← The Boswell Engine B2B outreach workflow · back to the site
Press play
not contacted 0 finished 0

The Boswell Engine's Architecture

- this blueprint demonstrates how the engine selects the most probable lead from a list · does a deep background · analyses how they think and speak · decides what to say and how · writes all the copy · checks its own work · and gets smarter the more you use it.
engine architecture
rev v0.5 / 2026-07-20
01No single LLM can do this.Five AI models, three engineering tools. Each one checking and improving the others.
02Most names never get contacted.Nobody else sells that.
who runs each stage
Human
Python
Exa
Postgres
Claude Opus 4.8
Claude Fable 5
Claude Haiku 4.5
Gemini 3.1 Pro
GPT-5.4

Anthropic reads, Google reads independently, OpenAI stress-tests and judges. Python is the plumbing. Exa is meaning-based search and the reason stage 3 works, and almost nobody outside the field has heard of it.

the rules of the engineer
L1Never let a model check its own work. It can't see its own tells.
L2Compress nothing until the last step. Each stage learns and builds on the complete work of the stage before.
L3Find out who wrote it before you use it. A chunk of what carries a person's name was written for them, not by them.
L4Every data point we use gets double-checked and verified. No source, no pass.
L5Nothing is binary. Three states, not two. Pass, kill, or we can't tell. The gray zone can go to a human.
L6If a machine can't check it, it isn't a test. And, it's not happening.
L7Every AI critique states a consequence. No defined outcome, no impact.
L8We never delete. We correct and file the original. You can't fix a mistake you didn't record.
stage 0 / once per client
The offer0.1
What they sell, and every claim we're allowed to make
Human
The rubric0.2
Who it's for, and who it's never for
Human
Client's voice0.3
How the client writes, and what they'd never say
Claude Opus 4.8
per person / from here downthe owner changes hands every stage
The list01
Names in, plus the relationship and function tags
Python
test · N rows in, N rows out. No silent drops.
First cut02
Pass, kill, or gray. No score.
Python + Claude Haiku 4.5 + Exa
test · right human and right place, before a dollar is spent.
not contacted
most of the list
Collect all public content03
Everything they've said in public
Exa deep search
test · collection only. No reading yet.
Sufficient data?3.1
Variety, not volume
Who wrote it?3.2
Their hand, or a ghostwriter's
Linguistic model04
How they say it
Claude Opus 4.8
test · absence detection: what they never say.
04 and 05 run blind to each other
World view model05
What they mean when they say it
Gemini 3.1 Pro
test · no claim without a source.
Cognitive model06
How they think about it
Claude Fable 5
test · confidence from independent sources.
Fact Checker07
A rival LLM verifies all claims
GPT-5.4
test · every correction links to a source.
what it catches
unverified items · projections · quotability bias · smoothed contradiction · cascading errors
All corrections are marked and preserved.
Go/no-go list08
If the data says no, they go
Claude Opus 4.8
the real exit
with reasons
The do-not-chase register a product, not a byproduct
Below the cut from 020
Carries the problem, cannot act from 080
illustrative · of a 1,000-name list, about 30 finish. more exits than finishes, by design.
The tactical assessment09
What to say, and why it lands
Claude Opus 4.8
test · the lineup: pick the target from five decoys.
Hyper-targeted content10
Messaging built from their cognitive profile
Claude Opus 4.8, new instance
test · bracket integrity. character limits respected.
Email
Written to their emotional frame
LinkedIn message
Survives the feed
Connection note
Built to be noticed
Ads
4 versions · tightens the noose
Brochure
For their cognitive profile, in their register
Editorial and copy analysis11
Two tests: i. usage violations, ii. poor writing
Python + Gemini 3.1 Pro
test · no style errors, then a reading for where the prose goes dead.
The record of it all12
A CRM and history of each person's progression
Python + Postgres
test · the flip rate, logged forever.
learning loop · feeds back to setup. the overrides are the training data.
project
The Boswell Engine
engineer
Atma
drawing
Engine Architecture
rev
v0.5
date
2026-07-20
for
internal use
Unpacking...