focusgroup

Focus Group Skill

Multi-persona review system that dynamically generates diverse reviewer personalities to provide comprehensive, evidence-based feedback on any target. Supports deep review protocols, iterative methodologies, automated quality control, and rich output formats.

Usage

/focusgroup https://example.com
/focusgroup ./src --deep
/focusgroup "a dating app for dogs" --delphi --full-report
/focusgroup https://mysite.com --panel startup --count 5
/focusgroup https://site.com --compare https://competitor1.com https://competitor2.com
/focusgroup https://mysite.com --walkthrough "sign up and make first purchase"
/focusgroup ./src --debate --strict --action-plan
/fg my-product --format json --html

Arguments

Argument	Description	Default
<target>	URL, file path, or idea description	(required)
--count N	Number of personas (3-15)	10
--panel <type>	Preset panel focus	(auto)
--sequential	Run reviews one at a time	parallel
--parallel	Run reviews in batches of 5	(default)
--format md|json	Output format	md
--personas <file>	Custom persona definitions file	(auto-generated)
--out <dir>	Output directory	.focusgroup/{timestamp}
--deep	Use full multi-phase review protocol per reviewer (8 phases for websites, 9 for codebases, 6 for ideas)	(standard)
--shallow	Skip deep site crawl, use homepage-only (NOT recommended for websites)	(deep)
--delphi	3-round iterative Delphi method	(none)
--debate	Red Team vs Blue Team structured debate	(none)
--walkthrough "goal"	Cognitive walkthrough for a user goal	(none)
--compare url1 [url2] [url3]	Comparative review against competitors	(none)
--strict	Raise quality threshold to 75/100	60/100
--no-quality-check	Skip automated quality validation	(enabled)
--action-plan	Generate prioritized action plan/backlog	(off)
--backlog-format	Action plan format: markdown|csv|github|linear	markdown
--swot	Generate SWOT analysis matrix	(off)
--risk-matrix	Generate effort/impact matrix	(off)
--html	Generate self-contained HTML visual report	(off)
--full-report	Shorthand for --action-plan --swot --risk-matrix --html	(off)

Preset Panels

• startup - Product-market fit, growth, MVP, competitive analysis
• enterprise - Scalability, compliance, security, integration, vendor risk
• accessibility - WCAG, assistive tech, cognitive load, inclusive design
• security - Attack surface, auth, data protection, threat modeling
• ux - User journeys, visual design, interaction patterns, usability
• performance - Load times, bundle size, caching, scalability
• content - Messaging, tone, SEO, readability, audience targeting

Input Types

• Website: URLs (http/https) — deep site crawl (all pages + linked properties) + browser tools + 8-phase protocol
• Codebase: File paths — Read, Grep, Glob, LSP, AST + optional deep 9-phase protocol
• Product: URL + local code — combines both approaches
• Idea: Plain text — web search + optional 6-phase research protocol

Website Research (CRITICAL)

For website targets, the orchestrator MUST run a deep research phase BEFORE generating personas or launching reviewers. A homepage-only scrape produces wildly incomplete reviews.

Research Phase (mandatory for all website targets)

1. Launch the website-researcher agent (agents/probes/website-researcher.md) to crawl the entire site:

   • Visit ALL internal pages (up to 25), not just the homepage
   • Follow outbound links to properties OWNED by the site owner (products, apps, research)
   • Read FULL text of all blog posts and articles
   • Run the technical JS probe
   • Output a comprehensive dossier to .focusgroup/site-dossier.md

2. Pass the dossier to ALL downstream agents:

   • The focus-director uses it to generate relevant personas
   • Each focus-reviewer receives it as their primary evidence source
   • The synthesis agent references it to verify reviewer claims

Why This Matters

Without deep research:

• Reviewers critique absence of things that actually exist on subpages
• Products linked from the site are invisible to reviewers
• Published research/articles are missed entirely
• Reviews become shallow impressions of a homepage, not informed assessments

The dossier is the foundation of review quality. Never skip it.

