Create a simulation from a structured report

You don't write YAML, code, or config files. You write four short sections about a workflow you know well. If you can explain the job to a new hire, you can create a simulation.

The whole process: fill in the report template (four sections below) → hand it to an AI agent connected to SkillSimm → the agent interviews you for anything missing, builds the formal template package, and submits it for review.

1What the workflow looks like

Describe the scenario and the sequence of work in plain language: who is involved, what kicks the workflow off, and what happens step by step. Think of it as narrating the job to someone shadowing you for a day. Aim for 5–12 steps.

Example A customer files an angry complaint about delayed onboarding. A support agent reads the ticket and diagnoses the core issue. If the customer is high-value or threatens to churn, the agent drafts an escalation summary for a manager. The manager decides the escalation path — refund, service credit, or executive callback — and the agent sends the final response.

2Materials needed

List every document, dataset, image, or reference a participant must read to do the work: the ticket, the policy manual, the financial spreadsheet, the org chart. For each one, say what it is and whether you can provide it or need help creating a realistic version. Materials make or break realism — vague scenarios produce vague decisions.

Example (1) The complaint email thread, 3 messages, escalating tone. (2) The customer's account summary: plan tier, contract value, support history. (3) A one-page escalation policy stating refund limits per role. I can write #1 and #3; I need a realistic-looking #2 generated.

3Human decision boundaries & standard operating procedure

This is the heart of a SkillSimm simulation. For each step, say who acts and where the human boundary sits:

Which steps must a human decide? These are the decisions the simulation exists to practice.
What may the AI do at each step? Suggest options, critique a draft, grade an answer, or decide outright.
What is the standard operating procedure? The rules a competent employee follows: thresholds, approval limits, escalation criteria, required checks before acting.

Example Diagnosing the complaint: human decides, AI may suggest candidate issues. Escalation decision: human decides alone — this is the graded judgment call. SOP: refunds under $200 need no approval; $200–$1,000 requires manager sign-off; any churn threat from an enterprise account escalates to the account executive within 4 business hours, no exceptions. Drafting the response email: AI drafts, human edits and approves before sending.

4Edge cases

What makes this workflow genuinely hard? List the tricky inputs, ambiguities, and failure modes a strong performer handles and a weak one misses. Good edge cases are what separate a training simulation from a quiz — including deliberately planted ambiguities participants should catch.

Example The customer's tone is angry but their actual ask is small (a $40 credit) — over-escalating wastes a manager's time. The account summary shows the contract renews in 3 weeks, which changes the risk calculus but is never mentioned in the complaint. The policy has no rule for customers who threaten to post on social media — participants must reason from principles.

Get the report template

Copy the blank report below, or download it as a Markdown file. Fill in each section — bullet points are fine.

# SkillSimm Simulation Report

## 0. Basics
- Working title:
- Who is this for (audience):
- What skill or judgment should it train (objective):
- Roles involved (e.g., support agent, manager):
- Solo or team simulation? Ideal team size:

## 1. What the workflow looks like
(Narrate the scenario and steps in order. Who does what, triggered by what?)

## 2. Materials needed
(Every document, dataset, or image a participant must read.
For each: what it is, and whether you can provide it or need it created.)

## 3. Human decision boundaries & standard operating procedure
(For each step: does a human decide, or the AI? If AI is involved, does it
suggest, critique, grade, or decide? What rules, thresholds, and approval
limits does a competent employee follow?)

## 4. Edge cases
(Tricky inputs, ambiguities, and failure modes. What do strong performers
catch that weak ones miss? Any deliberately planted traps?)

## 5. Grading (optional — the interview will help)
(How would you know a participant did well? Right answers, a rubric,
numeric targets?)

Submit it — no code, no git

Open Claude (or any AI agent that supports MCP connectors), connect the SkillSimm template server at https://mcp.skillsimm.com/mcp with your SkillSimm API token, then paste your finished report with this prompt:

Here is my structured report for a new SkillSimm simulation. Please:
1. Interview me about anything missing or ambiguous — one question at a time.
2. Build the full template package (steps, human/AI modes, materials, grading).
3. Show me a preview and let me revise it.
4. When I approve, validate it and submit it for review.

[paste your report here]

The agent walks you through the standard interview (goal → audience → scenario → roles → steps → human/AI mode → grading), assembles the formal package, and opens the review submission for you.

Comfortable with GitHub? You can also build the package by hand and open a pull request directly — see the contributing guide and the package format.

What happens after you submit

Automated checks validate the package structure, step logic, and grading rules the moment the submission opens.
A reviewer reads it for instruction clarity, sound step logic, honest grading, safety, and originality — and leaves comments if anything needs work.
You revise through the same AI agent (or directly on the pull request) until the reviewer approves.
A test run confirms the simulation produces a meaningful evaluation report.
Royalty terms are settled (revenue share, per-run, or open source — your choice, stated in your creator profile) and the template is published to the gallery with your name on it.

Nothing publishes without human review. The AI creates, GitHub reviews, SkillSimm tests, you agree, an admin publishes — in that order.