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.

State of AI Instruction Quality, Q3 2026: about 400 top repos ban the AI co-author line, and about 330 ask for one. The 10,000 most-starred repos, split on Co-Authored-By. Beside the text, a COMMIT_EDITMSG file whose last line, Co-Authored-By: Claude <noreply@anthropic.com>, is struck through.

Research

State of AI Instruction Quality: Repos Split on Co-Authored-By

About 400 tell the coding agent to leave its name off the commit, about 330 ask for an AI trailer, and a setting does the job in Claude Code

Gábor Mészáros · October 17, 2026 · 5 min read

Read the article →

State of AI Instruction Quality, Q3 2026: main files that say "You are" score lower, at every file size. Median 6.6 against 7.7; 285 of 20,836 main files have the line. Beside the text, a CLAUDE.md file with the line "You are a senior software engineer." highlighted on line 3, above a heading and three short rules.

Research

State of AI Instruction Quality: Files With Persona Lines Score Lower

Main instruction files that tell the agent who it is read like a prompt, and pick up more findings at every file size

Gábor Mészáros · October 16, 2026 · 4 min read

Read the article →

State of AI Instruction Quality, Q3 2026: a third of the top repos write their rules in capitals. 3,319 of 9,784 repos with an agent setup, 36,482 times between them. Beside the text, an AGENTS.md file with a Rules section where IMPORTANT, MUST, NEVER, ALWAYS, CRITICAL and DO NOT are highlighted.

Research

State of AI Instruction Quality: A Third of Repos Shout

The habits in the agent instructions of the 10,000 most-starred GitHub repos, and the setups they travel with

Gábor Mészáros · October 15, 2026 · 5 min read

Read the article →

State of AI Instruction Quality, Q3 2026: CLAUDE.md files that started from /init score higher, at every file size. Median 8.4 against 7.7 across 8,220 files. One in five has the line. Beside the text, a CLAUDE.md file with the /init line highlighted on line 3, above a Commands section with a code block and an Architecture section.

Research

State of AI Instruction Quality: /init Files Score Higher

Most of each file is a map of the repo, and the map is also where its rules get buried

Gábor Mészáros · October 15, 2026 · 4 min read

Read the article →

State of AI Instruction Quality, Q3 2026: 91% of the 10,000 most-starred GitHub repos give their AI agent vague instructions. No file, no command, no tool. The agent has to guess. Beside the text, a grid of 100 squares with 91 lit amber.

Research

The State of AI Instruction Quality: Q3 2026

How 10,000 of the most-starred GitHub repos instruct their coding agents

Gábor Mészáros · October 13, 2026 · 10 min read

Read the article →

A spec file, specs/migrations.md, marked delivered on turn 1, reading that migrations should be reversible where possible and docs should ideally mention which release a change goes out in. An arrow leads to the migration doc the agent wrote on turn 16: down drops the index, and a dashed slot holds Release: TBD.

Field Notes

Progressive Disclosure: Why Isn't Claude Code Following My Instructions?

One spec, worded two ways, loaded into the same 16-turn session. The named version got the release right on every migration; the vague one never did. What that says about what you load, and how to check it from the files.

Gábor Mészáros · October 7, 2026 · 9 min read

Read the article →

A reply's opening token, [status][migration], sends a line into a small central node. Dim rule cards fan out from it and only the ones whose reason matches light up green, above a quiet stream of log lines.

Field Notes

Progressive Disclosure: Claude Code Hooks, Events and Telemetry

Different moments in a Claude Code turn need different rules and different checks. One hook per rule stops scaling, so route every hook event through one small event bus, put telemetry on it first, and see what one entry point buys you.

Gábor Mészáros · September 30, 2026 · 8 min read

Read the article →

A dark field: a crowded stack of dim grey instruction lines on the left, one lit green, and several green lines sweeping right to converge into a single glowing answer node.

Field Notes

Progressive Disclosure: Shaping Claude Code's Output

Opus 5.5 ships less verbose, but the shape of the reply is still yours to set. One concern walked up through every place Claude Code loads an instruction from, to where the harness comes short and what one level up looks like.

Gábor Mészáros · September 22, 2026 · 9 min read

Read the article →

A title card reading Progressive Disclosure beside a dark context-window panel that has loaded only three rules, each tagged what (a skill), where (a path scope), and when (a session-start event), with 97 more rules held back and out of context this turn.

Field Notes

Progressive Disclosure: What, Where, When, and Why

Nested AGENTS.md files let you put a rule where the work is; path config made it finer. But where a rule lives is one lever of three, and getting what loads, where, and when right is what empties the crowded context window.

Gábor Mészáros · September 16, 2026 · 7 min read

Read the article →

A developer at a desk faced with an eleven-file diff behind a single green check, deciding whether to read four hundred lines to approve a one-line ask; three stacked layers guard the path from the agent's output to the main branch: a steering rule, a review pass that runs the code, and a deterministic gate that refuses a write outside the task's named files.

Field Notes

Opus 5: How to Review Generated Code

Opus 5 made writing nearly free and moved the cost onto review; a diff now lands eleven files deep behind a green check nobody read. The answer is not a smarter reviewer but defense in depth: steer the model, run a real review, and gate the rest.

Gábor Mészáros · August 27, 2026 · 10 min read

Read the article →

A developer hits Approve on a four-hundred-line AI diff behind a green check while four more agent pull requests stack up in the queue behind it; off to the side a second robot leans in to grade the first robot's diff, its verdict fraying line by line, while a plain gate stamps the same refusal on file 1 and file 100.

Field Notes

Opus 5: Review bottleneck

Coding agents made code a commodity; the scarce work now is reading it. The obvious fix, a second model checking the first, quietly rebuilds the problem one guess at a time. What actually scales is a gate that does not read the diff at all.

Gábor Mészáros · August 19, 2026 · 6 min read

Read the article →

A dark terminal shows Opus 5 answering a one-line request with a long, rambling reply; an arrow points to a green Stop-hook card whose word-count meter runs green to a 180-word budget then overflows amber to 217 words, marked 'answer first, then stop'.

Field Notes

Opus 5: How to Fix Verbose Output

Opus 5 runs verbose, and a plain `Be concise` rule does not hold it. Three levers get concise replies back, weakest to strongest: a reply-shape rule, an output style, and a Stop hook that refuses a long reply. A plugin carries the setup across projects.

Gábor Mészáros · August 14, 2026 · 7 min read

Read the article →