Claude Code is 100% written by Claude Code — confirmed March 7, 2026 by Boris Cherny, Claude Code lead. StrongDM ships production software with zero humans writing or reviewing code.
THE REST
A 2025 METR randomised control trial found experienced developers using AI tools completed tasks 19% SLOWER. They believed AI made them 24% faster. Wrong on direction AND magnitude.
KEVIN RYAN
03
The Framework — Section 02
NOT ALL AI CODING IS EQUAL.
Dan Shapiro's Five Levels of Vibe Coding maps where teams actually operate — vs where they think they are.
← 90% of developers are here
KEVIN RYAN
04
The Framework — Levels 0, 1 & 2
WHERE MOST DEVELOPERS LIVE
00
SPICY AUTOCOMPLETE
AI suggests the next line. You accept or reject. GitHub Copilot in its original format — a faster tab key. The human writes the software.
WHO: GITHUB COPILOT ORIGINAL
01
CODING INTERN
Hand the AI a discrete, well-scoped task — write this function, refactor this module. You review everything that comes back. AI handles tasks, human handles architecture.
WHO: MOST CASUAL AI USERS
02
JUNIOR DEVELOPER
AI handles multi-file changes, navigates the codebase, builds features spanning modules. You still read all the code. Shapiro estimates 90% of 'AI native' developers operate here.
WHO: 90% OF SELF-DESCRIBED AI-NATIVE DEVS
KEVIN RYAN
05
The Framework — Level 3
03
DEVELOPER AS MANAGER
The relationship flips.
You're no longer writing code with AI help. You're directing the AI and reviewing what it produces — at the feature level, the PR level. The model submits PRs for your review. You decide: approve, reject, redirect.
Most developers top out here. The ceiling isn't technical. It's psychological: letting go of the code.
MOST
developers hit the ceiling right here.
The barrier is psychological, not technical.
KEVIN RYAN
06
The Framework — Levels 4 & 5
BEYOND THE CEILING.
04 DEVELOPER AS PRODUCT OWNER
Write a spec, leave, come back hours later and check whether the tests pass. You're not reading the code anymore — you're evaluating outcomes. The code is a black box.
Requires spec quality almost nobody has developed yet.
05 THE DARK FACTORY
A black box that turns specs into software. No human writes the code. No human reviews it. The factory runs autonomously. Specification goes in — working, tested, shipped software comes out.
Almost nobody on the planet operates here. Yet.
KEVIN RYAN
07
Proof of Concept — Section 03
STRONGDM: THE DARK FACTORY IN PRODUCTION
"The most ambitious form of AI-assisted software development I have seen yet." — Simon Willis
HOW IT WORKS
Write markdown spec
Agent reads spec
Agent writes & tests code
Scenarios validate — agent never sees them
Software ships
3
Engineers. No one writes code. No one reviews code.
JUL '24
Running since July 2024. Inflection point: Claude 3.5 Sonnet, fall 2024.
$1k
Per engineer per day in AI compute spend. That's the benchmark.
KEVIN RYAN
08
Proof of Concept — The Architecture
TWO IDEAS THAT CHANGE EVERYTHING
EXTERNAL SCENARIOS (NOT TESTS)
Traditional tests live inside the codebase — the AI can see and game them. Teaching to the test.
Scenarios live outside the codebase. The agent builds the software without ever seeing the evaluation criteria. It can't cheat. Same principle as ML holdout sets.
DIGITAL TWIN UNIVERSE
Behavioural clones of every external service: simulated Okta, Jira, Slack, Google Docs, Drive, Sheets.
Agents develop against digital twins — full integration testing without ever touching real production systems, real APIs, or real data.
KEVIN RYAN
09
Proof of Concept — The Hyperscalers
THE TOOLS ARE BUILDING THEMSELVES.
Boris Cherny, Claude Code lead at Anthropic, confirmed on X (March 7, 2026): "Can confirm Claude Code is 100% written by Claude Code." His role has shifted entirely to specification, direction, and judgment. He hasn't personally written code in months.
100%
Claude Code written by Claude Code. Confirmed.
25%
Speed improvement building Codex 5.3 — by Codex itself.
4%
Of all GitHub public commits directly authored by Claude Code.
$1B
Claude Code annual run rate. Six months since launch.
KEVIN RYAN
10
Why Most Orgs Are Stuck — Section 04
THE J-CURVE
Productivity dips before it rises. Most organisations are sitting in the trough — and blaming the tool.
WHY THE DIP HAPPENS
Evaluating AI suggestions takes time Correcting "almost right" code is costly Context-switching between mental models Debugging subtle errors that look correct 39% of devs have low trust in AI-generated code — DORA 2024, 39,000+ professionals
KEVIN RYAN
11
Why Most Orgs Are Stuck — The Tool Trap
"COPILOT MAKES WRITING CODE CHEAPER BUT OWNING IT MORE EXPENSIVE."
— Senior engineer. Common sentiment across the industry.
