audit-session-metrics — Claude Code skill: token-usage audit
A Claude Code skill that reads a session-metrics JSON export and reports on token-usage waste. Invoke it with /audit-session-metrics quick or detailed against a finished export file.
MIT ★ 2,744 updated 2026-10-08
npx skills add centminmod/my-claude-code-setup --skill audit-session-metrics -a claude-codeWhat audit-session-metrics does
The skill runs in two modes. Quick mode checks token ratios, cache health, and the most expensive turns or sessions. Detailed mode covers all of that and adds a scan of CLAUDE.md, settings files, and re-reads — but only for session scope.
Three JSON scopes are supported: session_*.json for per-turn analysis, project_*.json for per-session, and instance/*/index.json for per-project. Scope is detected automatically from digest.scope. The skill calls python3 scripts/audit-extract.py, reads the digest from stdout, and never touches the raw .jsonl directly.
Each audit produces three artefacts: a JSON sidecar with structured findings, a Markdown copy, and inline chat output. Both files land in exports/session-metrics/. The author recommends switching to /model haiku before running on long sessions — on the Anthropic API that is Haiku 5.5 with a 1M-token window; on Bedrock and Google Cloud it is Haiku 4.5 with a 200k window, so only short sessions fit there.
What it helps with
- find the turns and sessions consuming the most tokens
- check cache health across a session-metrics export
- detect repeated reads of settings files and
CLAUDE.md - compare token spend across all sessions in a project
- produce a structured JSON findings report from an audit
Example prompts for the agent
- Run a quick audit:
/audit-session-metrics quick exports/session-metrics/session_abc12345_20250601.json - Run a detailed audit:
/audit-session-metrics detailed exports/session-metrics/project_20250601.json - Audit token usage across all projects:
/audit-session-metrics quick exports/session-metrics/instance/2025-06-01/index.json
How to install audit-session-metrics
1Skills CLI (Vercel)
npx skills add centminmod/my-claude-code-setup --skill audit-session-metrics -a claude-code2Manually: put the files in place
~/.claude/skills/audit-session-metrics/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/audit-session-metrics/
1Skills CLI (Vercel)
npx skills add centminmod/my-claude-code-setup --skill audit-session-metrics -a codex2Manually: put the files in place
.agents/skills/audit-session-metrics/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add centminmod/my-claude-code-setup --skill audit-session-metrics -a cursor2Manually: put the files in place
.cursor/skills/audit-session-metrics/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install centminmod/my-claude-code-setup .claude/skills/audit-session-metrics2Manually: put the files in place
.github/skills/audit-session-metrics/SKILL.md1Skills CLI (Vercel)
npx skills add centminmod/my-claude-code-setup --skill audit-session-metrics -a gemini-cli2Platform CLI
gemini skills install https://github.com/centminmod/my-claude-code-setup --path .claude/skills/audit-session-metricsManually: put the files in place
.devin/skills/audit-session-metrics/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about audit-session-metrics
How do I install audit-session-metrics in Claude Code?
Run npx skills add centminmod/my-claude-code-setup --skill audit-session-metrics -a claude-code; the skill will be available at ~/.claude/skills/audit-session-metrics/SKILL.md.
What does audit-session-metrics do?
It reads a session-metrics JSON export and produces a prioritised report on token-usage waste: ratios, cache state, and the most expensive turns. In detailed mode it also scans configuration files.
What is the difference between quick and detailed mode?
quick covers token ratios, cache health, and top expensive turns. detailed adds a scan of CLAUDE.md, settings.json, and .claudeignore — session scope only; any of those files exceeding 500 lines is itself flagged as a finding.
Source text
SKILL.md · MIT
---
name: audit-session-metrics
effort: low
description: >
Audit a session-metrics JSON export for token-usage waste and produce a
plain-English findings report. Trigger when the user runs
/audit-session-metrics, when session-metrics suggests an audit after an
HTML export, or when the user asks to audit / review / find waste in a
saved session-metrics JSON. Two modes: "quick" (ratios + cache health +
top expensive turns/sessions) and "detailed" (adds CLAUDE.md / settings /
re-read scan). Supports session, project, and instance JSON scopes.
Args: $ARGUMENTS[0] = quick|detailed,
$ARGUMENTS[1] = path to a session-metrics JSON export.
---
# Audit Session-Metrics
Reads a session-metrics JSON export and produces a prioritised, plain-English
audit of token-usage waste. It runs on the session's current model (it no
longer pins one — a hard model pin capped the usable context at that model's
window and broke invocation on long sessions). The work is mostly
summarisation over a small disk-read export, so for a much cheaper run
`/model haiku` before invoking. On the Anthropic API `haiku` is Haiku 5.5
(1M window, so long sessions fit; prompts over 100K tokens bill at its higher
rate). On Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS
it is still Haiku 4.5 with a 200k window — short/early sessions only there.
Supports three JSON scopes auto-detected from `digest.scope`:
- **session** — single session (`session_*.json`) — per-turn analysis
- **project** — all sessions for one project (`project_*.json`) — per-session analysis
- **instance** — all projects (`instance/*/index.json`) — per-project analysis
## Dispatch — how to route this invocation
**First positional argument received:** `$ARGUMENTS[0]`
**Second positional argument received:** `$ARGUMENTS[1]`
Read `$ARGUMENTS[0]` and match by **literal equality**:
| `$ARGUMENTS[0]` | Route | Then read (session scope) | Then read (project/instance scope) |
|-----------------|--------------------------|---------------------------|------------------------------------|
| `quick` | Quick audit | [`references/quick-audit.md`](references/quick-audit.md) | see Scope routing below |
| `detailed` | Detailed audit | [`references/detailed-audit.md`](references/detailed-audit.md) | see Scope routing below |
| *(empty / other)* | Print usage and stop | this file's "Usage" block below | — |
`$ARGUMENTS[1]` must be the path to a JSON export written by session-metrics.
