Imported from sumitake/agent-collab (plugins/agent-collab/skills/simulate-user/SKILL.md). Install upstream with npx skills add sumitake/agent-collab --skill simulate-user. Copyright stays with the author.

Unified runtime invocation

Resolve the plugin root from this loaded file: SKILL.md is at <plugin-root>/skills/<skill-name>/SKILL.md. Invoke only python3 "<plugin-root>/coordinator.py" and send one bounded JSON routing request on EOF-delimited stdin, without a PTY. Use the Python invocation example in the Routing request section in <plugin-root>/README.md and the co-packaged manifest's signed wire_contract; never invent fields or provider actions. Supply one caller-defined work unit per independently useful deliverable, with this skill's logical action and a bounded opaque payload. Use depends_on only for actual dependencies. Honor an operator-named provider with explicit_target. For an authorized independent review or governance task without an operator-named provider, also use that field to bind the caller-verified distinct reviewer selected by the caller or designated by the workflow. Carry the same target into planning and live dispatch; verify returned native lineage before accepting independence. Otherwise use normal untargeted routing. Choose quality and effort for the workload; include context/output token estimates when known. Read the current manifest digest and actual cwd device/inode; do not copy example values. The runtime owns its timeout; do not wrap it in a shorter fixed timeout. Repository identity, source-head verification, disposable copies, patch capture, and cleanup remain caller-owned where applicable. The shim runs standalone from the installed plugin and transports the routing client's bounded result without semantic interpretation. Never discover a provider executable, reconstruct a raw command, or replay, retry, or fail over a consumed work unit. Provider status, terminal records, receipts, telemetry, and other structured fields are optional diagnostics; none is a content-availability gate. Preserve every returned content record or recovered partial response and interpret it with ordinary model reasoning. Never synthesize approval, authority, or a receipt from process exit or missing diagnostics. A planning-only request sets dispatch_requested=false; a live request sets it true and consumes at most one provider attempt per work unit. Planning reports route eligibility, not live availability or authentication. Report a caller/client failure at that layer; provider state remains unknown unless native evidence establishes it. Content availability and each work unit's execution_status are separate facts.

Simulate-user — in-character persona reaction to an artifact

Unlike second-opinion or code-review, this skill is not an analytical critique. It is a strict in-character roleplay: the verifier becomes a specific persona reading the artifact and reacts as that persona would — confused, impatient, suspicious, enthusiastic, missing the ask, snagging on a phrase — without breaking character to deliver a "review." The point is to see how the artifact lands with the intended audience, not how a model thinks about the artifact in the abstract.

This is collaborative roleplay, not a governance review. A selected simulation partner may inhabit a persona or offer a reaction the primary would not. No family difference is assumed or required; same-family and unknown-lineage responses remain available as clearly labelled advisory roleplay, not independent governance evidence.

When to use

• The user explicitly asks for it — "simulate a skeptical engineer," "simulate an impatient board member," "how would a non-technical user react to this," "test this pitch on an executive," "play a confused customer reading this," "roleplay a skeptical engineer responding to this proposal," "what would a compliance officer say about this," "have the reviewer pretend to be a specific persona."
• A draft is about to ship to a specific audience — email, pitch, deck, marketing copy, instructional text, policy memo, internal announcement — and the framing matters.
• The audience is materially different from the author's defaults — non-technical readers, time-pressed executives, skeptical engineers, hostile reviewers, regulated-industry compliance, non-native-language readers.
• A UI flow or set of instructions needs a "first-time user with no context" pass to find where users will get stuck.

When to skip

• The artifact has no specific audience — internal notes, draft brainstorms, things the user is thinking through privately.
• The persona is too generic ("simulate a user") — produces generic output. Either narrow to a specific persona or use second-opinion for general critique.
• The user wants critique, not roleplay. Use second-opinion (analytic) or visual-review (visual) instead.
• The artifact is at draft-of-draft stage — wait for enough committed framing to have something concrete to react to.

Procedure

1. Identify the artifact and the persona

If the user said "test this email" without naming a persona, ask one short question: "Who should the reviewer pretend to be? An impatient executive, a confused customer, a skeptical engineer, a regulatory reviewer, a non-native-language reader?"

