00 / OPEN

AI Engineering
for Programmers

From Vibe Coding to Systematic Engineering.

ATL TECH· 2026· SADIG MURADOV

00 / OPEN

Who's talking?

Background

Sadig Muradov — Engineering Team Lead
15+ years across numerous industries: Healthcare · E-commerce · Payments · Logistics · Telecom · Defence · Fleet
www.sadig.dev

Credential

CLAUDE · ANTHROPIC

Claude Partner Network — Member

Recognized by Anthropic for production AI agent engineering practice.

01 / PART 1 — FOUNDATIONS

The Paradigm Shift

80/20 Flip — typing vs orchestration.

·REALITY

AI turns software development from manual code generation into agent orchestration.

·IMPLICATION

You are no longer a typist. You manage a fast junior with a limited context window.

·GOAL

Push effort upstream. If the spec is perfect, the code becomes trivial.

01 / PART 1 — FOUNDATIONS

Big Word Alert.

Quick glossary before we go deeper.

TOKEN

A sub-word chunk — roughly 4 English characters or one short word. Models read and write in tokens, not characters. Tokenizers (BPE in Claude / GPT) split encyclopedia into 3 tokens; hello stays 1. Token count is the bill.

INFERENCE

The phase where the model runs — reading your prompt and generating output, one token at a time. Distinct from training, where the weights are learned. Every API call is one inference. Inference is fast and cheap; training is slow and expensive.

CONTEXT

Everything the model sees at inference: your prompt, attached files, the conversation so far. Distinct from training data. Hard ceiling: the context window. Claude Opus 4.7: 1M tokens. Claude Sonnet 4.6: 200K. Frontier GPT: ~1M.

PARAMETERS · B

The "knobs" the network learned during training. 70B = 70 billion knobs. Bigger ≠ smarter — data quality and post-training (RLHF, Constitutional AI) matter more. Frontier models today: hundreds of billions to low trillions.

METHODOLOGY

Today's LLMs are transformers doing autoregressive next-token prediction, then post-trained with RLHF / Constitutional AI. Strong at: language, code, pattern recall. Weak at: precise math, long-horizon planning, novel reasoning.

HALLUCINATION

Confident-sounding output that's factually wrong — a non-existent API, a misremembered date, an invented citation. Not a bug; a property of next-token prediction. The fix isn't "smarter model" — it's verification: tests, MCP calls, docs lookups, your judgment.

01 / PART 1 — FOUNDATIONS

Context — finite resource

Managing the context window is the AI Engineer's job #1.

01 / PART 1 — FOUNDATIONS

MCP — Model Context Protocol.

The standard that bridges the silos.

CLI

claude mcp add — register a server
/mcp — manage servers in-session
.mcp.json at repo root — team scope, version-controlled

CONTEXT COST

MCP tool defs live in context, even unused.
Many servers → less room for your conversation.

DEFER LOADING · OFFICIAL

ENABLE_TOOL_SEARCH=auto
load only what fits 10% of context, defer the rest.
↗ code.claude.com/docs/en/mcp

01 / PART 1 — FOUNDATIONS

The Contractor
Mental Model.

Seven primitives, one analogy. Borrow this for your team.

The setup

You just hired a brilliant contractor. They're fast, capable, patient — but they need to know where they are, what to do, and what's off-limits.

CLAUDE.md + skills — what Claude knows and can do.

MCP — what it can reach.

Hooks — rules that enforce themselves.

Sub-agents & teams — workers you don't babysit.

Plugins — how you share a whole setup.

Primitive → Metaphor

CLAUDE.md
Briefing package. How the contractor works on this job.
SUB-AGENTS
Specialists. Isolated workers with their own fresh context window.
SKILLS
Trade manuals. Documented once, never re-explained.
MCP
Different suppliers. Gmail, Slack, your DB — outside resources brought to the job site.
HOOKS
House rules. Deterministic shell scripts that fire on events. If something needs to happen every time without fail, don't put it in a prompt — put it in a hook.
AGENT TEAMS
Solo → full crew. A house lead coordinates; the rest are specialists.
PLUGINS
Packaged setup. Bundle skills + hooks + MCP, hand the team an instant onboarding.

01 / PART 1 — FOUNDATIONS

LSP — beautiful theory,
reality differs.

·THE THEORY

IDE → LSP → structured context → AI
AI sees the same structural data the compiler sees
"Generation grounded in actual definitions"

·THE REALITY

Claude Code and most AI agents don't use LSP effectively
Request/response model doesn't fit the agent loop
Small projects: no difference — AI just reads files
Large projects: AI loses cross-file dependency, transitive impact, real call graph

· Don't rely on LSP for brownfield. Better tool — next slide.

01 / PART 1 — FOUNDATIONS

GitNexus — graph of your codebase.

The LSP replacement for big codebases. Queried over MCP.

02 / PART 2 — COMPOUND ENGINEERING

Anti-Pattern — Vibe Coding.

Why randomness doesn't scale.

Definition

Coding by intuition — random prompting, immediate gratification, no structure.

Why it fails

Entropy — context lost between sessions
Repetition — same mistakes, no learning loop
Ceiling — collapses on complex architecture

If you aren't systematic, you're just gambling with tokens.

Productivity vs Complexity

02 / PART 2 — COMPOUND ENGINEERING

Compound Engineering.

Compound interest for code.

The flip

TRADITIONAL — every task makes the next one harder.

COMPOUND — every task makes the next one easier.

The mechanism

02 / PART 2 — COMPOUND ENGINEERING

The Engineering Loop.

