perspective-research

Perspective Research

Explore open questions through diverse perspectives. Each perspective agent gets its own context window, independently explores the codebase and question, then a cross-pollination round lets perspectives build on each other to produce hybrid approaches no single lens would generate alone.

Core principle: Generative exploration benefits from dialogue between perspectives. Pre-mortem research shows prospective hindsight improves outcome prediction by 30%, but only when participants build on each other's scenarios.

When to Use

• "Should we use X or Y approach?"
• "What's the best way to handle Z?"
• "What are the risks of this architecture?"
• "What tech stack should we use for this?"
• "Are we following best practices for X?"
• Any open question where diverse perspectives would help before deciding

Not for: Evaluating an existing artifact (use perspective-review instead)

Process

Step 1: Understand the question

Read the user's question or topic. Gather relevant context from the codebase (existing implementations, constraints, prior decisions) using Read, Glob, Grep.

If the question is vague, ask one clarifying question.

Step 2: Recommend perspectives

Read the perspective catalogue at ../shared-perspectives/catalogue.md.

The catalogue has two families of perspectives:

• Role-based (Adversary, Operator, Maintainer, etc.) — who is looking
• Discipline-based (Design Principles, Testing Strategy, etc.) — what lens

Mix them based on the question. Tech stack evaluations benefit from role-based perspectives (Business/Strategy, Operator, Integrator). Best practices questions lean toward discipline-based (Design Principles, Conventions & Idioms). Many questions benefit from a mix. See the catalogue's selection guidance.

If the user requests a lens not in the catalogue, create a custom perspective with a specific analytical procedure (see catalogue's "Custom Perspectives" section).

Recommend 3-4 perspectives with reasoning. Present to the user:

"I'd recommend these perspectives for exploring [question]:

• [Perspective] (role/discipline/custom): [why it's relevant here]
• ...

Want to adjust — add, remove, or swap any?"

Wait for confirmation.

Step 3: Choose execution approach

Present the choice:

"Two execution approaches:

1. Subagent-Driven — Parallel subagents per perspective, file-based cross-pollination with agent resume. Fast, efficient, proven.

2. Team-Driven — Persistent teammate agents with direct messaging. Richer cross-pollination through real-time dialogue. Costs 2-4x more. Requires Opus 4.6+ and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

Which approach?"

Default to subagent-driven if the user doesn't have a preference.

Step 4: Set up workspace

Create a temporary directory for round outputs:

<workspace>/
  round-1/
  round-2/

---

Path A: Subagent-Driven

Step 5a: Round 1 — Independent Exploration

Read ../shared-perspectives/perspective-subagent-prompt.md for the Round 1 research template.

For each confirmed perspective, look up its full analytical procedure from the catalogue. Fill the Round 1 research template with:

• {PERSPECTIVE_NAME} — the perspective name
• {PERSPECTIVE_PROCEDURE} — the full analytical procedure from the catalogue
• {QUESTION_CONTENT} — the user's question
• {CONTEXT} — any relevant context gathered in Step 1
• {OUTPUT_PATH} — where to save position (e.g., <workspace>/round-1/adversary.md)

Spawn one subagent per perspective in parallel using the Agent tool. Each subagent independently explores the codebase and saves its position to the output file.

Capture each subagent's agent ID from the return value — you will resume these agents in Round 2.

Wait for all Round 1 subagents to complete.

Step 6a: Round 2 — Cross-Pollination (Primary Mechanism)

Resume each Round 1 subagent using the Agent tool's resume parameter with the agent ID captured in Step 5a. The resumed agent retains its full Round 1 context — every file it read, every constraint it found — so it can propose deeply informed hybrid approaches.

Read the Round 2 template from ../shared-perspectives/perspective-subagent-prompt.md.

Send each resumed agent:

• Paths to all OTHER perspectives' Round 1 output files, labeled by name
• {OUTPUT_PATH} — where to save cross-pollination response
• Instruction that its Round 1 position is LOCKED — react only, do not revise

Resume all perspectives in parallel.

This round is where the primary value emerges — perspectives build on each other to generate ideas no single perspective would produce alone.

Wait for all Round 2 subagents to complete.

Step 7a: Synthesis

Read ../shared-perspectives/synthesis-agent-prompt.md for the research synthesis template.

Fill the template with:

• {ROUND_1_PATHS} — paths to all Round 1 output files, labeled by perspective
• {ROUND_2_PATHS} — paths to all Round 2 output files, labeled by perspective
• {QUESTION_CONTENT} — the original question

Spawn a synthesis subagent. It reads all output files and produces the recommendation with confidence level and ADR template.

Save the final report.

---

Path B: Team-Driven

Step 5b: Set up team

Create the team:

TeamCreate(team_name: "perspective-research", description: "Multi-perspective research on [question]")

Read ../shared-perspectives/perspective-teammate-prompt.md for the research teammate template. For each confirmed perspective, fill the template with:

• {TEAM_NAME} — the team name
• {PERSPECTIVE_NAME} — the perspective name
• {PERSPECTIVE_SLUG} — kebab-case name (e.g., adversary, business-strategy)
• {PERSPECTIVE_PROCEDURE} — the full analytical procedure from the catalogue
• {QUESTION_CONTENT} — the user's question
• {CONTEXT} — relevant context
• {OUTPUT_PATH_ROUND_1} — e.g., <workspace>/round-1/adversary.md
• {OUTPUT_PATH_ROUND_2} — e.g., <workspace>/round-2/adversary.md

Spawn one teammate per perspective using the Agent tool with team_name.

Read ../shared-perspectives/synthesis-teammate-prompt.md for the research synthesis teammate template. Spawn the synthesis teammate with team_name.

Step 6b: Round 1 + Cross-Pollination

Round 1: Each perspective teammate independently explores the question and messages you when done. Wait for all perspectives to report completion.

Round 2: Once all Round 1 outputs are saved, send each perspective teammate the file paths of all OTHER perspectives' Round 1 outputs via SendMessage:

"Cross-pollination round. Read these other perspectives' Round 1 positions:

• Adversary: <workspace>/round-1/adversary.md
• Operator: <workspace>/round-1/operator.md
• ... Your Round 1 position is LOCKED. React only."

Wait for all perspectives to report Round 2 completion.

Step 7b: Synthesis + Shutdown

Send the synthesis teammate all file paths via SendMessage:

"All rounds complete. Produce the synthesis report. Round 1 files: [list with perspective labels] Round 2 files: [list with perspective labels] Original question: [question] Save report to: {OUTPUT_PATH}"

Wait for synthesis to complete. Then shut down:

1. Send shutdown_request to each teammate via SendMessage
2. Bash("sleep 30", run_in_background=true) — the harness blocks standalone sleeps; background and wait for the completion notification before step 3
3. TeamDelete

---

Step 8: Present results

Present the synthesis report to the user. Highlight:

• The recommended approach with confidence level
• Key tradeoffs they must accept
• The pre-formatted ADR template

Offer: "Want me to save the decision record to docs/decisions/?"

If the user wants to proceed:

• "Want to design a solution based on this?" -> brainstorming
• "Ready to plan implementation?" -> writing-plans

Integration

Invoked before:

• h-superpowers:brainstorming — research informs the design
• h-superpowers:writing-plans — if decision is clear, plan directly

Hands off to:

• h-superpowers:brainstorming — to design based on findings
• h-superpowers:writing-plans — if answer is clear enough to plan

Does NOT replace:

• h-superpowers:brainstorming — research explores the question space, brainstorming designs the solution