⚡ GPT-5 Mini Rewrite (V2)

Overview

V2 rewrites the existing seo_content from your service JSON using gpt-5-mini-2025-08-07. No web search — with a choice of 4 article plans (A–D), 13 sections including mandatory FAQ (§12), and 3 built-in improvements for data accuracy and quality.

Architecture

┌─────────────────────────────────────────────────────┐
│  SERVICE DATA (name, category, notes, keywords…)    │
└────────────────────────┬────────────────────────────┘
                         │
         ┌──────────────▼──────────────────┐
         │         SINGLE API CALL          │
         │                                  │
         │  Rewrite Prompt (Plan A/B/C/D)   │  gpt-5-mini
         │  + Service DATA block            │  max_completion_tokens=8000
         │  + SOURCE CONTENT (seo_content)  │  streaming=true
         │  + 13-section structure + FAQ    │
         │  Reformulate sentences, improve  │
         │  structure & Billoff angle       │
         └──────────────┬───────────────────┘
                        │
                  REWRITTEN ARTICLE
                  (HTML, ~1,800 words, 13 H2, 2+ tables, FAQ)

Prompt Context Injected

V2 assembles 4 layers of context for each article:

── SYSTEM MESSAGE ──────────────────────────────────────────────
══ VERIFIED SERVICE REGISTRY (improvement 1) ══
notice_period:      30     ← pre-populated from svc.notice_period_days
early_exit_fee:     null   ← null → "check current terms"
cooling_off_window: 14
cancellation_url:   "https://..."
regulator_name:     "ACCC"
Rule: null field → write "check current terms at {name} website" — NEVER estimate.

── USER PROMPT ─────────────────────────────────────────────────
══ EMPATHY OPENER (improvement 2) ══
Cancelling {name} in {country} should be simple — but between
contract lock-ins, notice periods, and unexpected fees, many people
find the process harder than it should be. This guide gives you
every step, every right, and every shortcut you need...

══ SERVICE DATA ══
Name: {name} | Keyword: {main_keyword}
Category: {category} | Website: {website}
Cancellation methods: phone, email, account-online
Country: {country} | Currency: {currency} ({symbol})   ← from locale
Language: {language} | Action: {cancel_word}         ← from locale

══ SOURCE CONTENT (rewrite — do NOT copy sentences verbatim) ══
{existing seo_content from service JSON, up to 4,000 chars}

After generation, a quality gate checks the output: min 1,800 words AND min 10 numeric data-points. If below either threshold and V3_FALLBACK_ENABLED = True, the article is automatically re-generated via V3 (Gemini), and the escalation reason is logged.

Heading capitalisation — enforced in system message + QUALITY CHECK self-scan: sentence case on every H1/H2/H3/H4. First word capitalised, proper nouns capitalised, everything else lowercase. Post-generation sanitiser applies as final safety net.

── DATA RETENTION RULE (improvement 4, injected before ══ OUTPUT ══) ───
══ DATA RETENTION RULE ══
The source content contains 37 specific data points. Your rewrite MUST
preserve a minimum of 14 of them verbatim.
RULE: NEVER generalise specific figures. If the source says "$150 exit fee",
write "$150" — NOT "a fee may apply".
KEY FIGURES FROM SOURCE: $14.99 | $24.99 | 30 days | 14 days | 10% | ...
✅ Self-check: count specific numbers/dates/durations — must be ≥ 14.
# min_required = max(10, floor(source_count × 0.4)) — e.g. 37 × 0.4 = 14
# Cost impact: neutral to +€0.001/article (token overhead <5%)

When to use V2

• Large-scale generation of 100–10,000 articles where cost matters
• Services where pricing stability is low (model knowledge may be slightly outdated)
• Initial content generation before manual review
• Categories with well-known services (streaming, software, SaaS)

When NOT to use V2

• Niche or local services the model may not know well
• Services with very recent pricing changes
• When you need verified, sourced competitor data

🌍 Multi-country & Locale Support

V2 generates articles in any language and currency — no code changes needed. Add locale fields to each service row in your JSON; the script resolves them automatically via get_locale(svc).

Fallback chain: svc.country → editor config → script defaults. Any missing field is silently defaulted — you can migrate countries incrementally.

Output JSON now includes cost_usd, cost_eur and elapsed per article, plus a total cost summary at the end of the run.

Cost Breakdown

Model pricing: Input $0.40/1M tokens, Output $1.60/1M tokens (gpt-5-mini-2025-08-07).

Python Script Reference

# Run V2 batch (all 20 sample services, 5 parallel workers)
python scripts/02_generate_v2.py

# Output: Billoff/data/results_v2.json

# Run 3-method comparison on 1 service
python scripts/test_compare_3methods.py --service "Netflix"

Parallelisation

# V2 runs with 5 concurrent workers (no rate limit concern for direct completions)
from concurrent.futures import ThreadPoolExecutor
MAX_WORKERS = 5

with ThreadPoolExecutor(max_workers=MAX_WORKERS) as ex:
    futures = {ex.submit(process_service, svc): svc for svc in services}
    for future in as_completed(futures):
        results.append(future.result())

Quality Results (Ocado test)

Pros & Cons

Writing Persona

A writing persona is appended to the system message for every article. Set PERSONA_ID at the top of the downloaded Python script to switch voice. Available personas: cancellation_specialist (default), consumer_rights_expert, contract_lawyer, financial_advisor. See V1 Docs for the full persona reference.

🔬 AI Comparative Analysis (after all 4 methods)

Once all 4 methods complete, the Lab runs a 2-phase parallel analysis and displays a full comparison report. The analysis is automatically saved to history alongside the article results.

Total cost per analysis: ≈€0.04–0.06. See V1 Docs for the full specification.