Claude Builds the Deliverable Before the Call Ends

You finish a sales or scoping call. You hang up. And then the worst part of the job begins: you scroll back through your own notes, try to remember what they actually asked for versus what you think they asked for, and start a blank doc. That gap — between the call ending and the work starting — is where momentum and accuracy both die. The fix is small and durable: your AI notetaker already has a perfect transcript. If Claude can read that transcript directly, it can extract the build spec from what was literally said and start the deliverable while the call is still warm. The key idea: the spec comes from what they said, not from your memory of it. You wire this once and every future call gets the same treatment.

Claude can reason brilliantly about a transcript — but only if it can reach one. An MCP (Model Context Protocol — Anthropic's open standard for letting an agent call external tools) is the bridge. The course uses Fathom, an AI notetaker that records and transcribes meetings and ships a real Claude connector; the same flow works for any notetaker that exposes an MCP. There are two paths, depending on whether your notetaker is already in Claude's connector directory.

1. Easy path (directory connector): In Claude, open Settings → Customize → Connectors (the menu label moved from 'Settings' to Customize on recent builds — if you don't see it there, check under Settings directly). Find your notetaker — Fathom ships an official Claude connector — click Connect, and complete the OAuth prompts to authorize your account.
2. Custom path (paste a URL): If your notetaker isn't in the directory, in the same Customize → Connectors screen click the + (Add custom connector), give it a name, and paste the server's remote MCP URL (always https://… with the correct path). Some servers also want an OAuth login or an API key.
3. Authorize when prompted. The connection runs from Anthropic's servers to the notetaker's — you're granting Claude read access to your meetings.
4. Restart Claude. New MCP tools register on launch — if you skip the restart, the tools simply won't show up and you'll think the connector is broken.
5. Verify: open a new chat and ask 'what tools do you have from <notetaker>?' — you should see meeting/transcript tools listed (e.g. list/search meetings, get transcript, get summary).

Once connected, Claude gets a set of meeting tools it can call on its own. The exact names depend on which server you use. Fathom's official connector exposes your meeting summaries, transcripts, and action items (it describes the capability rather than publishing tool names). A widely-used community Fathom MCP server (agencyenterprise/fathom-mcp-server) publishes its tools explicitly — handy to know so your prompt can name them:

This is the whole trick in one message. The instant the call ends, open Claude and paste the prompt below. It tells Claude to pull the transcript itself, extract only what was actually said, and turn it into a build spec — not a summary. Keep the framing strict: a spec is decisions and requirements, a summary is vibes.

• The prompt: Pull the transcript of my last meeting (use my notetaker connector). From the transcript ONLY — not your assumptions — extract exactly what to build. Output: (1) the one-line deliverable, (2) a numbered list of concrete requirements the client stated, each with the quote it came from, (3) explicit constraints (deadline, budget, must-haves, must-NOT-haves), (4) open questions where the call was ambiguous, (5) anything I implied I'd send that isn't a feature. Do not invent requirements. If something wasn't said, put it under open questions.
• Why 'with the quote it came from': forcing Claude to cite the line keeps it honest. If it can't point to where a requirement was stated, it's a hallucination — and you'll see it instantly.
• Why 'open questions' is mandatory: the most dangerous output is a confident, complete-looking spec that quietly papered over an ambiguity. Making Claude surface what was unclear turns gaps into your follow-up email instead of a wrong build.
• Name the meeting if 'last meeting' is wrong: back-to-back calls confuse 'last.' Say 'the call with Acme this afternoon' and Claude will use search_meetings to grab the right one.

A spec sitting in chat is still just notes. The payoff is feeding it straight into a scaffold so the deliverable literally starts existing before you've made a coffee. Keep the two steps separate — extract, then build — so a bad extraction never silently becomes a bad build.

1. Confirm the spec first: read Claude's extracted requirements, resolve the open questions (a quick reply to the client, or your own decision), and lock it. This 60-second check is the whole reason you forced citations and open-questions.
2. Then instruct the build: 'Using the spec above, scaffold the first version of the deliverable — <name the artifact: a proposal doc, a landing page, a project plan, a quote, a slide outline>. Follow the requirements exactly; where the spec has an open question, leave a clearly-marked TODO rather than guessing.'
3. Let Claude produce the actual artifact (a doc, an outline, code, a page) — not a description of one. A client opens a file; they don't open a transcript of your intentions.
4. Iterate against the spec, not your memory: 'requirement 4 says weekly reporting — the draft only has monthly, fix it.' The spec is now your single source of truth for the whole deliverable.

This workflow is fast, which is exactly why it can be confidently wrong. A 'good' run looks like: every requirement traces to a real quote, the open-questions list is honest (an empty one on a complex call is a red flag, not a win), and the draft deliverable matches the spec line for line. Here's where people get burned.

1. You forgot to restart Claude. The connector's tools don't appear until you relaunch — the #1 'it's not working' cause. Restart, then re-check the tools list.
2. OAuth not completed / token expired. If Claude says it can't access meetings, re-run the Connect/authorize step. Directory connectors silently lose auth; reconnect and try again.
3. The spec is from memory, not the transcript. If Claude starts listing requirements with no quotes, it's filling gaps from training instead of the call. Re-prompt: 'use the transcript tool — quote the line for every requirement or move it to open questions.'
4. 'Last meeting' grabbed the wrong call. On a busy day, name the meeting explicitly so it searches instead of guessing.
5. You shipped the draft as-is. The agent extracts and drafts at best-effort. You resolve the open questions and you approve before it reaches the client. Speed is the point; skipping the human check is not.