persona-review

name: persona-review description: "When a user wants a plan, design, document, PR description, code, or idea reviewed through expert personas, analyze the target and auto-select the matching read-only persona reviewer subagents (frontend, UX, accessibility, API, domain, scale, code, business, marketing, operations), dispatch them in parallel, and return each persona's native judgment with a short lead synthesis — no shared scorecard or fixed template. Falls back to Daniel when no axis clearly matches; honors an explicit --persona override." when_to_use: "Use before locking a non-trivial decision, when a target spans several expertise axes, or when the user names a persona lens. Auto-selection picks reviewers from the target's signals; override with --persona |all|auto|comma-list|. Trigger phrases: 'persona review', 'review through a persona', 'which experts should review this', 'frontend/UX/accessibility/API/domain/scale review', '/persona-review', '--depth interview'." argument-hint: " [--persona auto|daniel|evans|dean|martin|beck|fowler|abramov|evanyou|norman|rams|dunford|christensen|watson|fielding|all|comma-list|] [--depth quick|deep|interview]" disable-model-invocation: true allowed-tools: Read, Glob, Grep, Bash(git diff*), Bash(git log*), Bash(git status*), Agent, AskUserQuestion

/persona-review

Route the user's plan, document, PR description, design, code diff, or idea to read-only persona reviewer subagents. The goal is not to fill a scorecard. The goal is to expose weak assumptions, hidden costs, model problems, operational risks, and the next useful question through each selected persona's judgment style.

By default this skill auto-selects the personas that match the target. It analyzes the target, detects which expertise axes it touches, and dispatches the matching reviewers in parallel (up to three), then adds one lead synthesis. A single-axis target runs one persona; a multi-axis target runs several; a target with no clear axis falls back to daniel. An explicit --persona always overrides auto-selection.

The main skill is a lead/router, not the reviewer. Built-in persona reviews must be performed by the matching persona-<name>-reviewer subagent (the full map is in the Subagent Execution Contract), or by persona-custom-reviewer for a custom file.

If the user intent is unclear, do not critique prematurely. Ask one clarifying question before dispatch.

Core contract:

• Persona-native free-form review. Do not force a shared five-block template, shared table, or shared scorecard.
• Pressure Questions are an internal lens. Do not expand every question into rows or scores.
• Reviews are read-only. Neither the main skill nor subagents edit files, create files, stage, commit, or push.
• Built-in persona requests must go through the matching subagent.
• Do not guess. If evidence is missing, say what needs confirmation.
• End with one useful next question, next action, or handoff payload.

Standing Rules

• The main skill owns target collection, thin-input judgment, persona auto-selection, subagent dispatch, and aggregation.
• Preserve subagent output as much as possible. Add lead synthesis only for multi-persona runs.
• $ARGUMENTS overrides environment variables and defaults.
• When no --persona is given, auto-selection runs by default (see Phase 0). --persona auto requests it explicitly.
• ${PERSONA_REVIEW_DEFAULT} can pin a fixed default persona, replacing auto-selection.
• ${PERSONA_REVIEW_DEPTH} can set the default depth.
• Every mode is read-only. If implementation is needed, emit a handoff payload only.

Modes

--persona and --depth are independent. --persona evans --depth deep runs an Evans-style review with deeper evidence. Depth controls evidence and rounds; persona selection (manual or auto) controls who reviews.

Where this fits in the workflow

brainstorm -> persona-review lead -> persona reviewer subagent(s) -> lead aggregation -> decision-dashboard
                                                            \-> handoff payload -> improve / implement / fix

• Upstream: /brainstorm, a draft, a design, a PR description, a diff, or a short idea.
• This skill: choose persona, judge input density, ask if needed, dispatch subagent, relay or aggregate results.
• Downstream: use /decision-dashboard to lock decisions or /improve, /implement, /fix to execute changes.

Persona library

Auto-selection (Phase 0) chooses personas from the target by default. Use --persona to pin a specific lens, set, or all. A persona changes the judgment lens, not just the output format.

Recommended selection (auto-selection applies these automatically; this table is for manual --persona choices):

Custom personas

A custom persona should preferably use this structure:

# Persona: <name>

## Voice
## Hard Rules
## Decision Heuristics
## Priorities
## Forbidden
## Pressure Questions
## Blind Spots

Pressure Questions are prompts to think with, not a checklist to score. A custom persona may have its own output structure. The main skill reads the custom file and includes it in the persona-custom-reviewer prompt.

What's automated vs what needs your taste

Common Review Contract

This is the only shared contract. The rest belongs to the subagent's persona style.

1. Lead with the useful judgment. If the persona works better through a model, diagram, or number first, that is allowed.
2. Use the persona's native shape. Daniel focuses on operational mechanisms; Evans on model and language; Dean on bottlenecks and fault models; Martin on code smell, function structure, and tests.
3. No scorecard. Do not expand all Pressure Questions into rows, grades, or scores. Blend only useful questions into the review.
4. Evidence over vibe. Cite files, diffs, text, or explicit absence of evidence.
5. Use tables only when they clarify judgment. A table is optional, not a required format.
6. End with one next move. Use one next question, decision, strengthening request, or handoff payload.

Subagent Execution Contract

The main skill is the lead/router. Built-in persona requests must dispatch to the matching subagent.

Agent prompt payload

Each subagent prompt must include:

[ROLE]
Run a read-only persona review for /persona-review.

[USER INTENT]
<review intent, or "not specified">

[TARGET]
<file path, diff summary, or free text>

[DEPTH]
quick | deep | interview

[EVIDENCE]
<text, diff, file paths, or evidence already read by the main skill>

[CONSTRAINTS]
- Do not edit files.
- Do not create files.
- Do not run destructive commands.
- Do not output a scorecard or fixed five-block template.
- End with one next question/action/handoff.

