| name | dillylang-translate |
| description | Translate a Claude Code skill directory into a Dillylang recipe description — deterministic ingestion, 5-step recipe runner pipeline. Use when the user wants to translate, convert, express, or describe a skill in Dillylang vocabulary. |
Translate Skill to Dillylang
When to use
- "translate this skill to Dillylang"
- "what's the Dillylang recipe for [skill]?"
- "express [skill] in Dillylang vocabulary"
Execution
Phase 1 — Ingest (deterministic, no LLM)
Read the skill directory and build a JSON source bundle:
- Read SKILL.md. If absent, stop.
- Follow external references —
<invoke> tags, script paths, module imports.
Read the implementation files where the skill's actual logic lives.
- Read remaining files (README, prompts/, scripts/, configs). Cap at 20 files / ~50KB.
- Detect entrypoints, tool use patterns, and control flow.
- Format as JSON string for
skill_source_bundle.
Phase 2 — Recipe runner pipeline
Start the recipe runner with the ingested bundle:
python3 -m skills.scripts.recipe_runner start \
--recipe skills/scripts/recipes/translate.py \
--input '{"skill_source_bundle": "<ingested JSON>"}'
Follow the runner's emit/submit protocol for all 5 steps. The recipe file
(skills/scripts/recipes/translate.py) contains all prompts and operator
lineage. Each step's output is arbitrary JSON matching the schema described
in that step's prompt.
Reflection
After the pipeline completes, the synthesize output IS the product. Add
2–3 sentences: which steps produced sharpest signal, where classify and
evaluate agreed vs. diverged, and whether residue suggests vocabulary gaps
worth tracking.
