Articles

Measured findings on context engineering, loop engineering, and the instruction files that steer AI coding agents: CLAUDE.md, AGENTS.md, system prompts, and the checks around them.

A CLAUDE.md open to two rules under two different headings, one near the top reading scratchpad ships without tests and one far below reading write a test for every function, a line drawn between them showing they collide, and the lower rule marked as the one the model keeps.

Opus 5: Cost of Instruction Conflicts

Two rules in your instruction file disagree. The model picks one, follows it to the letter, and moves on without a word. The other goes dead, and where each rule sits in the file decides which one you lost.

Gábor Mészáros · August 5, 2026 · 8 min read

Read the article →

A 300-line CLAUDE.md with most of its lines greyed out under a DELETE stamp, and a small handful of lines left highlighted, each tagged with a property you can check without running the model: names a construct, does not conflict, loads where it applies.

Opus 5: Delete your CLAUDE.md?

Boris Cherny is right that most of your instruction file is dead weight. But deleting it treats a quality problem as a quantity problem, and cannot tell you which rules were load-bearing.

Gábor Mészáros · August 4, 2026 · 9 min read

Read the article →

A terminal shows the code the agent actually wrote, where charge(10000) returns 10000 at full price with the discount never added, while a success dialog beside it reports 'Task failed successfully' with a green check and the agent's claim that all tests pass.

Loop Engineering: Stop Failed Successfully

In a real agent loop the model ends the run by declaring its own success, and a stop that believes the report ships the bug behind it

Gábor Mészáros · July 29, 2026 · 12 min read

Read the article →

An agent loop: the model generates, the check returns red, the steer arm rewrites the next instruction to 'make the test pass' and hands it back, and the model turns green by editing the test's assertion to match the bug; the steer arm is the one lit up, carrying the wrong target back to the top of the loop.

Loop Engineering: How to Stop Your Agent Reward-Hacking Its Own Checks

The steer is the one instruction in the loop you never write down, and it is where an agent learns to game its own checks

Gábor Mészáros · July 22, 2026 · 12 min read

Read the article →

The agent loop drawn as a cycle: the context surface feeds the model, the model generates a reply, block-placation.sh checks it, a placated reply routes through steer and retry back to the model, a clean reply reaches continue; the context surface node is marked as the part this piece takes apart.

Loop Engineering: How to Stop the "You're Absolutely Right" Sycophancy

What an instruction costs to keep loaded, and how to load each rule only where it applies

Gábor Mészáros · July 14, 2026 · 9 min read

Read the article →

fixing check

Loop Engineering: Fine-Tuning the Guardrail That Fired Wrong

How to tell a guardrail that found nothing from one that quietly stopped looking

Gábor Mészáros · July 14, 2026 · 7 min read

Read the article →

Two channels feeding the same edit: a speech bubble labeled 'asks' feeding a prompt, and a stop sign labeled 'refuses' at a gate, with only the gate able to block the write.

Deterministic Guardrails: Prompts Steer, Hooks Enforce

A Prompt Asks; a Gate Refuses

Gábor Mészáros · July 10, 2026 · 10 min read

Read the article →

Two identical green checkmarks joined by an equals sign, one labeled safe and the other blind with a broken outline, under the line 'fix the instrument, not the reading'.

Prompt Engineering, Context Engineering, Loop Engineering: What Actually Changed

Three Renames, One Question: What Is the Unit You Are Actually Engineering?

Gábor Mészáros · July 7, 2026 · 8 min read

Read the article →

Black-and-white photograph of a brutalist concrete colonnade — vertical columns crossed at right angles by a horizontal beam and railing, two perpendicular axes meeting in a dark structural void.

Green Tests Don't Mean Better Software

Spec-Driven vs Hypothesis-Driven Is a False Choice — Build a Governance Selector Instead

Gábor Mészáros · June 30, 2026 · 9 min read

Read the article →

A black-and-white photograph of a brutalist concrete building shot upward and mirrored into a symmetric composition, rows of towers rising toward a bright sky at the center.

Instruction systems capability ladder: harness leveling

A maturity ladder for AI instruction setups, redrawn around the channel each rung runs on.

Gábor Mészáros · May 19, 2026 · 7 min read

Read the article →

The State of AI Instruction Quality (30k repo analysis)

The State of AI Instruction Quality (30k repo analysis)

28,721 repositories. 165,063 instruction files. 3.3 million instructions. What the corpus shows.

Gábor Mészáros · April 21, 2026 · 9 min read

Read the article →

The Undiagnosed Input Problem

The Undiagnosed Input Problem

The ecosystem inspects outputs. The input side has no equivalent diagnostic layer.

Gábor Mészáros · April 8, 2026 · 8 min read

Read the article →