Skip to content

Commit ebb4704

Browse files
committed
2026-02-25 23:07:03
1 parent d1b8c7e commit ebb4704

2 files changed

Lines changed: 67 additions & 0 deletions

File tree

eng_2026/02/2026-02-25-22-43.md

Lines changed: 33 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,33 @@
1+
**Does AGENTS.md Actually Help?**
2+
https://arxiv.org/abs/2602.11988
3+
The first large-scale empirical study testing whether repository-level context rule files actually help. Three scenarios were tested on real SWE-bench tasks and a custom dataset of repositories containing `AGENTS.md` files.
4+
5+
Main conclusion: Modern agents are excellent at finding necessary information directly in the code (package.json, README, schemas, types). Additional instructions often **hinder** rather than help.
6+
7+
Key downsides of such files: Increased costs as the agent reads more files, runs more tests, and performs redundant actions trying to "fulfill all requirements" in `AGENTS.md`, where outdated instructions often mislead the model.
8+
9+
If writing `AGENTS.md` manually — keep only minimal, specific requirements to fix recurring agent errors.
10+
11+
[https://www.youtube.com/watch?v=GcNu6wrLTJc](https://www.youtube.com/watch?v=GcNu6wrLTJc)
12+
13+
![YouTube Preview](https://img.youtube.com/vi/GcNu6wrLTJc/mqdefault.jpg)
14+
15+
**Practical recommendations from Theo:**
16+
- Better to invest time in clean architecture, strong typing, tests, CI/CD, and documentation **directly in the code**.
17+
- Blindly following "best practices" from agent developers can be harmful. Try removing CLAUDE.md / AGENT.md and compare the agent's speed and quality.
18+
- If a file is necessary — keep it **short** (15–30 lines) and focused on fixing **one** specific problem.
19+
20+
Special prompt engineering technique for AI agents: Instead of long rules in `CLAUDE.md`, add short, intentionally **false** but useful statements that guide the model's behavior much more effectively.
21+
22+
**Examples shown by Theo:**
23+
- **"This project is green"** → The agent stops searching for non-existent errors, doesn't run extra tests, and doesn't "fix" what isn't broken.
24+
- **"This is a brand new feature"** → The agent doesn't copy old code or try to "adapt" existing solutions, but writes clean code from scratch.
25+
- Other common variants: "All tests are passing", "We always write production-ready code".
26+
27+
**HN Discussion:**
28+
https://news.ycombinator.com/item?id=47034087
29+
Almost everyone agrees that LLM-generated context files (often via the `/init` command) worsen results. Well-written manual `AGENTS.md` files are useful only if they contain non-obvious domain knowledge that the model cannot infer from the code. Add them only **after** failed agent attempts.
30+
31+
**Critique of the study:** Lack of code quality measurement (only success rate), Python-only dataset, mostly small/LLM-generated repositories, and models change rapidly — results might differ in a month.
32+
33+
#prompts

ukr_2026/02/2026-02-25-22-43.md

Lines changed: 34 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,34 @@
1+
**Чи справді допомагає AGENTS.md**
2+
https://arxiv.org/abs/2602.11988
3+
Перше велике емпіричне дослідження, яке перевіряє, чи справді допомагають репозиторійні контекстні файли правил. Тестували три сценарії на реальних задачах SWE-bench та власному датасеті репозиторіїв з файлами `AGENTS.md`.
4+
5+
Головний висновок: сучасні агенти чудово самі знаходять потрібну інформацію в коді (package.json, README, схеми, типи). Додаткові інструкції частіше **заважають**, ніж допомагають.
6+
7+
Ключові мінуси таких файлів: зростання вартості так як агент більше читає файли, запускає тести, виконує зайві дії, бо намагається «виконати всі вимоги» з `AGENTS.md` де зазвичай вже застарілі інструкції вводять модель в оману.
8+
9+
Якщо писати `AGENTS.md` вручну — тільки мінімальні, точкові вимоги для виправлення конкретних повторюваних помилок агента
10+
11+
[https://www.youtube.com/watch?v=GcNu6wrLTJc](https://www.youtube.com/watch?v=GcNu6wrLTJc)
12+
13+
![YouTube Preview](https://img.youtube.com/vi/GcNu6wrLTJc/mqdefault.jpg)
14+
15+
Практичні рекомендації від Theo:
16+
- Краще інвестувати час у чисту архітектуру, сильну типізацію, тести, CI/CD і документацію **прямо в коді**.
17+
- Сліпе слідування «best practices» від розробників агентів може бути шкідливим. Спробуйте видалити CLAUDE.md / AGENT.md і порівняйте швидкість та якість роботи агента.
18+
- Якщо файл все ж потрібен — робіть його **коротким** (до 15–30 рядків) і тільки для фіксу **однієї** проблеми.
19+
20+
Спеціальна техніка prompt engineering для AI-агентів: замість довгих правил у CLAUDE.md додаєте короткі, свідомо **неправдиві**, але корисні твердження, які керують поведінкою моделі набагато ефективніше.
21+
22+
**Приклади, які показує Theo:**
23+
- **«This project is green»** (або українською: «Проєкт зелений / все в ідеальному стані») → Агент перестає шукати неіснуючі помилки, не запускає зайві тести, не «фіксить» те, що не зламане.
24+
- **«This is a brand new feature»** (або «Це абсолютно новий функціонал, якого раніше не було») → Агент не копіює старий код, не намагається «пристосувати» існуюче рішення, а пише чисто і з нуля.
25+
- Інші варіанти, які часто використовують: – «All tests are passing» – «We always write production-ready code»
26+
27+
Обговорення HN
28+
https://news.ycombinator.com/item?id=47034087
29+
Усі майже одностайно згодні, що файли контексту, згенеровані LLM (часто це команда `/init`), погіршують результат. Добре написані вручну файли `AGENTS.md` корисні, але тільки якщо містять неочевидні знання про домен, яких модель не може вивести з коду. Додавати їх тільки після невдалих спроб агента.
30+
31+
Критика дослідження: відсутність вимірювання якості коду (тільки success rate), датасет тільки Python, переважно маленькі/LLM-генеровані репозиторії, до того ж моделі швидко змінюються — результати можуть через місяць вже не бути таким самим.
32+
33+
34+
#prompts

0 commit comments

Comments
 (0)