python-vibe

ask, test, fix, add

GitHub

Investigation: skills written for the everyday 8B

The everyday brain is small. It does not follow a 7-step essay. It copies one Action: block — often the first example in the system prompt, or it pastes the whole menu. Skills are published only after scripts/measure/skill_probe.py on this machine.

Related: everyday-laptop.

Experiments (29 Aug 2026, llama3.1:8b)

Verbose skills (when-to / numbered steps):

Task First action Result
what does apply_source refuse? read random file Missed code.py
same + “start with grep” read then grep then done No refuse rules
add … multiply Find: def add(left Syntax break (harness now refuses)
same Action: write-tests Not an action

Tiny skills + harness prelude + short system prompt (scripts/measure/skill_probe.py):

Task prelude First parsed action Notes
apply_source refuse yes done Summary named empty draft, 2/3 length, syntax
apply_source refuse no grep + wrong read Invented a non-ASCII story
add multiply yes, long system prompt skill (menu dump) 8B pasted every Action example
add multiply yes, short system prompt patch + Append Intended first step

Then: if the model still dumps a menu, parse_turn_smart keeps one block — done/locate on a question, patch/edit on an add.

Third-party small tree (33 .py/.md, Mode: small, src/harness scope), same day, after locate shipped:

Task First action Result
what does listen_addr return? read the already auto-read file Correct tuple answer after a wasted step. start_hint still said grep/read.
what does complete do after two blocked drafts? (probe) read + Append: body Would have edited a question if applied.

Harness now: start_hint(..., located=True) says done only; questions refuse patch/edit/run; a second read of the auto-read file is refused. File: is an alias for Path:. A done that omits the -> type (for example tuple[str, int]) is refused once so the 8B quotes the signature instead of “a tuple”.

Add-feature on the kit fixture (29 Aug 2026): multiply landed, but the test was Append: after if __name__ with no import, so unittest ran 1 test. Harness now inserts def test_ before if __name__, adds the import, refuses a second locate after prelude, and injects the write-tests skill after the implementation patch.

What to publish

These kit skills, each a single copy-paste Action (no essays):

Ship is limited (src/harness/ship/git_ship.py): no --force, not main/master, unstage .env / credentials.json. Order: issue → branch → patch → commit → push → pr.

Harness: refuse_opaque_names (calc/tmp/x, CamelCase functions, snake_case classes) and refuse_layout (no impl in __init__.py or scripts/; split a module with 4+ functions).

Plus harness: prelude() locate before the model, Action: locate, parse_turn_smart, syntax reject, skill-name-as-action.

Live create a package for total_price (empty dir, 29 Aug 2026): pkg/__init__.py (export only) → pkg/total_price.py → a TestCase → unittest OK → done. A bare def test_*(self) is wrapped. done before a passing run is refused.

Do not publish a new skill until skill_probe.py shows the intended action with "prelude": true.

Do not