mozg.betaSign in

Skills & context files vs a brain

The most expensive answer
is the confident wrong one.

You gave your agent skills, CLAUDE.md files, a folder of carefully written knowledge. It reads them and answers with total confidence — and three months later half of those answers are quietly wrong. Here is exactly where that pain comes from, error by error.

Error № 1

You pay for every word, every session — needed or not.

contextthe whole file, every sessionbrain — only the 3 notes the task needed

A skill or context file loads front to back into the window whether today's task needs one line of it or none — and the folder only grows. Past a point it physically cannot fit, so people start deleting knowledge to make room, which is exactly backwards. A brain is searched: the agent asks, gets the three notes that answer, and the other seven hundred cost nothing. Knowledge scales; the bill does not.

Error № 2

The docs changed on Tuesday. Your agent finds out never.

file: same bytes, month after monthbrain: re-read, re-examined, correctedthe API changed on Tuesdayfile: kept teaching the old shapebrain: re-read it Tuesday night,re-sat its exam, score moved

Nothing in a file tells you it went stale — it keeps answering in yesterday's shapes, confidently, and you find out from a broken build. A brain re-reads its sources on a schedule, supersedes what changed, re-sits an exam generated from its stated goal — and any agent using it mid-task can flag a note the moment reality disagrees. Files decay silently; a brain files a complaint.

Error № 3

Nobody can answer "what does it actually know?"

Ask whether your knowledge folder covers the error codes of one endpoint and the honest answer is a shrug and a grep — so agents improvise, and improvisation reads exactly like knowledge until it ships. A brain answers with a number: an exam built from its goal, re-sat after every change, with the failures naming exactly which material is missing. Not a vibe — a score you can watch move.

a folder of files

coverage
?%

the same knowledge, as a brain

measured, re-sat after every change
92%

Error № 4

Every teammate has a copy. Every copy is different.

Hand someone a file and you have forked your knowledge: their copy and yours drift from that day on, and a correction lands in one of them. A brain is one URL — every agent on the team reads the same living thing, a fix approved once reaches every reader instantly, and it plugs into Claude Code, Codex, Cursor and whatever ships next month, because MCP is the socket.

Error № 5

You corrected the agent yesterday. It forgot by this morning.

The hardest-won knowledge — the pitfall that cost an afternoon, the convention you explained for the fifth time — lives and dies inside one conversation. With a brain, the agent writes the lesson back: you approve it once, and every future session of every agent starts already knowing it. Corrections compound instead of evaporating.

Where skills genuinely win

A skill that runs things — shell scripts, deploy gates, asset pipelines — is a tool, and a brain does not replace tools. The honest rule: procedures with code stay skills; knowledge becomes a brain. If your skill is mostly prose about how something works, it is a brain wearing the wrong container — and paying the token tax for it.

Cost per sessionwhole file in context, every timethree notes, only when asked
When docs changesilently wrongre-read, re-examined, score moves
Coverageunknowablea number, with the gaps named
Correctionsdie with the conversationwritten back, approved once, kept forever
Teamcopies that driftone URL, one truth
Selling itnot a thingstorefront, exam score, 95% to you
Running scripts✓ skills winnot a brain's job — keep the skill

columns: — · as a file · as a brain

Turn your knowledge folder into a brainThe gentler comparison: brain vs a fileSee measured brains