Skip to content

docs: prune AGENTS.md - #209

Merged
lesnik512 merged 1 commit into
mainfrom
docs/prune-agents-md
Sep 12, 2026
Merged

docs: prune AGENTS.md#209
lesnik512 merged 1 commit into
mainfrom
docs/prune-agents-md

Conversation

@lesnik512

Copy link
Copy Markdown
Member

85 lines to 66. No rule is dropped and no behaviour changes — what goes is boilerplate, lookups the
environment already answers, and one bullet that had outgrown its neighbours.

The warn_at_caller bullet: 12 lines to 5

#208 grew this rule to 12 lines against siblings of 2-3. Most of the bulk was an inventory: which
surviving literal stacklevel sits at 3, which at 4, and why one is a frame shallower than the
other.

That arithmetic is worth knowing and is not worth carrying on every turn. The rule an agent has to
follow is that neither depth is constant and the helper walks out to the first frame outside
lite_bootstrap; the measurements live in #202, so the bullet points there instead of restating
them.

## Agent skills: 13 lines to 5

Three ### subheadings, each wrapping a single sentence, to carry three pointers. Now a
three-bullet list under ## Agent docs. Same three pointers, same targets, none of the scaffolding.

Deletions

  • This file provides guidance to Claude Code (claude.ai/code)... — boilerplate that changes no
    behaviour.
  • just --list, or read it — the justfile answers this, and cannot go stale doing so.
  • Read them. and the lite_bootstrap/ prefix in Architecture. The directory is discoverable; the
    naming convention is the part worth stating, because it says where a new file goes.
  • Two of the three alias examples. One worked example carries the rule.

One rewording

The type-checking rule ended Do not add it, and do not act on its diagnostics. Steering by
prohibition keeps the forbidden tool in view, so it is now the positive target: ty is the only
supported type checker, act on its diagnostics alone, and adding Pyright yields noise rather than
findings. The guardrail is unchanged; there is just no ban to read past.

Say the word if you would rather keep the explicit prohibition — it is a one-line revert.

85 lines to 66. Every rule that was live stays live; what goes is boilerplate,
lookups the environment already answers, and one bullet that had outgrown its
neighbours.

The `warn_at_caller` rule went from 12 lines to 5. Its three siblings run 2-3
lines each, and most of its bulk was an inventory of which surviving literal
`stacklevel` sits at 3, which at 4, and the frame arithmetic behind each. The
rule an agent has to follow is that neither depth is constant and the helper
walks out instead; the arithmetic belongs in #202, which measured it, so the
bullet now points there.

`## Agent skills` spent three `###` subheadings and 13 lines carrying three
one-sentence pointers. Now a three-bullet list, 5 lines, same three pointers.

Deleted: the "This file provides guidance to Claude Code" opener; `just --list,
or read it`, which restates what the justfile answers; "Read them." and the
`lite_bootstrap/` prefix in Architecture, where the naming convention is the
part worth stating because it says where a new file goes; and two of the three
alias examples, since one worked example carries the rule.

The type-checking rule ended "Do not add it, and do not act on its
diagnostics." Naming the forbidden tool keeps it in view, so it is now stated
as the target behaviour: `ty` is the only supported type checker, act on its
diagnostics alone, and adding Pyright yields noise rather than findings. Same
guardrail, no prohibition to read past.
@lesnik512
lesnik512 merged commit d2f7be7 into main Sep 12, 2026
12 checks passed
@lesnik512
lesnik512 deleted the docs/prune-agents-md branch September 12, 2026 11:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant