The template package format

Reference for reviewers and hands-on creators. If you're creating through the structured report, you never touch these files — the AI interview builds them for you. Machine-readable schemas live in schemas/.

Package layout

Each simulation is one folder under templates/:

templates/
  your-template-slug/
    template.yaml          # identity: id, title, objective, audience, roles, team size
    steps.yaml             # the ordered steps, human/AI modes, and grading rules
    metadata.yaml          # domain, difficulty, duration, tags, license, learning goals
    creator.yaml           # who you are + royalty preference
    sample_inputs.json     # example inputs for a test run
    expected_outputs.json  # what a good run produces
    README.md              # human-readable description of the simulation
    LICENSE.md             # license text

The reference package to copy from: consumer-complaint-escalation.

template.yaml

Field	Required	Meaning
`id`	yes	Unique snake_case identifier, e.g. `consumer_complaint_escalation`
`title`	yes	Display name
`objective`	yes	The skill or judgment the simulation trains
`audience`	no	Who it's for, e.g. "MBA / enterprise training"
`roles`	no	List of participant roles, e.g. `support_agent`, `manager`
`version`	no	Semver, defaults to `1.0.0`
`team_size`	no	Default team size (integer, default 5)
`ai_chat_enabled`	no	Whether participants get an AI chat assistant (default true)
`group_communication_allowed`	no	Whether team members may talk to each other (default true)

steps.yaml — one entry per step

Field	Required	Meaning
`id`	yes	Unique snake_case step id
`title`	yes	Short step name
`actor_mode`	yes	Who acts — see table below
`task_type`	yes	Shape of the answer — see table below
`task_description`	yes	What the participant is asked to do
`role`	no	Which role performs this step
`depends_on`	no	List of step ids that must complete first (defines the workflow DAG)
`ai_assist`	no	What the AI may do: `suggests` · `critiques` · `grades` · `decides`
`grading`	no	Grading rule block — see below

Actor modes — the human decision boundary

Mode	Who decides
`human`	A person, working alone — no AI involvement
`human_with_ai`	A person decides; AI assists per `ai_assist`
`team_with_ai`	The team decides together; AI assists
`ai`	The AI acts autonomously; humans see the output downstream

Task types

Type	The participant produces…
`multiple_choice`	One choice from fixed options
`numeric_input`	A number (checked against value/range/tolerance)
`structured_response`	A structured answer, e.g. decision + justification fields
`open_response`	Free-form writing
`document_review`	A judgment on a provided document

Grading rules

`type`	Requires	Use when
`answer_key`	`correct_answer`, usually `options` and `distractors`, `points`	One objectively correct answer
`numeric`	`expected_value` or `expected_range`, optional `tolerance` / `formula`	Calculations
`rubric`	`criteria` list (each measurable)	Judgment quality scored by named criteria
`llm_judge`	`criteria` and `anchor_examples` (good/bad answer examples)	Open responses graded by an LLM against anchors
`none`	—	Ungraded steps (context-setting, handoffs)

metadata.yaml & creator.yaml

metadata.yaml carries discovery info: domain, industry, difficulty, duration_minutes, team_size (min/max), scoring_type, license, tags, and learning_goals. creator.yaml identifies you and states your royalty preference: fixed_fee · revenue_share · per_run · open_source · enterprise_only.

Validation

Every pull request runs scripts/validate_package.py over all packages: schema conformance, step-id uniqueness, depends_on references, and grading-rule completeness per the decision tree above. Run it locally before opening a PR:

python scripts/validate_package.py templates/<your-slug>