# Pasting the brain into a chat AI that has no tools Some chat AIs cannot read a repository: no connector, no sandbox, no files. They can still read this brain and hand changes back, if the message is built right. This is the third way in; the first two are a coding agent with the repo on disk and a chat AI with a connector or a clone. ## Build the message ``` python -X utf8 tools/brain.py paste --out paste.md ``` The message is: one opening line, the notation card, a `---` line, then the declaration blocks of `handoff/CURRENT.md` (frontmatter and the claim-and-close notes left out). Nothing else: no runtime, no spec link. Send it as the first message, as one message. The expected reply is just "done reading". The card follows the owner's language. `rules/GRAMMAR-CARD.md` is the English card, `rules/GRAMMAR-CARD.zh.md` the Chinese one; `--lang auto` picks by the language of the hand-off, `--lang zh` or `--lang en` overrides, and the opening line follows the card that was picked. Another language is one more file, `rules/GRAMMAR-CARD..md`, translated from the English card line by line; it is not measured until you measure it. The command refuses to build the message when any line looks like a secret; fix the hand-off first. Since 3.2 the opening line also says how a gap comes back: not filled from recall, but as an `::EXCLUDE{}` block with the candidates, what refuted each one, and what remains; `brain.py review` and `verify` read it like any other block. ## One check before real work Ask: "Restate the hand-off's progress in three parts: done, next, unverified. Item by item, verbatim, keep the numbering." A model that paraphrases, merges items or drops the numbering did not read the notation; use a stronger tier or a shorter hand-off. Do not fix it by adding instructions to the card; the card is measured and tested as it is. ## Taking the answer back Before you ask for the block, say that items are identifiers: an item moves between `done`, `next` and `unverified` whole, and news goes in a new item. Without that sentence a model rewrites the item it moves, and the review rejects the rewrite as a drop plus an addition. Whatever comes back is a candidate. Paste it into a file and run ``` python -X utf8 tools/brain.py handoff review returned.md --overlay --out candidate.md ``` against the current hand-off. `--overlay` lays the returned blocks over the baseline by identifier and carries the rest of the baseline in, so one `::PROGRESS{}` block is enough; `--out` writes the composed whole, which is what replaces `handoff/CURRENT.md` when you accept. It rejects a deleted or weakened boundary, a forbidden or unevidenced item in `done`, a missing key, an empty value, a changed identifier and a silent drop; it lists the rest for you to decide. Only an accepted candidate replaces `handoff/CURRENT.md` (see the skill, "Merging a hand-off another model wrote"). ## The ledger in the message Since 3.0 the message ends with the ledger: one `::STATE{@LEDGER…}` line and one row per door with its door, kind, verify command, how to get back in and the date it last opened, and **no value**. A chat AI with no tools cannot open a door, and this message leaves the private repository, so the values stay home. `--with-values` puts them back for the owner who means to hand the whole brain to a model that has a shell; nothing is masked on that message and the tool says so. Since 3.1 a key-shaped string left anywhere in the message is masked to `{{UNHARVESTED }}` instead of refusing the paste; `env/pepper`, `env/keys.env`, `log/` and `pair/` are never part of it. ## What was measured - The English card (`rules/GRAMMAR-CARD.md`, 1351 bytes): 381 to 410 tokens on ten models' tokenizers (MiniMax, Grok 4.20 and 4.7, GLM, Kimi, DeepSeek, Qwen, Gemini, GPT, Doubao), 539 on Claude's. Measured 2026-11-04; not yet tested in conversation. - The Chinese card (`rules/GRAMMAR-CARD.zh.md`, 1212 bytes): 352 to 392 tokens on the same ten, 542 on Claude's. Tested in conversation on twelve model tiers from weak to strong, 2026-11-03 and 2026-11-04. - `BRAIN.md`, the entry for a chat AI with a connector: 1552 bytes; 399 to 406 tokens on GPT, GLM and Kimi, 427 on Doubao, 588 on Claude's, net of each chat template, measured 2026-11-05 after the 2.1 edit of point 2. - One web endpoint, 2026-11-05: ChatGPT's web chat, signed out, free tier, no files and no tools. The Chinese card plus a 14-block hand-off went in as one message and came back as "done reading"; the check above restated 4, 6 and 9 items verbatim, confirmed by script. The first write-back rewrote the item it moved and the review rejected it; told the identifier rule, the second write-back passed (one flag, one transition, one addition) and was accepted.