Skill Atlas

Research protocol

Evidence before adjectives.

Skill Atlas separates what can be read from an artifact from what must be observed in a controlled run. The distinction shapes every dossier.

The unit of research is a fixed source snapshot.

Public repository files are collected as untrusted data and analyzed without execution. Records point to an owner, path, commit, license, and collection time. A changed upstream skill is a new reading, not a silent rewrite of the old one.

From source to dossier

  1. 01

    Collect

    Read public source files at a fixed commit. Preserve provenance and upstream license information.

  2. 02

    Inspect

    Parse frontmatter, inventory files, and identify structural and security-relevant signals without running downloaded code.

  3. 03

    Interpret

    Describe capabilities and a provisional workflow. Mark the analysis basis so interpretation is not confused with observation.

  4. 04

    Measure

    Only an explicit official token-counting run with a named model produces exact token values. Otherwise the field stays unavailable.

  5. 05

    Evaluate

    Controlled tasks require a recorded model, environment, baseline, and outcome. No fabricated tasks, passes, or performance comparisons.

The evidence boundary

Review status and evaluation status are independent. A source can be statically inspected without being executed; a controlled evaluation must identify its own conditions.

Static review can establish

  • Which files and instructions ship at a commit
  • Whether scripts, network signals, writes, or credential references appear
  • The apparent operating sequence and installation surface
  • Exact token counts only when measured with the specified official method

Static review cannot establish

  • That a skill is safe to run in your environment
  • That it completes a task correctly or consistently
  • That it improves quality, speed, or token efficiency
  • That an unobserved risk surface does not exist

Token layers are not a single price.

Discovery metadata and activated instructions enter context at different times. Bundled resources may be loaded selectively. We show each measured layer with its method and model, and never fill unavailable values with file-size estimates.

Evaluation is reported, not inferred.

Where controlled tests exist, the dossier lists tasks, baselines, outcomes, and available measurements. Where they do not, the interface says so plainly. A count of recorded tasks is coverage—not a universal benchmark.

Security findings are review prompts.

A network call, file write, or credential reference is a reason to inspect context and permissions, not an automatic verdict. Conversely, ‘no findings’ means no findings under the stated static method, never a guarantee.

Known limits

Source snapshots can age. Generated summaries and workflows may be provisional. Repository popularity does not measure skill quality. Token use and task behavior vary by model, prompt, environment, and resource-loading path.

Inspect the catalog