Keep file references one level deep from SKILL.md. Avoid deeply nested reference chains."

• Level 1 — SKILL.md references files in references/, scripts/, assets/ (valid)
• Level 2 — a referenced file itself contains references to other files (nested — should be flagged unless intentionally allowed)

For each file referenced from SKILL.md, read it and check whether it contains its own file references. If it does, report the nesting. Deep reference chains risk partial reads by agents, breaking progressive disclosure.

Step 4: Verify Content Structure

Check that the skill follows progressive disclosure:

Line count

SKILL.md should stay under 500 lines. Count the lines and report if the limit is approached or exceeded. When approaching the limit, detailed material should be moved to references/ files.

Reference files over 100 lines

Any referenced file over 100 lines should include a table of contents at the top to help agents navigate.

Description quality

The description field is the primary trigger mechanism. Check:

• Does it describe what the skill does? (concrete verbs and nouns)
• Does it describe when to trigger? (user intent phrases and keywords)
• Is it specific enough to distinguish from related skills?
• Is it keyword-rich, erring toward over-triggering rather than under-triggering?

Directory structure

The skill may contain these optional directories:

• scripts/ — executable code
• references/ — additional documentation
• assets/ — templates and static resources

Verify that files are organized in the appropriate directories. Scripts should not live in references/, documentation should not live in scripts/.

Step 5: Report Results

Structure the output:

## Skill Validation Report

**Skill:** <skill-name>
**Path:** <skill-directory-path>

### Frontmatter
- name: (valid/invalid — details)
- description: (valid/invalid — length, quality assessment)
- optional fields: (list any issues)

### File Reference Consistency
- Missing targets: (list or "None")
- Unreferenced files: (list or "None")
- Invalid paths: (list or "None")
- Forward-slash violations: (list or "None")
- Nested references: (list or "None — intentional/unintentional")

### Content Structure
- Line count: X/500
- Files over 100 lines without ToC: (list or "None")
- Description quality: (assessment)

### Overall
(Clean / Issues found — with summary of what to fix)