The brain Wolfberg runs on
CAPSTONE.
The brain Wolfberg runs on.
Versioned. Reviewable. Mergeable. Persisted. The way mature engineering teams treat code, applied to the context AI sessions run on.
What it is
Capstone is the strategy and synthesis instance — the brain Wolfberg runs on. Not a product. Not for sale.
The methodology that makes Capstone work is for sale, via the Curriculum. What's on this page: how it actually works, who runs alongside it, and the receipts that prove the discipline compounds.
The mechanics
Context as Code.
Most people work around AI's context loss by saving a prompt for next time, which doesn't carry the actual context. We treated context the way mature engineering teams treat code: versioned, reviewable, mergeable, persisted.
Memory files
Typed and indexed. Decisions and why. Corrections never given twice.
Session deltas
Every session ends with a structured handoff. Every session starts by reading the last three. Stop and Start hooks automate it.
Cross-instance state
Three Claude instances reading from one context bus. None work in isolation.
A canonical task board
So any instance, any session, knows exactly what's in flight.
Knowledge compounds instead of evaporating. Every session picks up where the last one left off — and so does every new operator. That's why this is an operating model and not a personal practice.
The team
Named components. Public architecture.
Three Claude instances, one human operator. None work in isolation. The architecture is public; the brain contents are private.
The human operator
Berg
Sets direction. Makes calls. Drives the work.
Strategy + synthesis
Capstone
Reads canonical state. Writes positioning, briefs, decisions. Talks to Berg, Code, and the brain.
Engineering execution
Code
Writes software. Touches files, git, AWS. Talks to Berg, Capstone, and the codebase.
Visual identity execution
Design
Less mature than the others; same protocol.
Canonical state
The shared brain
A Notion workspace + a markdown filesystem. The canonical state every instance reads from and writes to.
Continuity discipline
Session protocols
How a session starts, how it ends, how it hands off. The reason the model compounds.
The receipts
The model compounds. Here's the proof.
Live from the operator brain. Updated nightly. The shape is public; what's inside any of these surfaces is not.
~165
active hours
Operator work captured across 38 calendar days since 2026-04-17. The model compounds every session, not just the ones you remember.
87,522
lines of code
Engineering output. Three canonical repos. Built solo, alongside the AI instances.
1395
pages in the brain
Operator runbooks, memory files, session deltas, canonical pages. The substrate every instance reads from and writes to.
3
Claude instances
Capstone, Code, Design. None work in isolation. All read from one shared context bus.
From slice 1 of the shopizer pilot
The discipline on real engineering work.
$3.46
API spend
Migrator + Verifier end-to-end on 1,003 LOC of Java reference-data refactor. Anthropic Claude Opus 4.7.
2.9
hrs operator oversight
Build the harness, fire the run, author the parity report. Operator hours keep the loop honest; not the cost driver.
8
findings surfaced
1 blocker, 4 major, 3 minor. Each with severity, sub-classification, and suggested customer disposition.
2
Migrator-missed defects
Caught by the independent Verifier — a `@UniqueConstraint` lost in target, an audit-event create/update conflation. That's why the Verifier is independent.
The blocker is the receipt.
The Migrator emitted a Lambda surface with no auth wrapper because the Architect's mapping cited a primitive the configured Phase-1 dictionary lacked. The Verifier escalated it from a soft-pedaled gap to a structural blocker. The pipeline catches what the implementer punts on. Public artifact of the pilot — Migrator output, Verifier parity report, all eight findings with reproduction notes — publishes at launch.
Public artifact: github.com/atkinsonb2/shopizer-refactory-pilot Releases May 25, 2026
From the operator rebuild on 2026-05-19
The model caught its own near-miss.
0
data-loss events
37 business documents the destroy script's allowlist would have destroyed. Halt-and-surface caught the run mid-flight; recovery via delete-marker removal and archive-and-repopulate.
3m 36s
AWS API time
Active compute across destroy plus apply for the wolfberg operator. CloudTrail confirmed 98 plus 223 events across two bursts.
158
resources provisioned
keystone-wolfberg-prod-* end-to-end. Same code with operator_id=acme-hvac produces keystone-acme-hvac-prod-*. Multi-tenant by construction.
0
lines of remediation code
Berg authored. Ten minutes commanding plus twenty watching apply runs. The model handled the recovery; the operator gated the calls.
The differentiator isn't fast Terraform deployment. It's the operating model that catches its own near-misses and produces an audit trail dense enough to write the case study from primary sources after the fact.
The model is the asset. Keystone and Refactory are products of running it. Senior Advisory is the substrate the model grew from. Curriculum is this — the model itself — packaged for transfer.
When the brain points at engineering work, this is what comes out — Refactory and Keystone.
What's for sale
The methodology.
Not the instance.
Capstone-the-instance isn't for sale. The methodology that makes it work is. Day 4 of the Wolfberg Curriculum is Brain Process: how to build this.
Five days. Up to four seats. Same curriculum we use to onboard our own people — because the only honest way to teach an operating model is to be running on it.
The Deck
Get the formal deck.
The brain Wolfberg runs on. Operating model, mechanics, named components, and the receipts that prove the discipline compounds. Wolfberg LLC Proprietary & Confidential.
View Capstone deck (PDF)For operators
Want to run your work this way?
Tell us what you're trying to build. We'll tell you what the operating model looks like for your shape.
Start the conversation