Skills vs tree docs
Скилы или дерево документов?
C · Weak signal Short Telegram post on the LLM Under the Hood channel relaying one course participant's informal comparison; a useful weak signal, but no question set, accuracy figures or method are given.
Summary based on the full text·Original: about 2 minutes to read
In short
A post on the Russian Telegram channel LLM Under the Hood (12 October 2026) sharing a report from a participant in the channel author's AI coding webinar, Alexey S. He compared two ways of giving coding agents project knowledge: "skills" and "tree docs", meaning an instruction in AGENTS.md to list a docs folder whose file names describe their content, then open what is relevant. He tested several small projects and one large one (FastAPI) in Claude and Codex. Tree docs won on answer accuracy everywhere, and the gap grew with codebase size.
Key numbers
- 30%extra tokens Claude spent with tree docs compared with skills, per the tester (not seen in Codex)
- 1%apparent share of the maximum context window Claude allots to skill descriptions before truncating them
As reported by the source.
Russian·Web page·opens t.me
What it reports
- Tree docs beat skills on answer accuracy and on calling the docs or skills a question implied, in both Claude and Codex.
- On small projects both approaches work; as the codebase grows, tree docs pull further ahead.
- On utilitarian questions Claude often ignored skills and went straight to the code; this did not happen with docs.
- With tree docs Claude spent about 30% more tokens, which the tester attributes to how caching is marked; Codex showed no such overhead.
- Claude appears to truncate tool and skill descriptions: beyond a budget of about 1% of the maximum context window, skills keep only their name, and which ones get cut is not predictable in advance.
- The channel author's own setup is one AGENTS.md line telling the agent to run
tree docsfirst and to read a document-writing guide before editing docs.
How they got it
An informal side-by-side test by one practitioner, reported in a chat and reposted. The tree-docs setup was self-written, following the webinar's instructions. No number of questions, scoring method or accuracy figures are published.
Limits
- Single tester, unspecified question set and scoring, no raw results.
- The 1% truncation budget is inferred from observed behaviour, not documented.
- The channel author teaches the tree-docs method, so the post supports his own course material.
Why it matters for companies
Teams writing agent instructions for large repositories face the same choice. The signal here is that a plain, well-named document tree plus one routing instruction may retrieve better than a growing pile of skills, and that skill descriptions can be silently dropped once there are many. Cheap to test on your own repository before standardising either pattern.
Key passages in translation
Russian original
tree docs побеждает буквально везде (и в клоде, и в кодексе).
Tree docs win literally everywhere (in both Claude and Codex).
Russian original
по мере роста кодовой базы tree docs начинает выигрывать все заметнее
As the codebase grows, tree docs start to win more and more clearly.
Russian original
Судя по всему там бюджет 1% от макст контекстного окна на скиллы - все что не влезает получает только название скила, без описания.
Apparently there is a budget of 1% of the max context window for skills; anything that does not fit gets only the skill's name, with no description.
Summary based on the full text. Read the original for the full argument and data.Read the original on t.me