trace-simulation

name	trace-simulation
description	Produce low-quality, low-information end-user pressure scenarios and candidate trace packages for target role evaluation.
user-invocable	true
auto-invocable	true

Trace Simulation

Use this skill when the user asks UserCat to create, shape, simulate, or run realistic user traces for a target XiaoBa role.

Purpose

Generate candidate trace data from a low-quality end-user perspective, not benchmark decisions. The output is useful only when it preserves the separation:

UserCat creates candidate trace data.
InspectorCat diagnoses, groups, extracts Replay Cases, and routes.
ReviewerCat only executes a Replay Case after Inspector routing.
Benchmark harness owns fixture, verifier, replay, baseline, and release gates.

user_trace_run is the UserCat runtime tool for live target-role dialogue. In adaptive mode it sends the opening user message through the native Dashboard Chat/Pet entrypoint, reads the target role's visible reply, then decides the next low-information user message before continuing. Product traces and visible history land in the normal pet/chat locations; it does not judge the result.

Workflow

1. Establish the seed:
   • target_role
   • seed source
   • task summary
   • risk tags
   • privacy review requirement
2. Read the target role contract:
   • roles/<target-role>/role.json
   • roles/<target-role>/prompts/*
   • roles/<target-role>/skills/*
   • shared roles/README.md and docs/roles-skills/{SPEC,PLAN}.md
   • relevant eval or benchmark evidence when available
3. Write the role intent map before scenario design.
4. Design a low-quality, low-information end-user persona.
5. Design a 3-6 turn adaptive scenario:
   • vague opening;
   • one evidence challenge;
   • one missing detail or changed constraint;
   • one boundary or side-effect question;
   • a stop condition.
6. Produce an opening user turn plus fallback pressures for adaptive live dialogue.
7. If the user asks to run the trace for real, call user_trace_run with:
   • default entrypoint: dashboard_chat
   • default interaction_mode: adaptive for Arena / live exploration
   • target_role
   • seed
   • role_intent_map
   • persona
   • scenario_plan
   • messages as opening / fallback user pressures
8. Produce candidate case metadata:
   • capability tags;
   • expected artifacts;
   • verifier candidates;
   • replay readiness;
   • known gaps.
9. Produce a trace-quality self-check.

Quality Bar

A good candidate trace:

• comes from a real seed, failure log, eval gap, or explicit manual template;
• tests the target role's reason to exist;
• uses natural short user turns;
• is under-specified at the start;
• avoids developer-grade reproduction steps, architecture diagnosis, test plans, or fixes unless the seed explicitly requires a developer persona;
• includes at least one evidence demand;
• includes at least one misunderstanding, missing detail, or mid-course constraint;
• creates observable behavior, artifact, blocked reason, or failure;
• stays local until privacy review.

Reject or mark weak traces that:

• are single-turn prompts;
• are too helpful or too technical;
• ask the target role to do everything perfectly in one message;
• have no possible verifier;
• contain private data that cannot be sanitized;
• let UserCat decide pass/fail.

Output Template

target_role:
seed:
risk_tags:
privacy_review_required:

role_intent_map:
- role_exists_to:
- user_pain:
- must_demonstrate:
- must_not_do:
- fake_success_patterns:
- conversation_pressures:

persona:
- background:
- knows:
- does_not_know:
- temperament:

scenario_plan:
- opening_message:
- turn_plan:
- stop_conditions:

candidate_user_turns:
1.
2.
3.

candidate_case_metadata:
- capability_tags:
- expected_artifacts:
- verifier_candidates:
- replay_readiness:
- known_gaps:

trace_quality_self_check:
- covers_role_intent:
- realistic_low_information_user:
- multi_turn_pressure:
- observable_behavior:
- privacy_safe:
- worth_inspector_intake:

recommended_next_owner: