transcript-research

name	transcript-research
description	Orchestrate deep research across meeting transcripts in the local transcript vault. Use when the user asks to analyze client meetings, review sales calls, research a topic across meetings, build client personas, export transcripts for RAG, or says 'meetings', 'transcripts', 'call review', 'client analysis'.

Transcript Research

Orchestrate multi-step research workflows over the local transcript knowledge base (transcript-vault-kb MCP server). The store is a SQLite + LanceDB index built from your recorder's exports; all tools are read-only.

The core risk of this skill is not missing data. It is producing confident synthesis from a single slice of retrieval. The discipline sections below (Triangulation, Attribution, Temporal, Roleplay, Checkpoint) exist because the default failure mode is pattern-matching across meetings and silently dropping speaker, time, and source into a plausible-sounding narrative. Follow them.

Available MCP tools (transcript-vault-kb)

Tool	Purpose
list_meetings	List/filter meetings by date, person, company, meeting_type, relationships, call_types
get_meeting	Full metadata + summary + participants for one recording_id
find_person	Search participants by name/email substring
find_company	Search companies by domain substring
list_people_by_relationship	Enumerate every person with a given relationship tag
keyword_search	FTS5 over meeting titles and summaries (use for precise terms, names, acronyms)
semantic_search	Vector search across transcript chunks and summaries (use for conceptual questions). Supports min_distinct_meetings to diversify.
get_entity_timeline	Chronologically-ordered mentions of a person/company/project/topic
get_project_status	Latest known state of a named project/deal/initiative, with history
verify_claim	Scan for supporting, contradicting, and unclear evidence for a claim
get_transcript	All indexed transcript chunks for one recording_id, ordered by timestamp (optional speaker filter — limited until speaker attribution is improved)
build_context	Run a semantic search and return a markdown-formatted context block

Available prompts: client_persona, sales_coaching, content_from_meetings, account_prep.

Core discipline

Most failures of this skill are not retrieval failures — they are synthesis failures on top of thin retrieval. The five sections below are non-optional.

A1. Triangulation protocol

Any claim about user state, narrative, or a recurring pattern must pass through:

• At least 2 of: semantic_search, keyword_search, find_person / find_company, list_meetings, get_entity_timeline.
• Results from at least 3 distinct meetings (pass min_distinct_meetings: 3 on semantic_search), or explicitly note "only 1 meeting found" in the output.
• A temporal bracket: is this from the last 30 / 90 / 365 days? State it.

A single semantic_search call is only sufficient for "find me the meeting where X happened"-style retrievals. It is not sufficient for "what does the user think about X", "is this deal moving", or "what has the user said recently about Y". Those require triangulation.

If a pattern shows up in 1 meeting, it is an anecdote. If it shows up in 3+, it is a pattern.

A2. Attribution discipline

• Every quote must cite speaker + meeting title + date.
• If the transcript chunk shows Unknown: or Unknown as the speaker, do not attribute the line to anyone. Either:
  • Re-derive the speaker from context (e.g., two-party call where the other name is explicit), or
  • Drop the quote entirely.
• When describing someone's beliefs, positions, or narrative, strictly separate three categories:
  • "Things they said" — direct quote with citation
  • "Things said to them" — another participant's words, never recast as theirs
  • "Inferences" — explicitly labeled as inference, not fact

Meeting titles and meeting summaries are more reliable for attribution than raw transcript chunks. Prefer get_meeting(include_summary=true) when attribution matters and the transcript is ambiguous.

The classic failure mode: a coach says to a user "after six years of struggling you should…" — and later a search surfaces this quote, and it gets recast as something the user said about themselves. Don't do this.

A3. Temporal resolution

• meeting_date = when the conversation happened.
• event_date = when the thing being discussed happened (often earlier, sometimes later/planned).
• Status claims ("the project is X", "the deal is Y", "they are working on Z") must be based on the most recent meeting where the topic appears, not whichever meeting surfaced first in a search.
• For any ongoing entity (project, deal, relationship), run a "latest mention" query before stating status. Use get_project_status(project) for named projects, or get_entity_timeline(entity, ...) and read the last row.
• A summary line written on date D may describe an event that happened weeks earlier. Do not treat summary recency as event recency.

A4. Roleplay guardrails

When the user asks this skill to "be me", simulate their voice, or respond as them:

• Every factual claim about the user's past requires a citation.
• Every emotional or positional claim requires a grounded quote or an explicit "I'm inferring this from X."
• When the user corrects the agent, the first response is a targeted re-search (narrower query, more specific time window, direct get_meeting on the cited source), not a quick narrative reframe.
• Never attribute another participant's stance or narrative to the user just because it appeared in a meeting they were in.
• If you catch yourself generating dialogue in the user's voice without a source for the underlying claim, stop and search first.

A5. Checkpoint before synthesis

After any research burst of 3+ tool calls, before producing the final answer, produce an internal three-column note:

• Known — cited, attributed, dated.
• Inferred — plausible extrapolation from known.
• Unknown / would want to verify — gaps.

Only synthesize once this is clear. If "Inferred" is larger than "Known", do more searching. For any claim pulled up to the synthesis that lives in "Inferred" rather than "Known", either demote it ("likely…", "the pattern suggests…") or drop it.

For high-stakes claims (project status, deal status, user position on a topic), pass the claim through verify_claim before asserting it.

---

Workflow 1: Client / company analysis

1. find_company(query="acme") to confirm the domain.
2. list_meetings(company="acme.com", limit=50) for the relationship timeline.
3. get_entity_timeline(entity="acme", entity_type="company") for a chronological snippet-level view.
4. semantic_search with company="acme.com" + min_distinct_meetings=3 for each persona dimension:
   • "pain points and frustrations"
   • "goals and desired outcomes"
   • "objections and concerns"
   • "decision-making style"
   • "current priorities"
5. For the 2–3 highest-signal meetings, get_meeting(recording_id, include_summary=true).
6. Apply the checkpoint (A5) before writing the profile.

Use the client_persona prompt for a guided version.

Workflow 2: Sales call coaching

1. get_meeting(recording_id, include_summary=true) — context.
2. get_transcript(recording_id) — full conversation ordered by timestamp.
3. Analyze:
   • Talk ratio — count speaker turns per speaker (skip Unknown).
   • Question quality — open vs. closed from the rep.
   • Discovery depth — pain, budget, timeline, decision process.
   • Objection handling — semantic_search(query="objections concerns pushback", date_from=<meeting date>, date_to=<meeting date>).
   • Commitments / close — semantic_search(query="commitments next steps", ...).
4. Output a structured review with timestamps and quotes. Follow A2 strictly — if a line is Unknown, don't attribute it.

Use the sales_coaching prompt for a guided version.

Workflow 3: Topic / pattern research

When the user wants to investigate a topic across all meetings:

1. keyword_search(query=<exact terms>, limit=15) for precise hits (names, acronyms).
2. semantic_search(query=<phrasing 1>, min_distinct_meetings=3, limit=10) for conceptual hits. Run 2–3 differently-phrased queries.
3. get_entity_timeline(entity=<topic>, entity_type="topic") for chronological structure.
4. For top 3 meetings, get_meeting(recording_id, include_summary=true).
5. Apply triangulation (A1), attribution (A2), temporal (A3), and checkpoint (A5).
6. If the synthesis will make a strong claim (e.g., "the user's view is X"), run verify_claim(claim=<claim>) before asserting.

Cite every finding as [Meeting Title, YYYY-MM-DD](url) with speaker where known.

Workflow 4: Project / deal status

When the user asks "what's happening with project X" or "is the X deal still on":

1. get_project_status(project="X") — returns latest known status + history.
2. If status is unknown, fall back: get_entity_timeline(entity="X", entity_type="project") and read the last rows.
3. get_meeting(recording_id) on the most recent mention to confirm.
4. Never rely on an older meeting for status when a newer one exists. This is the Karim-cancellation class of error.

Workflow 5: Account prep

1. find_company or find_person to resolve the target.
2. list_meetings(company=..., limit=15) or list_meetings(person=..., limit=15).
3. get_meeting(include_summary=true) for the 3 most recent.
4. semantic_search with company / person filter + min_distinct_meetings=3 across:
   • "open action items commitments"
   • "unresolved concerns blockers"
   • "pending decisions approvals"
5. For any ongoing project mentioned, call get_project_status to confirm current state.
6. Apply the checkpoint (A5). Output brief.

Use the account_prep prompt for a guided version.

Workflow 6: Content mining

1. keyword_search with 2–3 precise terms (limit=10).
2. semantic_search with 3 different phrasings (limit=6 each, min_distinct_meetings=3).
3. build_context(query=<strongest>, limit=8) for a consolidated markdown block.
4. Pull 3–5 short direct quotes — each with speaker (A2) and date.
5. Do not invent quotes. Paraphrase is labeled as paraphrase; only verbatim strings in quote marks.

Use the content_from_meetings prompt for a guided version.

---

Query patterns

When single-slice is enough

• "Find the meeting where we discussed the pricing change" — semantic_search or keyword_search, one call.
• "What did this meeting cover" — get_meeting(include_summary=true), one call.
• "Who was in the X call" — get_meeting.

When you must triangulate

• "What does the user think about X" — A1 applies.
• "Is deal X still active" — A3 + Workflow 4.
• "What has person X been saying about Y" — get_entity_timeline(entity="X", entity_type="person") plus targeted semantic_search with person="X".
• "Summarize my recent business thinking" — multiple queries across business-model, prospecting, pricing, ideal-client themes; min_distinct_meetings=3 on each; checkpoint before output.

Query construction

• semantic_search queries should be natural-language phrases, not keyword bags. "Why my pipeline is slow" > "pipeline slow".
• keyword_search uses FTS5 syntax: AND, OR, NOT, quoted phrases, parens.
• If nothing returns, relax the query (drop qualifiers), then widen the date range, then try a different phrasing — in that order.

Known data-quality caveats

• Many transcript chunks show Unknown as the speaker. This is an upstream data issue. Respect A2 — do not invent attribution. Summaries (get_meeting(include_summary=true)) are more reliable than raw chunks for who-said-what.
• Entity resolution (e.g., "Karim" vs. "Kareem" vs. full name) is not always clean. For any search where the exact name matters, try multiple spellings.
• Project/deal status is not a first-class field — get_project_status infers from mentions. Verify with a direct get_meeting on the most recent hit before asserting.