Deep Market Research Brief

Setup and handoffs

Complete these in order. Availability reflects the reviewed setup; provider plans can change.

1. 1PerplexityFree tier

   Discover candidate evidence

   Use focused queries to locate candidate sources. Discovery citations are leads, not verified evidence.

   Open Perplexity

   Handoff in

   Research question, geography, period, definitions, competitor names and exclusion rules.

   Exact handoff out

   Candidate source list as `source_id | claim | source title | publisher | source URL | publication date | access date | quoted passage | caveat`. URLs and dates must remain verbatim in later handoffs.

   Numbered setup

   1. 1.1Write a narrow question with geography and cutoff date.
   2. 1.2List preferred primary-source types and exclusions.
   3. 1.3Run separate discovery queries for market, competitors and counter-evidence.
   4. 1.4Export candidate results into the stated table format.
   5. 1.5Treat every row as unverified until a person opens its URL.

   Complete AI prompt

   Find candidate public sources for the research question below. Prefer primary sources, regulator/government data, company filings, official pricing pages and reputable research with a clear date.
   
   RULES
   - Do not treat search snippets or your summary as evidence.
   - Return the direct source URL, not a search-results URL.
   - Never invent a publication date; write UNKNOWN when absent.
   - Quote only text visible in the source and keep quotations under 40 words.
   - Note paywalls, geography, sample limitations and possible conflicts of interest.
   - Assign stable IDs S01, S02, S03... that must survive the handoff.
   
   OUTPUT A TABLE
   source_id | candidate claim | source title | publisher | direct URL | publication date | access date | exact short quote | caveat | source type
   
   RESEARCH QUESTION: [QUESTION]
   GEOGRAPHY: [GEOGRAPHY]
   PERIOD/CUTOFF: [PERIOD]
   KNOWN COMPETITORS: [COMPETITORS]
   EXCLUDE: [EXCLUSIONS]

   Copy Perplexity prompt

   Human approval point

   A person opens every URL, checks the publisher/date/context and copies verified evidence into the canonical source register. Unopened links cannot enter synthesis.

   If this connection fails

   • Citation opens a different claim: reject it and search the primary source directly.
   • Source is paywalled: use only what your reviewer can lawfully inspect; mark inaccessible material as excluded.
   • Dates are unclear: record UNKNOWN and do not use the source for time-sensitive claims.
2. 2Source registerFree tier

   Inspect and freeze the evidence pack

   A person directly inspects every candidate source and creates the only evidence pack the synthesis tool may cite.

   Handoff in

   Candidate source table plus the opened original sources.

   Exact handoff out

   CSV or Markdown rows: `source_id | verified claim | title | publisher | direct URL | publication date | access date | exact evidence | page/section | caveat | verified_by`. Preserve URLs and dates exactly.

   Numbered setup

   1. 2.1Open every direct URL in a separate browser tab.
   2. 2.2Confirm the cited passage supports the proposed claim in context.
   3. 2.3Prefer the original dataset, filing or official page over secondary summaries.
   4. 2.4Record publication and access dates; mark missing dates UNKNOWN.
   5. 2.5Remove rejected sources and freeze stable source IDs before synthesis.

   Human approval point

   The named researcher signs the `verified_by` field for every retained source and confirms no source was accepted from an AI summary alone.

   If this connection fails

   • URL redirects or breaks: locate the publisher's canonical page and record the replacement URL plus access date.
   • Sources conflict: retain both, describe the methodological difference and do not force false consensus.
   • Evidence is only in a chart/PDF: record page, table or figure and manually verify the extracted value.
3. 3ClaudeFree tier

   Synthesize only the inspected source pack

   Create the brief using stable source IDs. Unsupported ideas belong under Unverified assumptions and may not appear as evidence.

   Open Claude

   Handoff in

   The frozen verified source register, research question, audience and decision to support.

   Exact handoff out

   Markdown with `EXECUTIVE SUMMARY`, `EVIDENCE`, `COMPETITOR TABLE`, `SWOT`, `UNVERIFIED ASSUMPTIONS`, `GAPS`, and `SOURCE REGISTER`. Every evidence statement cites `[S##]`; URLs and dates remain unchanged in the register.

   Numbered setup

   1. 3.1Start a new private chat and paste the complete synthesis prompt.
   2. 3.2Paste only the frozen, human-verified source register.
   3. 3.3Add the research question and intended decision.
   4. 3.4Generate the brief and compare its reproduced source register with the frozen original.
   5. 3.5Reject any section containing an unknown source ID or unsupported factual language.

   Complete AI prompt

   Create a decision brief using ONLY the VERIFIED SOURCE REGISTER supplied below.
   
   EVIDENCE RULES
   1. Cite every factual statement as [S##]. Use only source IDs present in the register.
   2. Do not change, shorten, reconstruct or replace source URLs, publication dates or access dates.
   3. Do not present your background knowledge, inference or plausible explanation as sourced evidence.
   4. Put any inference in UNVERIFIED ASSUMPTIONS with its reasoning and what evidence would confirm it.
   5. If sources conflict, describe the conflict and each source's scope; do not choose silently.
   6. Distinguish company claims from independent evidence.
   7. Write UNKNOWN when the source pack cannot answer a question.
   8. Never fabricate market size, growth rate, pricing, customer count, quote or source.
   
   OUTPUT EXACTLY
   # EXECUTIVE SUMMARY
   Five bullets, each cited or explicitly marked inference.
   # EVIDENCE
   Claim | evidence | source IDs | confidence (high/medium/low) | caveat
   # COMPETITOR TABLE
   Competitor | positioning | verified pricing/date | evidence IDs | unknowns
   # SWOT
   Each item must include evidence IDs or the label UNVERIFIED ASSUMPTION.
   # UNVERIFIED ASSUMPTIONS
   Assumption | reasoning | evidence needed
   # GAPS
   Questions not answered by the source pack.
   # SOURCE REGISTER
   Reproduce source_id, title, publisher, direct URL, publication date and access date exactly.
   
   RESEARCH QUESTION: [QUESTION]
   AUDIENCE/DECISION: [AUDIENCE AND DECISION]
   VERIFIED SOURCE REGISTER:
   [PASTE VERIFIED SOURCE REGISTER]

   Copy Claude prompt

   Human approval point

   A person clicks every citation in the finished brief, verifies the claim and context, then approves interpretations and recommendations separately from sourced evidence.

   If this connection fails

   • Model cites an unknown ID: reject the affected section and rerun with the allowed ID list repeated.
   • URLs or dates change: restore them from the frozen register; the synthesis output is not authoritative metadata.
   • Assumption appears as fact: move it to Unverified assumptions and lower confidence until evidence is added.

Test safely before real data

1. 1Choose a low-stakes historical question with three known primary sources.
2. 2Include one broken link, one conflicting source and one source with no publication date.
3. 3Confirm the workflow excludes uninspected sources and preserves every URL/date through synthesis.
4. 4Check that an intentionally unsupported claim lands only in Unverified assumptions.
5. 5Have a second person trace every final citation before using a live research question.