AI-Native Playbook · Bert Carroll
Doctrine

AI-Native Engineering Playbook

The operational mechanics behind the doctrine. How estimation, sessions, decisions, quality, communication, and scaling actually work.
Bert Carroll · Ask the Human LLC · v1.0 · March 2026

The AI-Native Engineering Doctrine covers what this discipline is and why it works. This playbook covers how.

Six operational systems make the doctrine real. Each one is simple on its own. Together they compound into a development methodology that lets one person operate like an engineering team.

Estimation Sessions Decisions Quality Communication Portfolio

1. Estimation

Story Points as Complexity Currency

Story points measure complexity, not time. They provide a consistent unit that translates between AI-native development speed and traditional stakeholder expectations.

Story points represent solved complexity, not hours worked.

SPCategoryScopeExample
1Trivial1 file, <50 LOCFix typo, update constant
2Small1-2 files, <150 LOCAdd form field, simple bug fix
3Medium2-4 files, <300 LOCNew UI component, API endpoint
5Large4-8 files, <600 LOCNew page, auth system, migration
8Very Large8-15 files, <1200 LOCMajor feature, architectural refactor
13Epic15+ files, >1200 LOCMust split into smaller tasks

What Story Points Measure

Complexity

How technically difficult is the problem? How many systems does it touch?

Scope

How many files, components, and integration points are affected?

Uncertainty

How well understood is the requirement? Are there unknowns?

Risk

What is the likelihood of rework? Could it break existing functionality?

Story points do not measure hours, business value, or developer skill.

Pricing Model

Clients buy value delivered, not time. Efficiency gains accrue to the builder as margin.

Project SizeRate per SPExample (8 SP feature)
1-50 SP$150$1,200
51-200 SP$125$1,000
200+ SP$100$800

Why Hourly Billing Breaks

Traditional consulting assumes human effort maps linearly to time maps linearly to output. AI-native development collapses this relationship. Two hours of architecture work can replace two weeks of traditional engineering effort.

When the relationship between time and output breaks, hourly billing punishes the builder for being efficient. Story points align incentives: the client pays for solved complexity, and speed accrues to the builder as margin.

Napkin Sessions

Structured 30-minute discovery that converts ideas to scoped plans.

Map the problem Define user flows Identify constraints Estimate SP Scope Phase 0

Phase 0 is always the smallest proof of value. Build the thing that proves the system works, then expand.


2. Session Lifecycle

How Work Gets Done

A session is a 2-4 hour focused work block with clear objectives, tracked outcomes, and documented decisions. Sessions are the atomic unit of AI-native development.

Every AI session starts cold. Documentation is the persistence layer.

Session Document Structure

Objectives

What we set out to do. Acceptance criteria defined before work starts.

Outcomes

What actually happened. Divergences from plan are documented, not hidden.

Decisions

Mini-ADRs captured inline. Why we chose X over Y. These feed future sessions.

Artifacts

Files modified, SP delivered, commits created, pages deployed.

Blockers & Next Steps

What stopped us, what to do next. Feeds directly into the next session's objectives.

The Session Workflow

graph TD A[Notes Repo
Problem Framing] --> B[Constraints
Definition] B --> C[Architecture
Decisions] C --> D[AI Dev Starter Kit
Project Bootstrap] D --> E[AI Execution
2-4 Hour Session] E --> F[Session
Documentation] F --> G[Pattern
Extraction] G --> H[Lessons Encoded
RCA / Runbooks / ADRs] H -.->|feeds next session| A

Cross-Session Persistence

Three files maintain continuity across sessions. Without them, AI is stateless and every session starts from scratch.

STATE.md Current sprint, decisions made, blockers, outcomes. Updated at end of every session.
CLAUDE.md Project context for AI agents. Stack, conventions, key paths, contacts, guardrails.
Daily Notes Cross-project journal. What happened, action items, carried forward tasks.

Morning Briefing

Automated system runs at 6 AM. Aggregates calendars, emails, project states, and uncommitted changes across all repos. Purpose: system awareness + entropy detection. Catch drift before it becomes debt.


3. Architecture Decisions

The Permanent Record

Architecture Decision Records preserve the why behind technical choices. In AI-native development, they serve double duty: they're documentation for humans and constraints for AI.

Architecture is the prompt.

AI performs best when operating inside clearly defined constraints. ADRs define those constraints. Without them, every session re-discovers decisions that were already made, debates recur, and compliance gaps surface late.

ADRs solve this by making decisions permanent and machine-readable. The AI reads the ADR before proposing changes, new sessions start with full context, and there is an audit trail from day one.

ADR Structure

Context What problem are we solving? What forces are at play?
Decision What did we choose? State it clearly and simply.
Consequences Positive, negative, and neutral outcomes. No decision is cost-free.
Alternatives What we rejected and why. Prevents relitigating.

Key rules: ADRs are permanent. Supersede, don't delete. A superseded ADR is still valuable because it records what changed and why.

Case study: Recovery Ecosystem (Orbiit) produced 42 ADRs across 174 documented sessions. These ADRs enabled a 12x productivity multiplier because every AI session operated inside well-defined architectural constraints instead of re-discovering decisions.