Methodology Modes

Delphi (--delphi)

3-round iterative review: blind review → informed revision with anonymized group feedback → final positions. Produces an evolution trace showing how opinions shifted. Best for: building consensus on complex topics.

Debate (--debate)

Red Team vs Blue Team with 3 rounds (Opening → Rebuttal → Closing) + Judge evaluation. Produces vulnerability report with defense ratings. Best for: adversarial quality testing, security reviews.

Walkthrough (--walkthrough "goal")

Each persona narrates their attempt to complete a specific user goal step by step, scoring friction per step. Produces a friction heatmap and drop-off funnel. Best for: UX evaluation, onboarding flows.

Comparative (--compare url1 url2)

Each persona reviews all targets on identical criteria. Produces competitive matrix and gap analysis. Best for: competitive analysis, vendor selection.

Methodology flags are mutually exclusive — use only one per session.

Quality Control

Reviews are automatically validated against quality standards:

• Specificity check: Catches vague phrases like "could be improved"
• Evidence check: Every claim must cite URLs, file paths, metrics, or quotes
• Rating consistency: Overall score must match individual area ratings
• Completeness: All template sections must meet minimum word counts
• Hallucination guard: Cross-references claims against probe data

Use --strict for higher quality threshold (75 vs 60). Use --no-quality-check to skip.

Output Formats

Action Plan (--action-plan)

Prioritized backlog with effort estimates (XS-XL), impact ratings, and measurable success criteria. Supports --backlog-format markdown|csv|github|linear.

SWOT Matrix (--swot)

Strengths/Weaknesses/Opportunities/Threats analysis with cross-reference strategies.

Risk/Impact Matrix (--risk-matrix)

2x2 effort-impact grid categorizing actions as Quick Wins, Strategic Bets, Fill-ins, or Avoid.

HTML Report (--html)

Self-contained visual report with radar charts, rating bars, SWOT grid, and print-friendly styles.

Full Report (--full-report)

Generates all four output types in one run.

Output

All output goes to .focusgroup/{timestamp}/:

• ##-{slug}.md — Individual persona reviews
• synthesis.md — Aggregate analysis with consensus, disagreements, ratings
• meta.json — Session metadata
• probe-data.json — Website diagnostic data (website/product targets)
• action-plan.md — Prioritized action plan (if requested)
• swot.md — SWOT analysis (if requested)
• risk-impact-matrix.md — Effort/impact matrix (if requested)
• report.html — Visual HTML report (if requested)
• Methodology-specific outputs (round directories, evolution traces, etc.)

Advanced Examples

# Deep website review with full report
/focusgroup https://mysite.com --deep --full-report

# Delphi consensus with strict quality, 5 reviewers
/focusgroup ./src --delphi --strict --count 5

# Debate mode with action plan in GitHub issues format
/focusgroup https://app.com --debate --action-plan --backlog-format github

# Competitive analysis with HTML report
/focusgroup https://mysite.com --compare https://competitor1.com https://competitor2.com --html

# Walkthrough with custom personas
/focusgroup https://mysite.com --walkthrough "find and purchase a product" --personas examples/personas-accessibility.yaml

Custom Personas

Create a YAML file with persona definitions:

personas:
  - name: "Jane Smith"
    slug: "jane-smith"
    title: "Senior Product Designer, 12 years experience"
    archetype: "The Pragmatic Designer"
    expertise: ["UX design", "design systems", "user research"]
    perspective: "Form follows function — beauty without usability is decoration"
    review_priorities:
      - "Visual hierarchy and information architecture"
      - "Interaction patterns and affordances"
      - "Design system consistency"
    tone: "Direct but constructive, always suggests alternatives"
    rating_bias: "Scores high for clear user intent, low for style over substance"

See examples/ for starter persona files: startup, security, investor, journalist, enterprise-buyer, accessibility.

Magic Keywords

• "focusgroup", "focus group", "persona review", "multi-perspective", "panel review"