deep-research-synthesis

name	deep-research-synthesis
description	Use to synthesize raw Apify dumps into the seven (eight including starter template) sub-task markdowns per niche. The Claude-side post-processing layer for Module 2B.

Deep Research Synthesis

Takes raw Apify JSON dumps from research/02-niche-research/{slug}/raw/ and produces the analyst-grade markdowns in the niche folder.

Inputs

• raw/agencies-search.json and raw/agencies-crawl.json → produces 01-agencies.md
• raw/maps-reviews.json + raw/reddit.json + raw/trustpilot.json (optional) → produces 02-customer-voice.md
• raw/fb-ads.json + raw/business-homepages.json → produces 03-copy-patterns.md
• raw/cro-crawl.json + screenshots → produces 04-cro-patterns.md
• Cross-references all of above → produces 05-trust-signals.md
• raw/serp-rankings.json + raw/gbp-density.json → produces 06-seo-landscape.md
• Recipes 1 + 2 outputs → produces 07-money-math.md
• After /pick-niche, synthesizes all of above → produces 08-starter-template.md

Synthesis rules

Lens: end customer first

Every output answers a question about the end customer (the niche business's customer, not the niche business itself or the student's agency).

• Customer voice extracts verbatim phrases the end customer uses about their pain or decision.
• Copy patterns identifies what hero language gets end customers to convert.
• CRO patterns identifies what wins end-customer trust and clicks.
• Trust signals identifies what end customers want to see to feel safe.

Verbatim over paraphrase

When pulling customer language from reviews or forum posts, quote verbatim. Include the source URL or post ID. Paraphrases lose the niche's actual voice.

Frequency × intensity × addressability

When ranking pains, hooks, or trust signals, score on:

• Frequency: how often it appears across sources
• Intensity: how strong the language is when it does appear
• Addressability: how much a website can move the needle on it

Top items in each output need high scores on all three.

Pattern over outlier

When analyzing competitor sites or ads, prefer the patterns that repeat across multiple winners. A single outlier hero or CTA copy line is interesting but not foundational.

Output discipline

• Every markdown ends with a "Source traceback" section listing which raw files were used.
• Every claim cites the source. Don't say "competitors all do X" without naming which ones and pointing to the raw file.
• Mark anything inferred (vs directly observed) with [inferred].

Use of LLM judgment

The agent uses Claude itself to analyze the raw JSON. Workflow:

1. Read raw JSON (in chunks if large).
2. Generate the analysis section by section, one section at a time.
3. After each section, validate against the rules above.
4. Write the final markdown.