Distill Genes, Skills, and Recipes
Turn completed work into reusable capability by retaining original evidence, extracting the method, and choosing a deliverable. This guide uses a narrow-screen overflow fix. To consume existing experience, see Recipes and experience reuse.
Choose the deliverable
| Deliverable | What it retains | Example |
|---|---|---|
| Gene | Signals, prerequisites, strategy, and validation requirements | Locate and fix overflow through actual layout constraints |
| Skill | Agent triggers, work steps, and tool conventions | Read, reproduce, and deliver an interface regression task |
| Capsule | One execution, its outcome, and evidence | A project's fix record and the Gene used |
| Recipe | Ordered orchestration of existing Genes/Capsules | Reproduction, repair, regression checking, and result preparation |
One task can produce several objects; changing extensions or copying a summary does not create them.
Extract a method from the task
Retain the request, screenshots, before/after diff, commands, verification output, and failed attempts. Identify the trigger, effective step, failed approaches, verification coverage, and prerequisites in another project.
The example method reproduces the reported viewport, identifies the element and width constraint, fixes the cause, then checks menus and desktop layout. Project filenames, accounts, and private content remain in local records; the reusable strategy must not depend on secrets.
Create a Gene candidate
This command supplies signals and strategy to standalone Evolver's distillation entry:
evolver distill --category repair --signals horizontal_overflow --strategy "Reproduce overflow; locate the width constraint; fix the cause; verify mobile interaction and desktop layout"
Inspect the candidate ID and content, then complete preconditions, constraints, and validation requirements. See GEP schemas for fields and formal validation. Candidate creation, approval, successful execution, and publishing are separate steps; creation provides no success evidence.
Author a Skill and Recipe
Use Build skills to define triggers, inputs, steps, and capabilities, then test normal, missing-input, and out-of-scope requests. A Recipe orchestrates existing assets; a SKILL.md path is not an asset ID.
For a two-step “diagnose overflow → verify the repair” Recipe, prepare two approved Genes and record their content IDs, prerequisites, outputs, and validation. Both assets must exist in the target Hub with matching types. Local presence does not establish remote presence.
Replace both IDs with real values and build in dependency order:
evolver recipe build --title "Repair and verify horizontal overflow" --genes "sha256:<diagnosis-gene-hash>,sha256:<verification-gene-hash>" --description "Diagnose the width constraint, then verify mobile interaction and desktop layout" --price 1
This command creates a DRAFT on the Hub, not an offline preview. --price 1 sets the execution price; it grants no purchase or execution consent. Omit --publish for ordinary draft creation, and obtain creation authorization and Hub identity first.
| Checkpoint | Required check |
|---|---|
| Step input | Full sha256 content IDs, correct dependency order, at most 20 steps |
| Type | CLI resolves local Gene/Capsule lists; an unknown ID defaults to Gene, so a Capsule must be locally identifiable |
| Create receipt | Status draft; retain the returned Recipe ID, separately from Gene hashes |
| Publication | Review steps, types, and price; use --publish only after a separate publishing decision |
| Create timeout | Reconcile the original operation and Hub record; the CLI derives an idempotency key from content, so preserve the complete original request, including title, description, price, assets, and step order; reconcile an unknown result before changing any field or retrying |
--assets is an alias for --genes. Neither flag publishes missing step assets automatically. For conditions, parameters, and input/output schemas, use the Recipe create payload in GEP schemas and the corresponding Hub interface; do not invent CLI flags.
Skill-based authoring uses evolver recipe from-skills --manifest <file>. It requires ordered steps with execution evidence, not a list of instruction files. Prefer build above for ordinary asset composition. Execution, leases, and outcome reporting belong in Recipes and experience reuse.
Test reuse in another task
Choose another task with matching signals and different file organization. Check whether the method still guides diagnosis and requires hidden manual assistance. Retain failures and exclusions as preconditions or counterexamples.
Acceptance includes original evidence, extracted objects, formal validation, reuse records, and applicability. A summary or model self-assessment alone does not complete distillation.