Research Report → Focus Group Questions

Extract key findings and controversial claims from industry research reports, auto-generate Focus Group questions that test whether target personas agree with the conclusions — turning static research into validated, audience-specific intelligence

Scenario

Take industry research reports and use Claude to extract key findings, controversial claims, and actionable insights. Then auto-generate Focus Group questions that test whether your target personas would agree with or act on the report’s conclusions — turning static research into validated, audience-specific intelligence.

Flow: Research report → Anthropic POST /v1/messages (extract findings) → Mavera POST /personas → Mavera POST /focus-groups (auto-generated questions) → Validated research

Code

Python

import os, json, time, anthropic, requests

MV = os.environ["MAVERA_API_KEY"]
MV_BASE = "https://app.mavera.io/api/v1"
MV_H = {"Authorization": f"Bearer {MV}", "Content-Type": "application/json"}
client = anthropic.Anthropic()

with open("gartner_martech_2025.txt") as f:
    report_text = f.read()
print(f"Report loaded: {len(report_text):,} chars (~{len(report_text)//4:,} tokens)")

# 1. Claude extracts findings and generates questions
extraction = client.messages.create(
    model="claude-opus-4-6-20250725",
    max_tokens=4096,
    input=[{
        "role": "user",
        "content": "Research analyst. Read this report and:\n\n"
            "A) Extract TOP 8 FINDINGS ranked by how controversial or actionable they are.\n"
            "Per finding: claim (1 sentence), evidence cited, confidence, counterargument.\n\n"
            "B) For each finding, generate 2 FOCUS GROUP QUESTIONS that test practitioner agreement.\n"
            "Questions should be specific, neutral, and grounded in experience.\n\n"
            "Return JSON: [{finding, evidence, confidence, counterargument, questions: [q1, q2]}]\n\n"
            f"REPORT:\n{report_text}"
    }],
)
raw = extraction.content[0].text
print(f"Extraction — {extraction.usage.input_tokens:,} input tokens")

try:
    findings = json.loads(raw[raw.index("["):raw.rindex("]") + 1])
except (ValueError, json.JSONDecodeError):
    findings = [{"finding": raw[:500], "questions": ["Does this match your experience?"]}]
print(f"{len(findings)} findings, {sum(len(f.get('questions',[])) for f in findings)} questions")

# 2. Create target personas
TARGET_PERSONAS = [
    {"name": "VP Marketing, Enterprise SaaS", "desc": "15 years B2B marketing. $5M budget. Team of 20. Skeptical of vendor claims. Data-driven."},
    {"name": "Marketing Ops Manager", "desc": "8 years experience. Manages tech stack day-to-day. Cares about integration and automation."},
    {"name": "CMO, Growth-Stage Startup", "desc": "Series B. $1.5M budget. 6-person team. Needs efficiency over features."},
]
persona_ids = []
for tp in TARGET_PERSONAS:
    p = requests.post(f"{MV_BASE}/personas", headers=MV_H, json={
        "name": tp["name"], "description": tp["desc"],
    }).json()
    persona_ids.append(p["id"])
    time.sleep(0.3)

# 3. Collect questions and run Focus Group
all_questions = [q for f in findings[:6] for q in f.get("questions", [])[:2]]
fg = requests.post(f"{MV_BASE}/focus-groups", headers=MV_H, json={
    "name": "Research Validation: Gartner MarTech 2025",
    "persona_ids": persona_ids,
    "questions": all_questions[:10],
    "responses_per_persona": 2,
}).json()

for _ in range(30):
    time.sleep(5)
    data = requests.get(f"{MV_BASE}/focus-groups/{fg['id']}", headers=MV_H).json()
    if data.get("status") == "completed":
        break

# 4. Map responses to findings
print(f"\n{'='*60}\nRESEARCH VALIDATION RESULTS\n{'='*60}")
for i, f in enumerate(findings[:6]):
    print(f"\nFinding {i+1}: {f.get('finding','')[:120]}")
    print(f"Confidence: {f.get('confidence','N/A')}")
    for resp in data.get("responses", []):
        if resp.get("question","") in f.get("questions", []):
            idx = persona_ids.index(resp["persona_id"]) if resp.get("persona_id") in persona_ids else -1
            name = TARGET_PERSONAS[idx]["name"] if 0 <= idx < len(TARGET_PERSONAS) else "Unknown"
            print(f"  [{name}]: {resp.get('answer','')[:200]}")

Example Output

Report loaded: 142,830 chars (~35,707 tokens)
8 findings, 16 questions — Focus group: 3 personas

RESEARCH VALIDATION RESULTS
============================================================
Finding 1: "By 2027, 75% of enterprise marketers will consolidate
martech from 10+ tools to 3-5 platforms." (Confidence: High)

  [VP Marketing]: Still at 8 tools after 2 years of consolidating.
  Every "platform" has gaps we fill with point solutions.

  [Marketing Ops]: I'm the one migrating 3 years of Marketo workflows.
  The timeline is aggressive.

  [CMO, Startup]: Started with 3, already at 7. Need speed, not consolidation.

Finding 2: "AI-generated content = 30% of marketing output by 2026."
  [VP Marketing]: 20% for first drafts, 0% published-as-is. Depends
  on what "AI-generated" means — if AI-assisted, we're past it.

Error Handling

Long reports

A 150-page report is ~35K tokens — well within limits. For 500+ page reports, use Claude Opus 4.6 with its 1M window. Analysis quality improves with full context versus chunked processing.

Question quality

If generated questions feel generic, add constraints: “Questions must reference specific data points” and “Start with ‘In your experience…’ or ‘At your organization…’”

Response mapping

The response-to-finding mapping relies on exact question string matching. If Mavera modifies question text, fall back to index-based mapping using question order.