Sales Machine — what's left
Three pieces of work that aren't finished, written so anyone can pick them up cold:
what it is, why it matters, what already exists, what blocks it, and the first move.
Everything else from this week is shipped and live. 13 Aug 2026.
Handoff
The short version
| Piece | State | What it's worth |
|---|
| A stack of outreach drafts | Engine works, only 2 drafts written |
Sean spends ~45 minutes per intro email today |
| Meeting-promise watcher | Not started |
Promises made on calls stop evaporating |
| Chat in the sales app | Not started; decision pending |
Ask the pipeline anything, in plain language |
Also worth knowing: the console, the inbox watcher, and the prospect data are all
finished and live. These three are the only open threads.
1 · A stack of outreach draftsHalf done
Give Sean twenty drafts waiting, not one at a time
Sean spends around 45 minutes writing each first-contact email. For the Situs
intro he clicked around the company's website until he found they run AppFolio — a fact we
already had on file. The machine should hand him a draft that already knows that.
What exists
- His real sent emails from this week are captured as the reference set — the examples the
machine imitates.
- A written breakdown of what makes his emails work: the structure, the register, and the
five places a personal detail slots in.
- A generator that writes one draft in his voice from a company's verified facts. It runs on
our Claude subscription, never a metered key, and it never sends — it writes a file
with a DRAFT banner for Sean to review, edit, and send from his own mailbox.
- Two finished sample drafts: Archway Equities and Baron Properties.
The blocker, and it's the whole job. The generator reads our deep company profiles,
and only 15 of those exist. Of the 15, exactly two were never contacted — Archway and Baron —
and both are already drafted. So the generator has literally run out of companies. It cannot
produce a stack until it can read the wide list too.
Where the volume actually is
The wide list has 514 researched companies. Of those:
| Slice | Count | Why it matters |
|---|
| Have a named contact person | 178 | An email needs someone to address |
| …and are owner-operators | 72 | Our best-fit buyer |
| …and we know their software | 46 | The personal detail that made Situs land |
Those 46 are the batch. That's a realistic first run.
First move
- Teach the generator to build its fact sheet from a wide-list row (company, contact, software,
portfolio size, city) when no deep profile exists — the deep profile stays the better input
when we have one.
- Add a batch mode that writes all 46 in one run.
- Put them in front of Sean as a reviewable stack, and ask him to mark each one
send / fix / bin. His edits are the training signal for the next version.
Guardrail that does not move: the platform drafts, a human sends. No draft leaves
the building without Sean pressing send himself.
2 · The meeting-promise watcherNot started
When Sean says "I'll follow up with them," that becomes a real task
Every call generates promises — "I'll send them the deck," "I'll check back after
the board meeting." They live in a transcript nobody re-reads, and they quietly die there. The
watcher reads yesterday's calls and turns each promise into a next step on the right company.
How it would work
Read yesterday's call transcripts
↓
Find every commitment — who owes what, to which company, by when
↓
Match it to a company we already track
↓
Write it as that company's next step, with a link back to the moment it was said
↓
One morning summary: "3 promises found, 2 already done, 1 overdue"
What exists to build on
- A read-only reader that pulls whole call transcripts from our transcript channel — already
written and used.
- Seven calls already captured and filed, so there's real material to test against.
- Every company record already has a "next step" field the watcher would write into. Nothing
new to invent.
- The same shape as the inbox watcher we shipped this week: read, decide narrowly, propose a
change a human approves. Copy that pattern.
The trap to avoid. Always read the raw, speaker-by-speaker transcript — never the
auto-summary at the top of the thread. Summaries drop exactly the throwaway sentence where the
promise was made. This has bitten us before.
Open call for Fede
Where should a found promise land — as the company's next step inside the sales app, or as a
card on the Trello board where the rest of your todos live? The app keeps it next to the deal;
Trello keeps it next to everything else you owe. My lean: the app, because these are deal
follow-ups, not personal todos.
3 · Chat in the sales appNot started
Ask the pipeline anything, the way you'd ask ChatGPT
You asked for an exact copy of the ChatGPT experience, inside the sales app: ask
about any company, ask who to call this week, ask it to rewrite an outreach draft — in
conversation, not through filters and columns.
What exists
- The data it would answer from: 514 researched companies, 15 deep profiles, seven call
transcripts, and every captured email thread.
- A working, tested way to talk to Claude on our subscription.
- A written spec of Sean's voice, so "rewrite this draft" already has a standard to hit.
Why this one isn't just "go build it." The console is a single static page — a document,
not an app. There's nothing behind it to receive a question. And our rule that all AI runs on
the subscription means a cloud server has no credential to think with. Chat needs both problems
solved, which is why it's Decision 5 in the v2 architecture doc and not yet in progress.
The three options, unchanged from that doc
| Option | Trade |
|---|
| A | Chat answers instantly while Fede's Mac is on; when it's off, it says so
and holds the question | Honest, no rule change, and the Mac is on during the hours Sean
would use it. My recommendation. |
| B | One carve-out from the subscription-only rule: a metered key for chat
alone, with a hard monthly cap | Truly always-on, but it changes a standing rule. Your
call, not a default. |
| C | Defer chat; ship the draft stack first | Lowest risk. Sean gets the
45 minutes back now and the conversation later. |
This is the only one of the three that's genuinely blocked on a decision from you.
The other two can start tomorrow.
Suggested order
Drafts first. It's the only one with a number attached — 45 minutes per email, and 46
companies ready to receive one. The engine already works; it just needs to read the wide list.
Watcher second. Small, self-contained, and copies a pattern we shipped this week.
Its value grows with every call, so starting it early compounds.
Chat third, once you've picked A, B, or C — and easier after v2 gives the console a
real backend, which it needs anyway for the board to save a drag.