Skip to content

Latest commit

 

History

History
66 lines (48 loc) · 2.16 KB

File metadata and controls

66 lines (48 loc) · 2.16 KB

Convert an existing prose skill

Use conversion after you have run one ordinary Yield workflow. Conversion is a separate job from the runtime itself: the converter helps extract and encode a policy; Yield then executes and verifies the resulting program.

The repository includes examples/convert-skill, a converter that is itself a Yield skill.

What moves, and what stays

Keep these in the thin SKILL.md:

  • the goal and when the skill should be used;
  • domain context the model needs;
  • judgment criteria and useful examples;
  • the instruction to start and resume the Yield program.

Move these into the program:

  • required order;
  • branches and retry limits;
  • commands whose output must be observed;
  • approval points;
  • claims required for completion;
  • Blocked and Refused outcomes.

Run the converter

From a checkout of the public repository:

go build -o /tmp/yskill ./cmd/yskill
cd examples/convert-skill
YSKILL=/tmp/yskill /tmp/yskill run .

The workflow asks for:

  1. the source directory containing SKILL.md;
  2. a target language;
  3. a destination directory.

The coding agent extracts the implicit flow and writes the generated files. The converter then runs the generated skill's own fixtures under yskill test. It allows two bounded repair attempts and returns Blocked if the fixture run still fails.

What “verified” means here

The converter verifies that the generated program executes its declared fixture path. It does not prove that the extracted policy is behaviorally equivalent to every reading of the original prose.

Review the conversion as a policy change:

  • map every load-bearing source instruction to code, retained model judgment, or an explicit exclusion;
  • add a positive fixture and at least one negative or refusal path;
  • test bypass attempts and failure states;
  • replay a completed run to check determinism;
  • keep performance or token-reduction claims separate from runtime correctness.

The implemented safety comparison is documented in locus-converter.md. Its narrow conclusion is that the converter cannot report success before the generated fixture run passes.