Research Reprompter

Researcher v2.0 (USACF Research Generator)

Voice-to-research engineering for Claude Code. Transform rough questions into executable USACF swarm configurations.

Changelog

Purpose

Turn your rough research questions into complete, executable multi-agent research prompts using the USACF framework.

The Problem:

• Research questions are often vague and unstructured
• Manual setup of research swarms is tedious
• Missing adversarial review leads to blind spots
• No systematic algorithm selection

The Solution: Smart interview → USACF super-prompt with all phases, agents, and commands.

Process (4 steps)

Step 1: Receive raw input

Accept the user's rough research question (dictated, typed messily, or incomplete).

Trigger words: research, investigate, deep dive, analyze, research this

Step 2: Complexity detection

Auto-detect complexity to select algorithm:

• Simple (< 20 words, single topic) → CoT, 1-3 agents
• Medium (branching, comparison) → ToT, 4-8 agents
• Complex (comprehensive, multi-domain) → GoT, 9-15 agents

Step 3: Smart interview (gather user input)

Gather:

1. Research Title - Name for this research
2. Subject - What we're researching
3. Subject Type - Product / Software / Business / Process / Organization
4. Research Type - Competitive / Gap / Technical / Due Diligence / Market
5. Objectives - What to find out (1-20, one per line)
6. Constraints - Focus areas, limitations (optional)
7. Depth - CoT / ToT / GoT
8. Output - Brief / Full Report / Action Plan / Raw

Step 4: Generate USACF super-prompt + score

Generate complete executable configuration with:

• Initialization commands
• All phase agents (Discovery, Analysis, Adversarial, Synthesis)
• Memory operations
• Final report generator
• Quality score comparison

CRITICAL: MUST GENERATE COMPLETE SUPER-PROMPT

After interview completes, you MUST immediately:

1. Select algorithm (CoT/ToT/GoT) based on complexity
2. Generate full USACF super-prompt with ALL phases
3. Include claude-flow commands for every operation
4. Add adversarial review agents (red-team, fact-checker)
5. Show quality score (before/after comparison)
6. Offer to execute or copy

WRONG: Generate simple prompt without agents
RIGHT: Generate full USACF config with all phases, agents, memory ops

Algorithm selection matrix

Complexity Indicators:

• Simple: Single topic, factual question, < 20 words
• Medium: "compare", "vs", "evaluate", "options"
• Complex: "comprehensive", "gaps and opportunities", multiple domains

USACF phases (all required for complex research)

Phase 0: Initialization

npx claude-flow@alpha init --force
npx claude-flow@alpha swarm init --topology {topology} --max-agents {N}
npx claude-flow@alpha memory store "session/config" '{...}' --namespace search

Phase 0.5: Meta-analysis

• Step-back prompting (principles, criteria)
• Self-ask decomposition (15-20 questions)
• Research planning (ReWOO)

Phase 1: Discovery (Parallel)

• component-identifier
• hierarchy-analyzer
• interface-mapper
• flow-tracer

Phase 2: Analysis (Parallel)

• 6 gap hunters (quality, performance, security, structural, capability, UX)
• 4 risk analysts (FMEA, edge cases, vulnerabilities, reliability)

Phase 2.5: Adversarial review (critical)

• red-team-reviewer: Challenge ALL findings
• fact-checker: RAG verification with web_search
• coordinator: Integrate feedback, update confidence

Phase 3: Synthesis (Parallel)

• quick-win-generator (0-3 months)
• strategic-generator (3-12 months)
• transformational-generator (12-36 months)
• pareto-optimizer (multi-objective portfolios)

Phase 4: Final report

• Ultra-brief (3 sentences)
• Executive summary
• Top 10 findings with confidence
• Recommended actions by horizon
• Limitations & uncertainties

Memory namespace convention

All agents store to namespaced memory:

# Session
session/config          # Research configuration

# Meta
meta/principles         # Core principles
meta/questions          # Decomposed questions
meta/research-plan      # Planned tasks

# Discovery
discovery/components    # Identified components
discovery/hierarchy     # Structural map
discovery/interfaces    # APIs/contracts
discovery/flows         # Data/control flows

# Gaps
gaps/quality            # Quality gaps
gaps/performance        # Performance gaps
gaps/security           # Security gaps

# Risks
risks/fmea              # Failure mode analysis
risks/edge-cases        # Edge cases
risks/vulnerabilities   # Security vulnerabilities

# Adversarial
adversarial/critiques   # Red team challenges
adversarial/fact-check  # Verified claims

# Opportunities
opportunities/quick-wins        # 0-3 month wins
opportunities/strategic         # 3-12 month plays
opportunities/transformational  # 12-36 month bets
opportunities/pareto-recommendation  # Optimal portfolio

# Output
output/final-report     # Comprehensive report

Quality scoring

Always show before/after metrics:

Example

Before (rough input):

"look into what solana is doing with AI and how we compare"

After (USACF super-prompt):

# USACF Research: Solana AI Competitive Analysis

## Configuration
- Algorithm: ToT (medium complexity - comparison)
- Topology: Hierarchical
- Agents: 8
- Output: Executive Brief

## Phase 0: Initialization
[claude-flow init commands]

## Phase 1: Discovery
[4 parallel agents with memory stores]

## Phase 2: Analysis
[Gap hunters + risk analysts]

## Phase 2.5: Adversarial
[Red team + fact checker]

## Phase 3: Synthesis
[Opportunity generators + pareto optimizer]

## Phase 4: Report
[Final report generator]

Quality: 1.2/10 → 9.3/10 (+675%)

Tips for best results

• Be specific about competitors - Name them in the input
• Mention constraints early - "focus on Q1", "executive-level"
• State objectives - Even rough ones help
• Say "expand" - For full interview on simple queries
• Say "quick" - To skip interview for simple research

Comparison: Reprompter vs Researcher