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.

Field Notes

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

A deterministic check that comes back quiet means one of two things: the signal is absent, or the instrument went blind. How to tell them apart.

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.

Field Notes

Deterministic Guardrails: Prompts Steer, Hooks Enforce

A Prompt Asks; a Gate Refuses

Gábor Mészáros · July 10, 2026 · 10 min read

Most guardrails beside a generate-check loop only watch what an agent already wrote. The one worth building is a deterministic gate that refuses a forbidden edit before it lands.

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'.

Field Notes

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

Prompt, context, loop engineering: three renames of the same corner of the job. At each step the unit you actually engineer changed — here is what shifted.

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.

Field Notes

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

Green CI tells you a change conforms to its spec; measuring whether it improved the system is a separate job. On that gap and a per-change governance selector.

Read the article →

Research

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

Everybody has opinions about AGENTS.md/CLAUDE.md files. Best practices get shared, templates get copied, folk knowledge dominates. We analyzed 30k repos and measured the actual state.

Read the article →

Research

The Undiagnosed Input Problem

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

Gábor Mészáros · April 8, 2026 · 8 min read

The AI agent ecosystem has built a serious industry around controlling outputs — guardrails, classifiers, retries, human review. The input side has no equivalent diagnostic layer.

Read the article →

Research

Do NOT Think of a Pink Elephant

Ordering effects in LLM instruction-following — directive first, restriction last.

Gábor Mészáros · March 31, 2026 · 5 min read

You thought of a pink elephant, didn't you? Same goes for LLMs. "Do not use mocks in tests" — and the next file imports unittest.mock on line 3.

Read the article →

Research

Instruction Best Practices: Precision Beats Clarity

Formatting is the signal. Precision is the target.

Gábor Mészáros · March 24, 2026 · 5 min read

Two rules in the same file. Both say "don't mock." One gets followed, one doesn't. Precision matters more than clarity for AI instruction files. We measured the gap.

Read the article →

Field Notes

CLAUDE.md Best Practices: 7 Formatting Rules for the Machine

Don't just write more instructions. Format the ones you have.

Gábor Mészáros · March 3, 2026 · 8 min read

I watched an agent ignore a rule I wrote two hours earlier. Not a vague rule — a specific one. The formatting itself matters. Seven rules for writing CLAUDE.md the machine can follow.

Read the article →

Field Notes

Why Bootstrap Should Be the First Command in Every Agent Session

Don't start a session. Bootstrap it.

Gábor Mészáros · February 24, 2026 · 6 min read

After a 2.5-hour session you accidentally close your coding agent's terminal mid-session. The output is there, the commits are there, but the loaded mental model is gone.

Read the article →

Field Notes

CLAUDE.md Best Practices: Mermaid for Workflows

Don't describe the path. Draw the map.

Gábor Mészáros · February 17, 2026 · 6 min read

A picture says a thousand words. I wanted to see my system — not the code. I wanted to see the workflows. Use mermaid to make CLAUDE.md instructions machine-readable as flow.

Read the article →

Field Notes

CLAUDE.md Best Practices: The backbone.yml Pattern

Don't drop your agent in the dark. Give it a map.

Gábor Mészáros · February 10, 2026 · 5 min read

There's a Dutch scouting tradition called 'dropping' — kids get driven into an unfamiliar forest at night and have to find their way back. backbone.yml is the map for your agent.

Read the article →