NightDraft at Claude Founder House

On Wednesday 23 September I am taking NightDraft to Claude Founder House in London: the afternoon Builder Cohort run by the Claude Startups team at Anthropic. A masterclass, a guided build from a blank editor to working software, one-to-ones with the Applied AI team, and then a community showcase and pitch-off where selected builders demo live and the room votes. This page is the demo in writing - what I would say if the wifi died.

NightDraft is an AI that drafts a restaurant's booking emails in the owner's voice. It has been live at Lorne, a neighbourhood restaurant in Victoria, since June 2026, and Lorne has been a paying customer since August. Nothing it writes is ever sent by a machine.

The problem, in one email

A table of twelve on a Friday, with wine, is worth a couple of thousand pounds to a small restaurant. The enquiry arrives by email, often in the evening, and whoever replies first tends to win it. At Lorne the owner answered every one of those emails herself, after service, and she cared far more about never getting a detail wrong than about the time it took. That is the job. Not faster typing: a first reply that is quick, right on every policy, and unmistakably hers.

What it does

A poller watches the bookings inbox. Each new email is classified - enquiry, follow-up, cancellation, supplier, noise - and the ones that matter go through a deterministic rules layer that applies the restaurant's playbook: party-size tiers, deposits, private-dining minimums, which menu to attach. Claude then drafts the reply in the owner's voice, and a verifier checks the draft against the playbook before it lands in her Gmail drafts folder, addressed and ready. She reads it, edits if she wants to, and presses send. Enquiry to draft takes well under a minute.

Values before rules

The first version did what most people do with AI writing: it gave the model writing rules. Be warm, no exclamation marks, sign off like this, quote these prices. It followed every rule and still made the wrong calls, because rules tell you what to do and never why. We measured it the honest way, against what she actually sent, and only about two thirds of each draft survived. She was rewriting a third of everything.

So we flipped the order. Before the model sees a single writing rule it now has to understand who it is writing as and what the restaurant stands for, and only then how she writes. The same rules sit on top of a person, and the model can finally make the judgement calls between them. On a small early sample, draft survival went from roughly 65-70 per cent to about 85 per cent, and her edits halved.

The cold start

Onboarding is the part I am proudest of. Connect the mailbox and NightDraft reads the restaurant's own history - thousands of sent emails - and reverse-engineers the playbook from the best of them, not the average. A model trained on everything converges on the lazy version of someone; the drafts have to sound like her good emails. Before go-live, a fresh model instance is handed that playbook plus a real historical enquiry and scored against what the restaurant actually sent. The free trial is a demonstration, not a training period. The value is there before the owner has sent a single reply.

Why it never sends

Assist-only is not a limitation, it is a decision about where the liability sits. If an AI sends on its own and gets an allergen or a price wrong, the software is on the hook. If the owner presses send, she is, and she is the one person who should be. So the system is built to stop: the hard cases are drafted but flagged, and the verifier holds anything it does not trust rather than surfacing it. Speed versus risk, settled in favour of the human, every time.

Where it stands

Live in production at Lorne since June 2026, first paid contract in August. Multi-tenant on Railway with Postgres since July, so the next restaurant is a config change rather than a rebuild. Python end to end; the drafting model is Claude. Further pilots are in progress. The clearest thing I have learned from selling it is that the customer names the value before I do: group enquiries go to whoever answers first, and she wants to be first.

What I want to talk about on the day

Three things I would like time on with the Applied AI team.

First, approval-gate design for autonomous agents. I run a headless Claude Code loop on a Mac Mini that triages my inbound tasks, does the reversible work itself, and stops at a Telegram button for anything that touches another person. The hard part has not been capability. It has been deciding what the agent may do unsupervised, and stopping it quietly widening its own scope over time.

Second, grounding against confident wrong answers on structured data. My agent used to invent free calendar slots by eyeballing a list of events. The fix was to ban the judgement entirely and force a free-busy call before any slot could be proposed. I suspect that generalises: where else is the right move to take the reasoning away from the model rather than improve the prompt?

Third, evaluating rule-following rather than quality. NightDraft drafts against a codified playbook with party-size tiers and a sign-off ladder. Grading "is this well written" is easy. Grading "did it obey the rule it was given" is not, and that is the eval I want to get right.

The showcase slides go in within the next 48 hours. This page updates when they do.

← Back to Events