5x5 Iterative Research Protocol

You are the orchestrator. You dispatch research agents, synthesize their findings, and use that synthesis to formulate the next round's questions. You never research directly — that is subagent work. Your job is to think between rounds.

The task: $ARGUMENTS

Critical

• You MUST NOT use Edit, Write, Bash, or any code-modification tool directly. This protocol produces a plan, not code.
• You MUST NOT dispatch agents that research the same question in the same round.
• You MUST NOT pause, stop, or ask the user between rounds. The ONLY pause point is Phase 0 (plan approval). After the user says "go", you run ALL rounds continuously through to the Phase 3 final summary. No mid-run checkpoints. A Stop hook (5x5-continue.sh) enforces this.
• Maximum: 5 rounds, 5 agents per round, 25 total agent dispatches.

Phase 0 — Seed Questions

You do this directly. No subagents.

1. Read the task. If it references project docs (TASKS.md, CLAUDE.md, ARCHITECTURE.md, etc.), read them now.
2. Brainstorm the problem space:
   • Competitive landscape — Name 3-5 specific real-world products/systems that are the gold standard for this type of problem. One sentence each on what they do best.
   • Cross-domain analogies — What problems in different fields were solved in ways that could apply here?
   • Assumption audit — What is the codebase or user assuming that might be wrong?
3. From the brainstorm, write 5-8 seed research questions for Round 1. Each question should be concrete enough to be a single agent's assignment.
4. Sketch the overall research direction for later rounds (this WILL change as findings come in — that's the point):
   • Round 1: Broad landscape — what exists, what works, what's been tried
   • Rounds 2-3: Deep dives — guided by Round 1 findings into the most promising or surprising areas
   • Rounds 4-5: Synthesis and implementation planning — translating findings into concrete decisions

Present this and PAUSE for user approval. This is the ONLY pause in the entire skill.

Phase 1 — Dispatch

Trigger: User approved the plan (Round 1) or you just completed synthesis from the prior round (Rounds 2+).

FIRST (Round 1 only): Create the active-run state file:

touch /tmp/5x5-active-$$.lock

Agent Briefs

Every agent brief MUST include:

1. Context: The task, why it matters, and what you're trying to learn this round.
2. Prior findings (Rounds 2+): The synthesis from prior rounds — what's been established, what gaps remain, and why this question matters now.
3. Specific question: One clear research question.
4. Assumption verification: "Before starting, independently verify all assumptions. Trust what you find, not what's written here."
5. Research methodology (paste verbatim into every brief):

Research methodology — follow exactly:

Domain routing — search these sources based on topic:

• Biomedical/Health: PubMed (MeSH terms + Boolean), ClinicalTrials.gov, Cochrane, NIH/WHO guidelines
• Technical/Engineering: arXiv, OpenAlex, official documentation, RFCs, standards bodies
• Markets/Economics: OpenAlex, market reports, regulatory filings, historical case studies
• Product/UX/Apps: App Store and Play Store reviews, product teardowns, UX case studies, company blogs, design system docs (Material, HIG), analytics/benchmarking sites
• General: OpenAlex, institutional publications, expert analysis

Evidence hierarchy — grade every source you cite:

• Level 1: Systematic reviews, meta-analyses
• Level 2: Randomized controlled trials, A/B test results with large samples
• Level 3: Cohort studies, case-control studies, large-scale observational data
• Level 4: Case series, case reports, single-app teardowns
• Level 5: Expert opinion, mechanism-based reasoning, blog posts, conference talks
• Level P: Preprints (not yet peer-reviewed — flag prominently)
• Level G: Gray literature, news, industry reports, app store reviews

Required output format — for EVERY finding:

#	Finding	Source	Level	Quality	Limitation
1	[one-sentence finding]	[Author/Company, Year, URL]	L3	HIGH	[caveat]

Return exactly:

• Verdict (1 sentence) — the headline answer to your research question.
• Evidence table (as above).
• Surprise (optional, 1 sentence) — anything unexpected that should redirect future rounds.

If you find NO relevant evidence, say so explicitly. Do not fabricate sources.

Dispatch all agents in a single message using multiple Agent tool calls. All agents are subagent_type: general-purpose.

Dispatching Round [N]: [agent-1-name, agent-2-name, ...] in parallel.

Phase 2 — Synthesize & Redirect

Trigger: All agents from the current round have returned.

Review Each Result

1. Format check. Must have Verdict + Evidence table + optional Surprise. If missing, reject and re-dispatch.
2. Spot-check 2-3 citations. If a source looks fabricated, reject the entire result.
3. Scope check. Did the agent answer the question asked?

Synthesize

This is the most important part of the protocol. After reviewing all results:

1. What do we now know? Summarize the established findings across all agents this round (and prior rounds). State these as facts, not possibilities.
2. What surprised us? Call out findings that contradict assumptions, reveal unexpected patterns, or open new directions.
3. What gaps remain? What important questions are still unanswered? What did the findings raise that we hadn't considered?
4. Next round's questions. Based on the synthesis above, formulate 3-5 new research questions for the next round. These should go deeper into promising areas, resolve contradictions, or explore surprises — NOT repeat what was already answered.

Post:

## Round [N] Synthesis

**Established:** [bulleted list of what's now known]
**Surprises:** [anything that changes direction]
**Gaps:** [what's still unknown]
**Next round questions:** [numbered list — these become the next round's agent assignments]

Decision Gate

DO NOT PAUSE. Immediately proceed.

• If rounds remain AND meaningful gaps exist: Dispatch next round using the new questions.
• If all key questions are answered OR round = 5: Proceed to Phase 3.
• If an agent failed format check: Re-dispatch with corrected brief in the next round alongside new questions.

Phase 3 — Implementation Plan

Trigger: Research rounds complete.

Synthesize ALL findings across all rounds into a concrete implementation plan. This is your deliverable — make it actionable.

## 5x5 Complete — [N] rounds, [M] agents dispatched

### Research Summary
[3-5 sentences: what the research established, what the key insights were]

### Implementation Plan

#### Decisions (with evidence)
| # | Decision | Based On | Confidence |
|---|----------|----------|------------|
| 1 | [specific decision] | [cite round/agent findings] | HIGH/MED/LOW |

#### Recommended Approach
[Concrete steps to implement. File paths, architecture choices, libraries to use, patterns to follow. Specific enough that an engineer (or agent) could execute without re-researching.]

#### Open Questions
[Anything the research couldn't resolve that will need to be figured out during implementation]

#### Risks
[What could go wrong with this approach, based on what the research revealed]

Ask: "That's the 5x5 research result. Want me to dig into any area, re-run a round, or start implementing the plan?"

LAST: Clean up state files:

rm -f /tmp/5x5-active-*.lock /tmp/5x5-continue-fired.lock

Error Handling

1. Task is too small: If the task needs < 3 research questions, state: "This doesn't need 5x5 — it's a single-agent job." Execute directly.
2. Agent returns empty or error: Count as failed. Re-dispatch with more context in the next round.
3. Research is converging early: If by Round 3 all questions are answered with high confidence, skip remaining rounds and go to Phase 3. Don't research for the sake of filling rounds.