Transcript Synthesis

Step 1 — Get the transcript

Accept whichever form the user provides:

• Pasted text — already in context, proceed
• File path — Read the file (supports .txt, .md, .json, .vtt, any text format)
• Freeform notes — ask the user to share what they remember from the call

If nothing has been provided yet, ask:

“Share the transcript or notes — paste them here or give me a file path.”

Step 2 — Extract ICP attributes

Read the full transcript and extract every ICP signal mentioned. Organize into these dimensions:

Firmographics

• Industry / vertical (specific is better than broad: “mortgage servicers” not “finance”)
• Company size (employee count, revenue, or both if mentioned)
• Geography (country, region, state — be specific)
• Funding stage (bootstrapped, Series A, enterprise, etc.)
• Other company attributes (public/private, age, growth rate, etc.)

Technographics

• Specific tools, platforms, or tech stack mentioned
• Integrations or compatibility requirements

Buying signals & triggers

• Events that indicate readiness (hiring, funding, expansion, regulatory change, etc.)
• Timing triggers (“companies that recently…”)

Disqualifiers

• Explicit exclusions mentioned (“not X”, “we don’t want Y”)
• Deal-breaker criteria

Target contacts

• Job titles, seniority levels, departments
• Decision-maker vs. champion distinctions if mentioned

Reference companies

• Any example companies or domains mentioned as “companies like these”
• Existing customers named as exemplars
• Competitors’ customers mentioned

Quality vs. volume preference

• Did the prospect indicate they want a tight high-precision list or broad coverage?
• Any stated list size targets?

Step 3 — Present the extraction table

Show all extracted attributes in a single table:

Then ask:

“Does this capture everything? Add anything missing, correct anything wrong, then I’ll ask you to prioritize each attribute.”

Wait for confirmation before proceeding.

Step 4 — Classify must/should/nice

For each attribute, determine priority:

• must — non-negotiable; a lead without this doesn’t qualify
• should — improves fit; worth filtering for but not a hard gate
• nice — enrichment only; good to have, not required

When the transcript is explicit (“we absolutely need X”, “X is critical”, “X would be nice to have”) — capture that classification verbatim.

When ambiguous — mark as unclassified and surface those to the user.

Present the unclassified attributes and ask the user to classify each:

“These attributes weren’t explicitly prioritized in the transcript. Classify each as must / should / nice:”

Do not silently default. Never assume must for everything (over-conservative) or should for everything (under-conservative).

Once all attributes are classified, show the final table for confirmation:

Ask: “Anything to promote, demote, or change? Reply with edits or ‘looks good’ to save.”

Step 5 — Surface next steps

Return the classified attribute table to the calling context. If invoked standalone (not from icp-capture):

• Capture the full ICP → run icp-capture to validate scope, check for similar builds, and write icp-definition.md
• Build a prospect list → use prospect-builder with the classified attributes
• Find lookalike companies → if seed domains were captured, use lookalike-expansion

Red Flags — STOP

• About to classify all attributes as must → that’s over-conservative; ask the user to prioritize
• About to skip Step 4 because the transcript seemed clear → always confirm the table with the user
• Transcript mentions a company by name but no domain → try to infer the domain or note it as unresolved
• Quality vs. volume preference not mentioned → ask explicitly; it affects how the search query is constructed