Recipe Engine
The Recipe Engine turns one-off analysis and generation work into repeatable workflows. It has four building blocks:
| Concept | What it is |
|---|---|
| Skill | A reusable analysis or generation unit: prompts, expected inputs and output artifacts, model settings, and validation criteria. |
| Recipe | A workflow template that chains published skills into ordered steps with references and approval gates. |
| Reference pack | A bundle of curated context documents (architecture guidelines, domain catalog, endpoint catalog, security patterns, and similar) that steps can read and update. |
| Execution | A single run of a recipe against a repository and selected code symbols. |
Open the Recipe Engine from Agentic → Chat → Personalize and choose Skills, Recipes, Reference Packs, or Recipe executions.
When to use
- Standardize a multi-step modernization task so it runs the same way every time.
- Apply the same analysis across many code symbols in one repository.
- Keep human review gates in a workflow that would otherwise be manual.
Skills
Personalize → Skills lists your skills with their status (draft, active, deprecated).
Create a skill with New skill or Create with AI. A skill defines:
- Identity: name, slug, version, description, category, and tags.
- Inputs and output artifacts.
- Agent prompts: system prompt and user prompt template.
- Model configuration: temperature and maximum output tokens.
- Validation criteria: a judge mode (heuristic, LLM, or hybrid) and one or more criteria.
- Human review and retry: whether a step needs approval, whether it auto-approves when all criteria pass, whether editing is allowed before approval, and the maximum number of retries.
Only published skills can be used in recipes. You can revert a skill to draft or deprecate it.
Recipes
Personalize → Recipes lists workflow templates. Create one with New recipe or Create with AI (which assembles a draft from your published skills).
A recipe defines identity and classification, required and optional references, execution defaults (output language, maximum context tokens, whether eligible steps run in parallel, auto-approve behavior, export format), and an ordered list of steps. Each step selects a skill and can declare a condition, the references it needs, input mappings, whether it requires human approval, and whether it creates or updates a reference document.
Reference packs
Personalize → Reference Packs manages the context documents executions rely on. For each pack you can add documents (Markdown, SQL, YAML, JSON, or text), view them rendered or raw, download the pack as a ZIP, and review a change history showing every version an execution created or updated.
Executions
Personalize → Recipe executions monitors runs. Start one with New execution and follow four steps:
- Recipe — choose the workflow to run.
- Target — choose a documented repository and the symbols to process. Pick an execution mode:
- Single symbol — one execution for one symbol.
- Unified context — several symbols loaded into one execution.
- Parallel batch — a separate execution for every selected symbol.
- Context — attach a reference pack and optional JSON configuration overrides. The wizard flags missing required reference types.
- Review — confirm the setup and start.
An execution shows per-step status (pending, running, awaiting approval, approved, rejected, skipped, completed, failed), the validation (judge) result and evidence, generated artifacts, and any reference-pack changes it made. At a review gate you can Approve, Reject and regenerate with feedback, Skip, or Edit and approve the artifact. Finished executions can be exported as a ZIP.
Expected result
A library of published skills and recipes, reference packs with a tracked history, and executions that run your workflow consistently with review gates where you need them.