The organisations seeing 25–30%+ gains didn't just install Copilot and call it done. They redesigned their entire development workflow around AI capabilities.
20M
GitHub Copilot users 42% market share
55%
Faster code completion in isolated lab tasks
↑PRs
Larger PRs, higher review costs in production
41%
Higher churn in AI-generated code (GitClear 2024)
KEVIN RYAN
12
The Reckoning — Section 05
YOUR ORG WAS DESIGNED FOR A WORLD WHERE HUMANS WRITE CODE.
CEREMONY
WHY IT EXISTED
STATUS
Daily Standup
Developers need to synchronise
FRICTION
Sprint Planning
Humans can only hold so many tasks in memory
FRICTION
Code Review
Humans make mistakes other humans catch
FRICTION
QA Team
Builders can't evaluate objectively
EVOLVING
Write Spec / Eval
Only thing left that actually matters
CORE
KEVIN RYAN
13
The Reckoning — Talent
THE APPRENTICESHIP MODEL IS BREAKING.
67%
Decline in US junior developer job postings
46%
Fall in UK graduate tech roles in 2024
53%
Further UK drop projected by end 2026
9–10%
Junior employment drop per 6 quarters of AI adoption (Harvard, 2025)
THE NEW BAR FOR JUNIOR ENGINEERS
The junior of 2026 needs the systems design understanding expected of a mid-level engineer in 2020 — because the entry-level work got automated. The bar is rising toward exactly the skills that have always been hardest to develop.
KEVIN RYAN
14
The Reckoning — New Org Economics
THE AI-NATIVE ORG IS ALREADY HERE.
REVENUE PER EMPLOYEE
Cursor
$3.3M
Midjourney
$3.0M
Lovable
$2.2M
Avg SaaS
$610K
5–6× the average SaaS company
WHAT CHANGED
No traditional eng / product / QA / DevOps teams
Small group — exceptional at understanding user needs
Exceptional at translating needs to clear spec
Directing AI systems that handle implementation
Org chart is radically flat
Middle management layer deleted
KEVIN RYAN
15
The Path — Section 06
THE BROWNFIELD PROBLEM
You cannot dark factory your way through a legacy system. The system IS the specification — and it was never written down.
01
Start at Level 2–3
Use AI to accelerate what developers already do — new features, bug fixes, refactoring. Expect the J-curve dip.
02
Document your system
Use AI to generate specs from existing code. Build scenario suites capturing real behaviour. Create the holdout sets a future factory will need.
03
Redesign CI/CD
Handle AI-generated code at volume. Different testing strategies, review processes, deployment gates.
04
Shift new development to L4–5
Autonomous agent patterns for greenfield work while maintaining legacy in parallel.
KEVIN RYAN
16
The Path — What To Do
THE BOTTLENECK HAS MOVED.
From implementation speed → to specification quality.
Spec quality
The ability to describe what needs to exist precisely enough that machines can build it — without the human filling gaps.
Domain depth
Understanding your system, your customer, and your problem deeply. The dark factory doesn't reduce this demand — it makes it an absolute law.
Judgment over coordination
Shift roles from managing who is rowing to defining where the boat should go with enough precision for machines to navigate.
Invest in your people
Engineers must upskill monthly, not annually. The person still on a January 2026 model needs a February model. This loop won't slow down.
KEVIN RYAN
17
The Opportunity
THE TOTAL ADDRESSABLE MARKET FOR SOFTWARE IS EXPLODING.
Every time the cost of computing has dropped, total software produced didn't stay flat — it exploded. Cloud didn't make existing software cheaper; it created SaaS, mobile, streaming, real-time analytics.
We are dropping the cost of software production by an order of magnitude. The unmet need is becoming addressable — not theoretically, now.
Regional hospital
Custom patient portal: was $300k+. Now addressable.
Mid-market manufacturer
Custom inventory system: was $500k+. Now viable.
Family logistics co.
Bespoke ops tools: were unaffordable. Now buildable.
KEVIN RYAN
18
THE DARK FACTORY DOES NOT REPLACE GREAT ENGINEERS.
IT AMPLIFIES THEM.
The constraint moves from can we build it — to should we build it. Should we build it has always been the harder and more interesting question.
Kevin Ryan & Associates
kevinryan.io
sddbook.com
KEVIN RYAN
19
What Next
THE BOTTLENECK IS SPEC QUALITY AND EXECUTION.
THAT'S WHERE WE WORK.
01 — TALK
START HERE. ONE CONVERSATION.
If you're a CTO, engineering lead, or platform owner — let's talk. No pitch. We figure out together whether what we do is relevant to where you are.
kevinryan.io →
02 — ASSESS
WHERE DOES YOUR ORG SIT?
A structured diagnostic — where your team sits on the five levels, what's blocking the transition, and what the migration path looks like for your specific stack, team, and codebase.
kevinryan.io →
03 — EMBED
WE COME IN. WE MAKE IT WORK.
We sit inside your team and make the transition happen. Not a report. Not a deck. Working pipelines, working practice, shipped. Field engineering for the AI-native transition.