A specific persona produces a specific reaction. Generic personas produce generic reactions; that defeats the skill.

2. Frame the prompt strictly

Instruct the verifier to stay entirely in character and not break the fourth wall. Specifically:

• No meta-commentary ("As an executive, I would...")
• No phrase-by-phrase critique ("The third paragraph is unclear because...")
• No "review" framing
• Show internal monologue if confused / bored / annoyed — that is the signal
• React exactly as the persona would in real life, with the constraints the persona actually has (time pressure, attention, background, motivation)

3. Call the verifier

Invoke python3 "<plugin-root>/coordinator.py" with quality_profile='economical' and effort_class='minimal' (economical quality with minimal effort through an eligible provider — the skill default; raise both closed profiles for nuanced personas — short in-character responses do not need depth). Use frontier/maximum only for personas requiring nuanced reasoning, such as a litigator parsing a contract clause or a detail-focused engineer reading a specification.

Prompt template:

You are [specific, opinionated persona — title, background, current context, mood, time pressure]. You have [time/attention budget — "2 minutes between meetings," "skimming on a phone on the bus," "reading aloud to your team," etc.] to read this [artifact type].

React IN CHARACTER. [Word limit, e.g., "Under 150 words"]. No "review," no "As a {persona}..." preambles, no meta-commentary, no critique-mode. If you get confused, bored, distracted, irritated, or excited — show that as internal monologue. Respond exactly as you would in real life.

--- ARTIFACT ---
[paste the artifact verbatim]

Out-of-character artifact. If the response slips out of character (delivers a "review" instead of a reaction; says "As {persona}, I would..."), Surface the out-of-character simulation as incomplete. Do not issue a second provider request or ask the verifier to repair the artifact. A later caller-authorized request is a new attempt. If the caller authorizes one, include the stricter persona framing in that new request. The failure-to-inhabit is itself information.

4. Surface the in-character reaction, then synthesize

Present the verifier's in-character response to the user as-is, framed as the persona's reaction. Then step out of character yourself and offer a one-paragraph synthesis: what the simulation revealed, what concrete change to the artifact follows from it. Examples:

• "The executive completely ignored the third paragraph because the bottom line was buried. The ask should move to the top."
• "The confused customer got stuck on the word 'provision' — they thought it meant 'food allotment' rather than 'configure.' Reword to 'set up.'"
• "The skeptical engineer immediately challenged the latency number without seeing the source. Add the benchmark methodology inline."

The simulation is the raw data; the synthesis is what the user can act on.

Examples across domains

The pattern is constant: name a specific, opinionated, time-budgeted persona; cast strictly; capture the in-character reaction; synthesize what to change.

Anti-patterns

• Letting the verifier slip into "polite critique" mode. "I think this could be improved by..." defeats the skill. Surface the current artifact as incomplete; do not replay it automatically.
• Simulating a generic "user" instead of a specific opinionated persona. Generic personas produce generic reactions. Always name the persona's title, context, mood, time budget.
• Skipping the synthesis step. The raw in-character reaction is data; the user wants the actionable change. The skill is not done until the synthesis is delivered.
• Using frontier/maximum reflexively. Short in-character reactions favor economical/minimal; frontier/maximum is for nuanced-reasoning personas (litigator, detail engineer, or compliance officer parsing a regulation).
• Simulating personas the verifier may have content-policy issues inhabiting (hostile, prejudiced, criminal personas). If the persona's reaction is the actual question, frame the persona's role (e.g., "skeptical adversarial reviewer") rather than the persona's identity; or use second-opinion framed adversarially instead.
• Treating the simulation as ground truth for the actual audience. It is a prediction of audience reaction, not a focus-group result. For high-stakes audiences, run a real focus group or A/B test in addition.
• Running the same persona simulation 3 times without changing the artifact. No new information; the verifier saturates.
• Mixing simulate-user and visual-review on the same artifact in one call. Roleplay is fundamentally different from analytic critique; mixing produces neither cleanly. Two separate calls.