4. Quality Gates

Where Humans Stay in the Loop

AI handles execution. Humans handle judgment. The line between them is the quality gate. Cross it in the wrong direction and you get vibe coding. Respect it and you get velocity with integrity.

AI writes the code. Humans own the consequences.

Architecture Decisions Stack selection, design patterns, trade-offs between competing approaches Human
Compliance HIPAA, SOC 2, regulatory requirements, data handling Human
Business Logic Does the implementation solve the actual problem the user described? Human
Code Implementation Writing functions, components, tests, migrations, API endpoints AI
Code Review Security audit, performance review, correctness verification Both
Client Communication Expectations, scope, priorities, relationship calibration Human

Operational Checkpoints

CheckpointWhenWhat Happens
Pre-PR GrillBefore any pull request"Grill me on these changes. Don't PR until I pass."
Post-CorrectionAfter any mistakeUpdate CLAUDE.md so the same mistake never repeats.
High-Stakes ReviewArchitecture, security, dataSecond AI instance reviews as a staff engineer.
Session EndBefore closing a sessionUpdate STATE.md with decisions, blockers, outcomes.
Mediocre ShipAfter shipping a compromise"Knowing everything now, scrap and implement the elegant solution."

The goal is not zero mistakes. It is zero repeated mistakes. Every correction gets encoded into the system so it cannot happen again.


5. Communication Layer

Shareable Pages

The engineering knowledge lives in the repo. ADRs, session docs, patterns, commit history. But the people who need to understand what you built don't read repos.

Documentation that nobody reads is documentation that doesn't exist.

Shareable pages bridge the gap. Single-URL, self-contained HTML pages that open instantly on any device, render cleanly on mobile, and print to PDF. Hash-based URLs serve as lightweight access control.

When to Create a Shareable Page

TriggerExample
Stakeholder needs to understand a featureFeature brief for investor or partner
Partner onboardingPlatform capabilities overview
Release communicationWhat shipped, what changed, what's next
Decision affects non-engineersADR summary for business stakeholders
Sales materialOne-pager, proposal, demo walkthrough
Post-incidentSimplified RCA for affected parties
Thought leadershipArticle or whitepaper for distribution

When NOT to Create One

Rule of thumb: If the audience is an AI agent or engineer on your team, it stays in the repo (ADR, runbook, session doc). If the audience is anyone else, consider a shareable page.

SDLC to Communication Mapping

Repo ArtifactBecomes
ADRDecision summary for stakeholders
Session docsSprint/release summary for leadership
RCAIncident summary for affected parties
ROADMAP.mdFeature brief for partners/investors
PatternsBest practices article for community

Design Principles

The multiplier: A partner reads a capabilities page before a call. They ask informed questions. The meeting is productive instead of introductory. 30 minutes creating a page saves hours of meetings, emails, and misunderstandings.


6. Portfolio Management

Scaling Across Projects

The methodology above works for one project. Scaling to 5-7 concurrent projects requires consistent tooling and automated awareness.

Same methodology. Different repos. Consistent tooling.

Project Bootstrap

Every new project starts from the AI Development Starter Kit, which provides standardized scaffolding.

CLAUDE.md

Project context, stack, conventions, key paths, guardrails. AI reads this at session start.

ADR Templates

9 starter decisions covering auth, database, hosting, CI/CD, testing, and more.

Session Docs

Template with objectives, outcomes, decisions, artifacts. Consistent format across all projects.

STATE.md

Living document: current sprint, recent decisions, blockers, and context that AI needs.

Portfolio Awareness

Command Center Parallel scan of all repos, 3 email accounts, 3 calendars. Surfaces blockers and uncommitted work.
Signal Dashboard Real-time project tracking. Backlog, velocity, burn-down. Client-accessible progress.
Morning Briefing Automated 6 AM report. Calendar + email + repo state + entropy detection.
Daily Notes Cross-project journal. What happened everywhere today, carried forward items.

The Fractional CTO Pattern

One person, multiple companies, same methodology. Each project gets its own repo with the full starter kit. The portfolio layer (command center, daily notes, briefings) sits across all of them in a central notes repo.

graph TD N[Notes Repo
Portfolio Layer] --> P1[Project A
CLAUDE.md + STATE.md + ADRs] N --> P2[Project B
CLAUDE.md + STATE.md + ADRs] N --> P3[Project C
CLAUDE.md + STATE.md + ADRs] N --> P4[Project D
CLAUDE.md + STATE.md + ADRs] N --> CC[Command Center
Parallel Scan] CC --> P1 CC --> P2 CC --> P3 CC --> P4

The methodology is replicable. The IP is in the system, not in any individual's head. New projects inherit the full discipline from day one through the starter kit.

Proven results: Recovery Ecosystem achieved 160,000+ lines of code, 1,315+ commits, 174 documented sessions, 42 ADRs, and a 12x productivity multiplier vs. traditional development. All from one architect and AI execution.


Six systems. Each one simple. Together they compound into a methodology that lets one person with AI operate like an engineering team.

The Doctrine explains the thesis. This playbook explains the mechanics. Neither is complete without the other.

Build the system. Then let the system build.