context-audit

1# Audit the context window
2 
3Theory: [How models read context](https://secondbrainos.dev/#/read/course-1-context/how-models-read)
4and [The four places](https://secondbrainos.dev/#/read/course-1-context/the-four-places).
5Every piece of context belongs in exactly one of four places — a byte-stable
6system prompt, a frozen tool set, a compacting history, and a short live tail —
7and most agent problems trace back to something sitting in the wrong one.
8 
9## Core rule
10 
11Report and rank; never edit. The output is an audit, and the user decides
12what to apply. If they ask you to apply fixes afterwards, that is a normal
13edit session, not this skill.
14 
15## Workflow
16 
171. **Find the context sources.** Locate what actually reaches the model:
18   system prompt (or CLAUDE.md / AGENTS.md for a Claude Code setup), tool
19   definitions, and — if the project logs requests — one full mid-conversation
20   payload. If nothing is logged, say so and audit the static files; ask for
21   one captured request only if the user can produce it cheaply.
222. **Measure the five numbers.** Approximate token counts for: system prompt,
23   tool definitions, message history, retrieved content, current query.
24   A rough count (chars / 4) is fine; write all five down.
253. **Hunt cache killers.** Flag every dynamic value in the system prompt —
26   timestamps, user names, injected memories, "current task" lines. Each one
27   is a guaranteed cache miss on every call. Check whether the tool list can
28…