hyth.ai · Operating Manual · Appendix A
Transcript 001 — Building the Sales Coach
June 8–9, 2026. One session, lightly cut for length. Everything in Robert's voice is verbatim, typos intact — that's the point. Tool work is summarized in brackets; nothing else is touched. The salesperson's name is abbreviated; prospect names are removed.
I'd like to ingest S—'s sales calls to be transcribed and dissected to spot patterns and optimize our sales training process. I think I will need to give you 5-10 closed sales transcripts to train on and then we can create a end-2-end sales coach tool that grades a call, documents the results and provides insights on how to optimize our sales framework - continuously improving on itself as it notices patterns that increase our close rate. We already have the Zoom API setup, I just need to get S—'s sales recordings to you and build the machine that will be able to understand, process, document, analyze & coach.
This is a meaty initiative. Let me pull up the brainstorming framework before we scope it.
[Claude inventories the existing codebase: finds the Zoom integration already built — OAuth, recording access, transcript download with speaker attribution — and a cron that ingests community calls but has no concept of sales calls.]
The recording/transcription infrastructure is largely built. The new work is: identifying S—'s sales calls, analyzing them against a framework, grading, storing results, and building the feedback loop. […] My recommendation: Start with A, design the schema to support C. Build the grading pipeline with Zoom transcripts + Claude. Store enough structured data (per-stage scores, timestamps, outcome) that a pattern-detection layer can be added in Phase 2 without schema changes. […] Does this framing make sense? Want to adjust before I design the details?
highlights include the things to improve - the coaching is critical.sounds good
Understood — the coaching output is the point, not just a score. "Here's what to do differently next time" matters more than "you got a B+." […]
[Claude walks through the pipeline design: ingestion, transcript extraction, a single structured grading call, storage, an admin panel.]
yes we can do that but the framework and scorecard docs that we have now need some work. we need to review the closed calls to see how we can improve it
That changes the sequencing. The framework and the transcripts need to inform each other before the automated pipeline makes sense. So there's a calibration phase before the grading machine runs […] The transcripts teach us what "good" actually looks like for this product at this price point. […] What's the Zoom topic name on his sales calls?
I think his calls are always title with the prospect's name in it. we will need to investigate
[Claude scans 30 days of Zoom recordings, filters out the community calls, and finds the pattern: two call types, "Discovery Call" and "Strategy Session," each titled with the prospect's name. 16 calls with transcripts in the last month. It asks which ones closed.]
can you cross reference against stripe? he closed most of his deals in december and january
[Claude cross-references months of call recordings against the payments database — matching call titles to paying customers to separate the wins from the losses. 28 sales calls in December, 41 in January.]
[a prospect] was closed long ago, that can be skipped. keep research to find all the closed deals (or at least enough to train our system on
[The detective work continues across several rounds — including Robert correcting the machine's name-matching:]
[a prospect's full name] is [his nickname in the records]. what's next? before we write it? any more prep that would be effective
Good catch on the nickname — that's 12 closed deals, 17 transcripts. Before writing the spec, two things would sharpen it:
1. Pricing discrepancy. You told me earlier what the current offer is. The framework docs say something different — the pricing has clearly changed since they were written. I need the current numbers so the grading system uses the right benchmarks. 2. Run one transcript through the framework. Grading one real call against the full 13-section rubric would prove the transcript quality is sufficient, surface framework gaps, and give us the exact output shape the database needs to store.
the docs will have to be updated for the new price - was [the old pricing] in the call.
Got it — grade historical transcripts against the pricing that was live when the calls happened; the system going forward uses current pricing. I'll build the framework as configurable so pricing updates don't require code changes. Let me run one transcript through the full rubric now.
[First attempt: wrong call. The "sales call" Claude picked turns out to be an existing member — the entire conversation is deal-flow discussion, not selling.]
This is actually an important finding for the tool: not every "Strategy Session" is a sales call. Some are existing-member service calls. The system will need to detect this — cross-reference the prospect against the users table before grading. Let me try a real close call instead.
[Claude grades a real 55-minute closed-won call against the 13-section rubric.]
any further prep to do first? would a codex and opus subagent review now be useful?
Two parallel subagents would be high-value right now. Agent 1: audit the admin panel architecture so the new tab follows the real patterns. Agent 2: run the full 13-section framework analysis on the closed-won transcript — that gives us the exact output shape the schema needs. Both run in parallel. Want me to launch them?
yes
[Two agents launch. While they run, Claude files what it's learned to memory so it survives the session.]
Both agents are back. Here's the grading — it's a goldmine for the spec.
[The graded call lands low on the rubric: the salesperson skipped half the framework's stages and jumped from rapport into a 25-minute product education session — and still closed. The framework and reality clearly disagree about something.]
[…design spec written; two independent adversarial reviews run against it; the plan is revised twice — several hours elided…]
i think we should take note of any initial pattern matching of the closed calls. after following the framework more strictly (see calls from May), he closed less deals
That's the most important data point in this entire session. The adversarial reviewers assumed "follow the framework → better results." If May shows the opposite — more framework compliance, fewer closes — then the framework itself is the problem, not adoption. Let me save this and quantify it.
pull may data, let's look at amother 10 calls. not cherry pick wins
[Claude pulls every May sales call and cross-references against the payments database.]
A losing month against a winning quarter — that's the data. The winter and spring months produced a steady run of closes; May, on similar call volume, produced none worth counting. This is the real A/B test. Now let's grade a batch of these lost calls against the same rubric — that's the comparison the adversarial reviewers said we needed.
[…the build itself: schema, grading pipeline, admin tab, framework viewer — six pull requests over the following hours, including two more adversarial review rounds — elided…]
should S— have a UI where he can mark a "won/lost" toggle
let's add an export button with a few options (last 10 calls, last 30 day, all, (use your best judgement) to download a csv
run the prod migration for me
hte new page /sales-coach doesn't work yet - error something went wrong loading this page. let's also add it to a tab after Challenge & Workshop in /funnel
[The error gets diagnosed and fixed; the work continues through the unglamorous part — merges, a re-merge, cache behavior, a blank score section caught from a screenshot.]
can we cache fetched calls so if I fetch now, S— will be able to see them without another APi fire?
if the user clicks off of the page that says Grading - takes 30s - 60s - will it abort the process or continue to work in the background?
— 18 hours pass —
any housekeeping to do before wrap
All clean. Here's the session summary: What shipped: Sales Coach tool — 6 PRs merged. Live in the admin panel: Zoom call picker, 13-section grading, won/lost toggle, CSV export, framework viewer. Framework v1.0 seeded with current pricing. Production migration applied. What we learned (20 graded calls): Screen sharing correlates with closes. Framework compliance didn't change between the winning and losing periods — the close behavior did. The revised framework's money check gates too early. Six data-backed framework update proposals are documented for the sales team.
ok write a kickoff prompt for a new session for phase 2. then a copy-paste terminal warm-up
[Claude writes the Phase 2 kickoff doc and hands over a one-line prompt for the next session. The next morning, a different session picks it up and starts building the auto-grading cron.]
What this was: roughly a day, start to finish. A business idea typed as one paragraph became a production tool — and along the way the data quietly overturned the assumption the tool was built on. The machine was supposed to enforce the framework. Twenty graded calls suggest the framework needs to learn from the salesman.