Result handling

• single persona: relay the subagent result. Do not rewrite it except for obvious duplicate headings.
• multiple personas / all: preserve each persona result separately and add only one ## Lead synthesis section. Synthesis covers conflicts, common risks, and one next decision.
• interview: choose the one uncertainty that changes the conclusion, ask the user, then rerun the same subagent set after the answer.
• custom persona: if the custom file is missing or too thin, say so and ask whether to fall back to daniel.

Workflow

Phase 0: Persona selection or recommendation
Phase 1: Target collection and input-density judgment
Phase 2: Persona to subagent dispatch decision
Phase 3: Subagent prompt construction and Agent execution
Phase 4: Subagent result relay or aggregation
Phase 4.5: Interview convergence loop (--depth interview)
Phase 5: Handoff or close

Procedure

1. Parse target, --persona, and --depth from $ARGUMENTS.
2. If persona is missing or auto, run auto-selection (Phase 0). ${PERSONA_REVIEW_DEFAULT}, if set, pins a fixed persona instead.
3. If depth is missing, use ${PERSONA_REVIEW_DEPTH} or quick.
4. Read the target and judge input density.
5. Map the chosen persona(s) to subagent_type(s).
6. Build a read-only evidence payload and run Agent for each persona (in parallel when more than one).
7. Relay a single persona result; add lead synthesis when more than one persona ran.
8. In interview mode, ask one conclusion-changing question and stop.

Phase 0: Persona selection (auto by default)

When the user pins --persona <name>, all, a comma-list, or a custom path, honor it directly. Otherwise — no --persona, or --persona auto — auto-select: read the target, detect which expertise axes it touches, and dispatch the matching personas. Do not default to a single persona before looking at the target.

Signal map — match the target's path, content, and the user's intent against these axes:

Selection rule:

• 1 axis matches → run that one persona.
• 2–3 axes match → run those personas in parallel and add one ## Lead synthesis.
• More than 3 match → run only the 3 strongest (most central to the target) and note which axes were dropped so the user can re-run with a comma-list.
• 0 axes / ambiguous → fall back to daniel.

Never dispatch more than three personas in auto mode. The user forces the full roster with --persona all or a specific set with a comma-list. State the auto-selected personas (and the signals that triggered them) in one line before dispatching.

Phase 1: Target collection and input-density judgment

Ask before review when input is thin or ambiguous.

• Thin: one or two lines, unclear goal or scope.
• Ambiguous: document exists but goal or completion criteria are unclear.
• Sufficient: goal, scope, constraints, and completion criteria are available.

Phase 2: Persona to subagent dispatch decision

• Auto (no --persona or --persona auto): use the Phase 0 selection; map each selected persona to its subagent_type and prepare parallel runs when more than one.
• Name arg: choose the matching subagent_type.
• all or comma-list: prepare parallel read-only subagent runs.
• Path arg: read the custom persona file and choose persona-custom-reviewer.
• Missing custom file: ask whether to fall back to daniel.

Phase 3: Subagent prompt construction and Agent execution

The main skill reads only the needed evidence and passes it to Agent. Subagents may perform additional read-only exploration when the prompt and depth permit it.

Agent({
  subagent_type: 'persona-evans-reviewer',
  prompt: '<payload>'
})

Phase 4: Subagent result relay or aggregation

For multiple personas, preserve persona outputs and add one lead synthesis:

## daniel
<persona-daniel-reviewer output>

## dean
<persona-dean-reviewer output>

## Lead synthesis
- Common risk: ...
- Conflict: ...
- Next decision: ...

See references/example-review.md for examples. The example is guidance, not a fixed output template.

Phase 4.5: Interview convergence loop

quick and deep stop after one subagent run. interview can run up to three rounds.

1. Pick the biggest uncertainty that changes the conclusion.
2. Ask only that question.
3. Rerun the same subagent set after the answer.
4. Prefix each round with (interview R{n}/3).
5. Stop when the conclusion is clear, the user stops, or round 3 is reached.

Phase 5: Handoff or close

If the review should lead to implementation, emit a handoff payload only. Do not execute /improve, /implement, or /fix from this skill.

(handoff -> improve/implement/fix <target>)
- persona: <name>
- judgment: <one-line judgment>
- change intent: <what should change>
- evidence: <key evidence>
- risks to preserve: <risks that must not be lost>
- out of scope: <what not to do now>

Pitfalls

Forbidden

• Code edits, file creation, staging, commits, or pushes.
• Forcing a common five-block output.
• Outputting score tables, grade tables, or exhaustive checklist rows.
• Mechanically evaluating every Pressure Questions item.
• Writing built-in persona reviews directly in the main skill.
• Flattening subagent results into one generic voice.
• Praise, motivational tone, marketing tone, emojis, or apologies.
• Conclusions based on unverified assumptions.
• Automatic external URL fetch.
• More than three interview rounds.
• Executing the handoff payload inside this skill.

Validation

python3 plugins/elian-store/skills/persona-review/scripts/validate_skill.py
python3 plugins/elian-store/skills/persona-review/scripts/validate_skill.py --json

The validator checks frontmatter, read-only guardrails, Agent tool use, persona reviewer agents, persona references, example links, absence of fixed five-block or scorecard requirements, persona override support, and interview mode.

Pre-flight

• Target and persona are clear.
• Thin input was clarified before review.
• Built-in persona was dispatched to its matching subagent.
• Subagent prompt includes user intent, target, depth, evidence, and constraints.
• No scorecard or fixed shared template was produced.
• Confirmed evidence and missing evidence are separated.
• The response ends with one next question or next action.