Perspectives

Perspectives

Evaluate anything through the eyes of multiple relevant personas. Dynamically casts perspectives that fit the situation. Domain-agnostic.

Announce at start: "Using perspectives to evaluate [subject] from multiple angles."

The Iron Law

PERSPECTIVES ARE GENERATED INDEPENDENTLY BEFORE SYNTHESIS.
NO COPY-PASTE OPINIONS. EACH PERSONA THINKS FOR THEMSELVES.
IF ALL PERSPECTIVES AGREE, RE-CAST WITH MORE CONTRAST.

Three Tiers

T3 is not "deeper opinions" — it's a fundamentally different operation: real agent dispatch with file I/O, web search, and tool use. Treat it as such.

Escalation Rule

If the current tier finds nothing alarming → Done. Stop. No upsell. Never routinely ask "want me to go deeper?"

Process

Step 1: Intake

Phase 1A: Identify the Subject

• Domain — code, game design, business, UI, product, personal decision, etc.
• Subject — what specific thing are we evaluating?
• Tier — user-requested, else default T1.

If the user says just "perspectives" with no subject, infer it from conversation context.

Phase 1B: Calibration

Match depth to subject:

• Casual subjects ("should I get a cat?", "is this color OK?") — skip Phase 1B. Trust the user. The intake gate exists to prevent shallow analysis on serious questions, not to interrogate small ones.
• Most subjects — ONE calibration question is enough:

  "Quick check before I cast — are you [questioning a choice / seeking validation / blind-spot checking / just curious]? Anything I should know about constraints (budget, audience, hard requirements)?"

• Technical subjects (hardware, code, performance) — confirm hard facts before assuming defaults. If you catch yourself writing "assuming X..." — STOP. Ask.
• Personal/emotional subjects — the backstory IS the subject. Ask about intent and feelings.

Step 2: Cast

Select 4-6 perspectives. See persona-engine.md for the dimension catalog and selection rules.

For each persona present:

• Name — short, vivid label (not "User A")
• Who — 1-2 sentences: background, mindset, what they care about
• Why here — why their opinion matters for THIS subject

At least one adversarial perspective. At least one the user probably hasn't considered.

Wait for user to confirm, modify, or swap personas.

Step 3: Analyze (T1)

For each persona, independently generate:

1. Who I am (1 line)
2. Honest reaction — speak AS this person. Authentic > diplomatic. Don't soften.
3. What I'd actually do — behavior, not just opinion. Bounce? Complain? Workaround? Love it?
4. Verdict — 👍 / 👎 / 🤷 with a one-sentence reason

3-5 sentences per persona. This is a quick council, not an essay.

Step 4: Synthesize

• Consensus — what most agree on (if anything)
• Conflicts — where perspectives clash. These ARE the interesting design tensions.
• Blind spots — what the user likely didn't consider

Step 5: Recommendation

Direct stance based on the full picture. Take a position.

T2: Deep Analysis

Triggered by explicit request OR earned escalation from T1.

See frameworks.md for the framework catalog.

Additional steps:

1. Framework Selection — pick 1-2 that fit:

   • Decisions → Six Thinking Hats / Devil's Advocate
   • Plans/launches → Pre-Mortem
   • UX/flows → Empathy Map + Cognitive Walkthrough
   • Security/robustness → Red/Blue Team
   • Multi-stakeholder → Stakeholder Analysis
   • Creative features → Disney Strategy

2. Structured findings — typed output per the framework, not free opinions.

3. Action items — see action-translator.md. Every finding becomes: FINDING → IMPACT → SEVERITY → SPECIFIC ACTION → EFFORT

4. Prioritized roadmap:

   • 🔴 Fix now (Critical)
   • 🟡 Fix next (Major, workarounds exist)
   • 🟢 Schedule (Minor)
   • 💡 Consider (Enhancement)

T3: Execution Tier

Triggered by explicit request OR earned escalation from T2.

See agents/ for each agent's prompt and protocol.

T3 Cost Gate

Heads up — T3 is heavy. About to dispatch [N] agents in parallel. Each independently explores the codebase and runs searches. This means:

• ~[N × 30-70] tool calls total
• Large context usage — each agent builds its own deep understanding
• ~2-5 minutes wall clock
• Long synthesis report

The most expensive operation in this skill. Worth it for real projects, overkill for quick questions.

Ready to dispatch?

"go", "yes", "do it" = confirmed. Anything else → stop and clarify.

Dispatch

1. Select agents based on T1/T2 findings.
2. Present selection for user confirmation.
3. Show the cost warning above.
4. On confirmation, dispatch in parallel via the Agent tool.
5. Each agent gets constructed context inline — never tell them "read the conversation."
6. Each agent reports back: Findings, Severity, Recommended Actions.
7. Synthesize across all agents — patterns, conflicts, compound issues.
8. Optionally execute fixes if user approves.

Rules

• Dynamic casting — every subject gets fresh personas. No reuse without rethinking.
• Authentic voice — frustrated people sound frustrated. A 70-year-old sounds different from a 15-year-old. Don't homogenize.
• Concise by default — T1 is 3-5 sentences per persona.
• User controls depth — they can stop, escalate, or add personas at any tier.
• No domain lock — works for code, UI, game design, business, personal choices, writing, architecture.

Anti-Patterns

Intake

Analysis

Integration

• Chains well with: superpowers:brainstorming (during design phase), superpowers:writing-plans (review a plan before execution), superpowers:test-driven-development (T3 test-writer agent).
• After perspectives: user decides next step. Don't auto-chain.