# Anthropic Compatibility Assessment‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌⁠⁠‍‍‍⁠⁠‍‌‍⁠‌‌‍‌‌‍‌‍‍‌​‍⁠​‌‍‍‌⁠⁠​‌⁠​‌⁠‌​‌​‌‍​

Documentation checked: 2026-09-15.

Scope: Claude custom skills and Claude Code's common Agent Skills format. The pack deliberately avoids product-specific permissions, hooks, model routing and tool identifiers so the same skill directories can be used across those surfaces.

## Verified format checks‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌⁠⁠‍‍‍⁠⁠‍‌‍⁠‌‌‍‌‌‍‌‍‍‌​‍⁠​‌‍‍‌⁠⁠​‌⁠​‌⁠‌​‌​‌‍​

| Requirement or recommendation | Implementation | Verification |
| --- | --- | --- |
| Named skill directory containing SKILL.md | One directory per skill; exact filename | Automated |
| Required YAML name and description | Exactly those two scalar fields | Automated plus native skill validator |
| Name length and characters | Lowercase ASCII kebab-case, at most 64 characters | Automated |
| Reserved names avoided | No anthropic, claude or synced skill names | Automated |
| Description is nonempty and at most 1,024 characters | Capability plus specific activation context | Automated; wording reviewed |
| Third-person discovery description | Describes the capability, not “I can” or “you should” | Manual |
| Lean entrypoint | All SKILL.md bodies under 500 lines | Automated |
| Progressive disclosure | Detailed methods/examples/provenance linked directly | Link checks plus manual review |
| Safe relative paths | Resources resolved from the actual skill directory | Link checks and instructions |
| Clear degrees of freedom | Judgment stays flexible; arithmetic and invariants are explicit | Manual |
| Executable helpers work | Five offline standard-library helpers | 45 tests plus CLI checks |
| Single-skill ZIP layout | Exactly one root folder named after the skill | Archive checks |
| No unnecessary product configuration | No OpenAI metadata, hooks or install scripts in exported skills | Packaging filter |
| License stated for redistribution | CC BY 4.0 for pack contents, with cited research excluded | LICENSE |
| Evaluation and iteration | Seeded tests, limited fresh-agent trials and a per-case model-test rubric set | See validation report |

## Sources and interpretation

[Anthropic's authoring best practices](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices) supplies naming, metadata, progressive-disclosure and evaluation guidance. The under-500-line body guidance is a recommendation, not a universal parser limit.

[Claude custom skill creation](https://support.claude.com/en/articles/12512198-how-to-create-custom-skills) specifies the upload-folder packaging. The outer ZIP is a distribution bundle; inner ZIPs are the actual single-skill upload units.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌⁠⁠‍‍‍⁠⁠‍‌‍⁠‌‌‍‌‌‍‌‍‍‌​‍⁠​‌‍‍‌⁠⁠​‌⁠​‌⁠‌​‌​‌‍​

[Claude Code skills documentation](https://code.claude.com/docs/en/skills) documents personal and project installation locations. Copy the skill folder, not its loose files.

## Explicitly unverified

- Upload acceptance in the user's live Claude account.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌⁠⁠‍‍‍⁠⁠‍‌‍⁠‌‌‍‌‌‍‌‍‍‌​‍⁠​‌‍‍‌⁠⁠​‌⁠​‌⁠‌​‌​‌‍​
- Trigger selection in any particular Claude model/version.
- Behavior on every case in the supplied evaluation set.
- Comparative accuracy versus no skill on Claude.‌‌​⁠‌​‌⁠‌​⁠​​⁠​‌​​​‌⁠⁠‍‍‍⁠⁠‍‌‍⁠‌‌‍‌‌‍‌‍‍‌​‍⁠​‌‍‍‌⁠⁠​‌⁠​‌⁠‌​‌​‌‍​
- Prospective real-world calibration, sharpness or superiority to expert forecasters.

The structural validator is custom, not Anthropic's own validation service. Runtime and behavioral claims require testing on the actual target. “100% aligned” can responsibly mean every listed structural check passes; it cannot mean a guarantee that subjective best practices or model behavior are perfectly satisfied for all future requests.

## Why these are multiple skills

Each technique has a discriminating trigger and an independent operating procedure. Detailed theory is not duplicated in all entrypoints. The workflow orchestrates by need and has its own standalone path, avoiding a requirement to load thirteen skills for a simple calculation.