Nothing goes out unless you send it. The AI preps every lead overnight; you review each morning. · 2026-07-17
Nothing goes out unless you send it. The AI preps every lead overnight; you review each morning.
Master summary — the gist in 30 seconds
TL;DRWe flip the pipeline's default: no automatic emails, ever. Instead, every night at 3 a.m. an AI reads each pre-sale lead's full history and prepares the next move — usually a ready-to-send email draft in your voice. Each morning, purple banners on the board cards (plus a filtered morning view) show you everything; you edit lightly and hit Send. Every suggestion AND what you actually did gets logged, so the drafts keep getting more 'you'.
Input: your pipeline + all past conversations + your own past replies (Missive + Gmail, last 12 months). Output: a morning board where every actionable lead already has a suggested next step and a draft waiting.
Why this mattersYou stay 100% in control of what clients see (the AI stays invisible to them), but the tedious 'what do I say to this lead next?' thinking is done while you sleep. Triage of ~40 leads should take ~10 minutes.
flowchart LR
N["3 a.m. nightly pass"] --> R["Reads each lead's<br/>full history"]
R --> D["Drafts email or<br/>suggests next step"]
D --> B["Purple banner on<br/>the board card"]
B --> Y["You: review, edit,<br/>Send (or Skip)"]
Y --> L["Both sides logged:<br/>suggestion + your action"]
L --> N
1 · One brain instead of four
TL;DRToday three separate robots already draft follow-ups, chases and no-show emails on timers. The new nightly AI REPLACES them as the single drafter — their proven timing rules become hints it considers, not competing drafts.
Input: the old follow-up/chase/no-show timing rules + the lead's situation. Output: exactly one suggestion per lead per morning, in one consistent voice.
Why it mattersWithout this you'd wake up to duplicate, conflicting drafts on the same card — the biggest hidden trap we caught during planning.
flowchart TD
A["Old: 3 separate<br/>draft robots"] -->|become timing hints| C["ONE nightly AI drafter"]
B["Lead history +<br/>your past replies"] --> C
C --> E["One suggestion<br/>per card"]
2 · The winning design: the card IS the cockpit (B-card, 8.50/10)
TL;DRNine AI planners with different personalities designed 3 competing frontends; a judge scored them. Winner: keep everything on the board card you already read — a rich expandable purple banner (history on top, draft below, action buttons) — plus a morning view that's just the board filtered to cards with suggestions. Best ideas from the losers were grafted in: read-context-before-draft layout, a triple warning when a suggestion goes stale, a '12/41 done · ~6 min left' progress counter, and keyboard shortcuts.
Input: the existing board card. Output: the same card, now with a purple next-step banner that expands into history + editable draft + Send/Book meeting/Wait/Mark lost/Answer-the-AI buttons.
Why it mattersSmallest change to what already works = fastest to ship, least risk of breaking the dashboard — while still hitting the 10-minute-triage goal.
3 · An AI that writes like you — without fancy infrastructure
TL;DRNo Pinecone, no vector database. We tag your last 12 months of real replies (only ones YOU wrote to real prospects) by situation — first touch, price pushback, ghosting, scheduling… — and show the AI 3–8 of your best matching replies as examples each time it drafts. It must use your templates when one fits; if none fits it drafts freehand but flags it visibly, and repeated freehands become a proposed new template for you to approve.
Input: your curated past replies + the lead's situation. Output: a draft that sounds like you and uses the right template.
Why it mattersAt your scale (~57 leads, a few thousand past emails) a vector database is over-engineering; matching by SITUATION beats matching by topic for capturing your voice. Cheaper, simpler, better.
flowchart LR
M["12 months of<br/>your sent mail"] --> F["Filter: only real<br/>prospect replies"]
F --> T["Tag by situation<br/>(objection, nudge, pricing...)"]
T --> X["Pick 3-8 matching<br/>examples"]
X --> G["AI drafts in<br/>your voice"]
4 · Live and never stale
TL;DRNew client emails pop onto the cards in real time (the push channel already exists — nothing new to build). If a lead replies during the day, their 3 a.m. suggestion instantly greys out with an 'outdated — new reply received' warning so you never send yesterday's answer.
Input: a lead replies at 10:00. Output: card updates live, old draft disabled, fresh draft next night.
Why it mattersA stale suggestion is worse than none — this was a gap we caught and closed in the plan.
⏭️ Next steps
TL;DRThe plan bundle is complete. Next: open a FRESH Claude Code chat and paste the prepared prompt (plans/review-first-nightly-ai/PROMPT_to_techspec.md) — that instance turns this plan into a concrete build checklist, then a third instance builds it through the normal staging→QA→prod pipeline.
Input: the copy-paste prompt. Output: a build checklist, then the shipped feature.
Why it mattersNothing needs a decision from you right now — all 9 interview questions are answered and baked in. Three technical checks (is 3 a.m. safe for the other nightly chores · how the old robots hand over · nightly AI cost vs the $30 cap) are assigned to the next instance.
💡 Fun facts & practical stuff
TL;DR(1) The design competition cost ~1.1M tokens across 13 agents and finished in ~9 minutes — and the 'boring' smallest-change design beat both flashier ones on the judge's weighted score. (2) The biggest risk wasn't the UI at all: three existing robots already draft emails nightly, and a fourth would have meant duplicate drafts every morning. (3) Research verdict worth remembering: semantic (vector) search finds emails about the same TOPIC, not the same MOVE — for copying someone's voice, situation tags beat embeddings. (4) The 3 a.m. slot was free the whole time: no new cron needed, the existing daily sweep just moves from 8:00 to 3:00. (5) Your Approve button already had a 'feedback' field nobody was using — it becomes the learning hook for free.
Five things from this session worth keeping.
Why it mattersReusable lessons for future AI-drafting projects.