Audit an agent's context layout against the four places: system prompt, tools, history, tail. Use when the user asks to audit, review or fix their agent's context, prompt caching, token spend per turn, or CLAUDE.md / AGENTS.md layout. Read-only: reports findings and a fix list, changes nothing. Do NOT use for writing evals (evals-bootstrap) or for loop design (goal-test).
Start a fresh conversation. Replace the highlighted placeholders.
Use context-audit to read-only audit [agent project] context sources and return token estimates plus a ranked fix list for [system prompt, tools, history, and tail].
Required inputs
[agent project]
[system prompt, tools, history, and tail]
User-approved starter prompt based on a pinned source version; check current skill requirements before running. Task evidence ↗
SKILL.md
name: context-auditdescription: >- Audit an agent's context layout against the four places: system prompt, tools, history, tail. Use when the user asks to audit, review or fix their agent's context, prompt caching, token spend per turn, or CLAUDE.md / AGENTS.md layout. Read-only: reports findings and a fix list, changes nothing. Do NOT use for writing evals (evals-bootstrap) or for loop design (goal-test).
Audit the context window
Theory: How models read context
and The four places.
Every piece of context belongs in exactly one of four places — a byte-stable
system prompt, a frozen tool set, a compacting history, and a short live tail —
and most agent problems trace back to something sitting in the wrong one.
Core rule
Report and rank; never edit. The output is an audit, and the user decides
what to apply. If they ask you to apply fixes afterwards, that is a normal
edit session, not this skill.
Workflow
Find the context sources. Locate what actually reaches the model:
system prompt (or CLAUDE.md / AGENTS.md for a Claude Code setup), tool
definitions, and — if the project logs requests — one full mid-conversation
payload. If nothing is logged, say so and audit the static files; ask for
one captured request only if the user can produce it cheaply.
Measure the five numbers. Approximate token counts for: system prompt,
tool definitions, message history, retrieved content, current query.
A rough count (chars / 4) is fine; write all five down.
Hunt cache killers. Flag every dynamic value in the system prompt —
timestamps, user names, injected memories, "current task" lines. Each one
is a guaranteed cache miss on every call. Check whether the tool list can
change mid-run; a changing list invalidates the whole prefix.
Hunt window bloat. Find the largest single item in the history. A
verbatim tool result over ~2,000 tokens should have been a file path plus
a one-line receipt. Count tools: past twenty (or ~10K tokens of
definitions), recommend deferred tools + tool search.
Check the tail. Locate the current goal. If it appears only in the
opening message, it lives in the weak middle of the window — recommend
restating it in the tail every three to five steps.
Check guides. If CLAUDE.md / AGENTS.md exists: it should be a few
lines of constraints the code cannot show, not four pages of narrative.
Flag anything the repo already records (structure, history).
Output format
Context audit — <project>
system prompt <n> tok <clean | N dynamic values: ...>
tools <n> tok <n> tools <frozen | mutates mid-run>
history <n> tok largest item: <what, n tok>
tail <present | goal only in opening message>
Top fixes, in order of saved tokens per turn:
1. ...
2. ...
Each fix names the file and line where possible, states what moves to which
of the four places, and estimates the saving. Close with the one-line rule:
stable prefix, frozen tools, compact history, live tail.