Standard operating procedure.

02 / PART 2 — COMPOUND ENGINEERING

Plan & Work.

Measure twice, generate once.

·PHASE 1 — PLAN · 40%

Action. Research, codebase analysis, detailed Markdown spec.

Tool. Use Plan Mode — press Shift + Tab until you see "Plan Mode" under the input. Claude can read but not edit.

"If the spec is right, the code is inevitable."

·PHASE 2 — WORK · 10%

Action. Approve the plan. AI executes against it.

Reality. The "easy" part. The AI acts as a junior developer following a rigid blueprint.

Don't start Work until Plan is validated.

02 / PART 2 — COMPOUND ENGINEERING

Review & Compound.

Closing the learning gap.

·PHASE 3 — REVIEW · 40%

Action. Verify output, check against constraints, catch hallucinations.

Role. You are the senior engineer reviewing a junior's pull request.

Method. Automated tests + manual audit.

·PHASE 4 — COMPOUND · 10%

Action. Document the learnings for the future.

Mechanism. Update CLAUDE.md or system instructions.

Goal. Never make the same mistake twice.

# CLAUDE.md update
- Rule: always check ownership on DELETE
- Context: planted-bug PR caught in review

02 / PART 2 — COMPOUND ENGINEERING

Live Demo.

Notes API on the Loop — full cycle in ~9 min.

LIVE · NOTES API ON THE LOOP

Run the cycle. End to end.

01 / PLAN

Read CLAUDE.md + ai.docs · audit · write spec

~3 min

02 / WORK

Apply the spec · minimal diff

~1 min

03 / REVIEW

uv run pytest -v · bug FAIL → PASS

~3 min

04 / COMPOUND

CLAUDE.md update · lock the lesson

~1 min

Demo recording Shown live during the session — not bundled with this hosted deck.

backup recording

03 / PART 3 — PRACTICAL PATTERNS

Multi-Agent Orchestration.

Specialized agents in parallel — not one serial conversation.

·RESEARCHER

Scans docs, finds patterns, surfaces prior art.

·IMPLEMENTER

Writes code based on the researcher's findings.

·CRITIC

Audits against security, style, and project conventions.

· Strategy — swarms for complex architecture, sequential for simple fixes.

03 / PART 3 — PRACTICAL PATTERNS

Setup & Environment.

The engineer's stack.

Tooling

Editor — Cursor · VS Code (Copilot) · Claude Code ←
Model — Claude 4.6 Sonnet / 4.7 Opus
Interface — MCP servers (filesystem · git · gitnexus · sentry)

CLAUDE.md hierarchy

Project — repo root, committed, team-shared
User — ~/.claude/CLAUDE.md, your personal preferences across all projects
Reference docs inline with @README.md
Bootstrap with /init — let Claude generate the first draft

CLAUDE.md is source code.
Version-control it.

PROJECT ROOT

/
├── src/
├── .cursorrules
├── CLAUDE.md            ← THE BRAIN
├── ai.docs/
│   ├── business-context.md
│   ├── user-flows.md
│   └── permissions.md
└── .claude/
    ├── commands/
    └── skills/

03 / PART 3 — PRACTICAL PATTERNS

Handling Failure.

Mitigating hallucinations and risks.

Reality

AI is probabilistic, not deterministic. It will hallucinate.

If the AI says "Good catch!" or "That's a great question!" — context is corrupted. Drop the branch. Start fresh.

Mitigation

RECOGNIZE — plausible but non-existent API? verify via MCP / docs
DEBUG — ask for the reasoning before the code
OVERRIDE — know when to stop prompting and write it by hand
SECURITY — never give secrets / credentials to generated code
RESET — "good catch" signal → discard branch, restart context

05 / CLOSE

The Engineer's Manifesto.

Five rules to engrave.

01

CONTEXT IS KING

High-signal context beats the "perfect" prompt every time.

02

PLAN FIRST

40/10/40/10 loop. Value is created before and after coding, not during.

03

COMPOUND KNOWLEDGE

If it isn't written into the system config, you haven't learned it.

04

ORCHESTRATE

Think in workflows and agents — not chat messages.

05

GROUND REALITY

MCP + GitNexus connect the AI to real tools and your real codebase.

05 / CLOSE

Stay in touch.

Questions, follow-ups, or just to say hello.

SPEAKER

Sadig Muradov

Engineering Team Lead · Claude Partner Network Member

LINKEDINlinkedin.com/in/sadigm
GITHUBgithub.com/sadigm
EMAILmail@sadig.dev

Slide deck & example project will be shared after the session at linkedin.com/in/sadigm.

SCAN · LINKEDIN.COM/IN/SADIGM

AI Engineeringfor Programmers

Who's talking?

Background

Credential

The Paradigm Shift

Big Word Alert.

Context — finite resource

MCP — Model Context Protocol.

The ContractorMental Model.

The setup

Primitive → Metaphor

LSP — beautiful theory,reality differs.

GitNexus — graph of your codebase.

Anti-Pattern — Vibe Coding.

Definition

Why it fails

Productivity vs Complexity

Compound Engineering.

The flip

The mechanism

The Engineering Loop.

Plan & Work.

Review & Compound.

Live Demo.

Multi-Agent Orchestration.

Setup & Environment.

Tooling

CLAUDE.md hierarchy

Handling Failure.

Reality

Mitigation

The Engineer's Manifesto.

Stay in touch.

AI Engineering
for Programmers

The Contractor
Mental Model.

LSP — beautiful theory,
reality differs.