Accepted patterns:
- `exports/session-metrics/session_<id8>_<ts>.json` — session scope
- `exports/session-metrics/project_<ts>.json` — project scope
- `exports/session-metrics/instance/<datedir>/index.json` — instance scope
If `$ARGUMENTS[1]` is missing, empty, or the file does not exist, print:
> Usage: /audit-session-metrics {quick|detailed} <path-to-session-metrics.json>
>
> Accepted JSON types:
> session: exports/session-metrics/session_*.json
> project: exports/session-metrics/project_*.json
> instance: exports/session-metrics/instance/*/index.json
…and stop without further work.
## Steps
1. Run the extract helper once with the input path:
```
python3 scripts/audit-extract.py $ARGUMENTS[1] --mode $ARGUMENTS[0]
```
The helper emits a single JSON digest to stdout. Read `digest.scope`
from the output — it will be `"session"`, `"project"`, or `"instance"`.
**Do not** read the raw `.jsonl`, and **do not** re-derive numbers the
digest already carries.
2. **Scope routing — read the matching reference file:**
| `digest.scope` | `$ARGUMENTS[0]` | Reference file |
|----------------|-----------------|----------------|
| `session` | `quick` | [`references/quick-audit.md`](references/quick-audit.md) |
| `session` | `detailed` | [`references/detailed-audit.md`](references/detailed-audit.md) |
| `project` | `quick` | [`references/project-quick-audit.md`](references/project-quick-audit.md) |
| `project` | `detailed` | [`references/project-detailed-audit.md`](references/project-detailed-audit.md) |
| `instance` | `quick` or `detailed` | [`references/instance-quick-audit.md`](references/instance-quick-audit.md) |
Follow the playbook step-by-step. Do not improvise additional phases.
3. For `detailed` mode **on session scope only**, the playbook also asks
you to read the user's config files (`~/.claude/CLAUDE.md`,
`./CLAUDE.md`, `~/.claude/settings.json`, `./.claude/settings.json`,
`./.claudeignore`). Each is capped at ≤500 lines — if a file is
bigger, that itself is a finding.
4. **Output contract — three artefacts.** All playbooks specify the
same three-artefact contract:
- **JSON sidecar** at `<project>/exports/session-metrics/audit_<id8>_<ts>_<mode>.json` — structured findings (versioned schema, enum'd metrics).
- **Markdown copy** at `<project>/exports/session-metrics/audit_<id8>_<ts>_<mode>.md` — same content rendered for humans.
- **Inline chat output** — the markdown content printed in your reply.
`<id8>` and `<ts>` come from `digest.session_id_short` and
`digest.ts_str` (the helper parses the input filename and
normalises all three filename patterns).
5. **Write order.** Populate the JSON object first using the digest
values, write the JSON sidecar, render the markdown using the
template in the playbook, write the markdown copy, then print the
markdown inline (without the H1 heading — the chat client already
shows context above the audit). Finish with two stderr-style lines
on their own:
`[audit] saved → <json-path>`
`[audit] saved → <md-path>`
**IMPORTANT:** Use the **Write tool** directly for steps 3 and 5.
Do NOT generate a Python script (e.g. writing to `/tmp/audit_synthesis.py`
and executing it) — this adds unnecessary failure modes (syntax errors,
f-string escaping) and is never required. Build the JSON in the AI's own
context; call Write. Render the markdown in context; call Write again.
## Tone
- **Direct and specific.** Cite the exact ratio, dollar figure, or turn
index. No motivational language, no LLM-theory padding.
- **Prioritise by impact.** Sort findings so the costliest fix is first.
- **Quote sparingly.** Snippets capped at 5 lines each.
- **Honour the playbook.** If quick-audit.md asks for 5 rows, produce 5
rows — don't invent a 6th to look thorough.
## Why this is a separate skill
The audit is a distinct, user-initiated analysis step over a finished
export, not part of generating one. It reads only the on-disk JSON — never
the conversation — so it stands alone as its own turn, which is why
session-metrics *suggests* `/audit-session-metrics` rather than invoking it
programmatically. Running it as a fresh slash command keeps the turn focused
and lets the user decide when to spend on it. The work is summarisation-heavy
and the input is tiny, so the cost path is `/model haiku` before invoking
(Haiku 5.5 with a 1M window on the Anthropic API; on other providers
`haiku` is Haiku 4.5 with a 200k window, so short/early sessions only) — far
cheaper than a frontier model, with identical output (every dollar figure is
pre-computed by `audit-extract.py`, not guessed by the model).
## Reference files
### Session scope
- [`references/quick-audit.md`](references/quick-audit.md) — Distilled
ratios + cache health + top expensive turns. Read when scope=session,
mode=quick.
- [`references/detailed-audit.md`](references/detailed-audit.md) — Quick
audit findings plus config + re-read scans. Read when scope=session,
mode=detailed.
### Project scope
- [`references/project-quick-audit.md`](references/project-quick-audit.md) —
Per-session cost outliers, cache health, breaks, weekly trend. Read when
scope=project, mode=quick.
- [`references/project-detailed-audit.md`](references/project-detailed-audit.md) —
Project quick findings plus per-session turn-level drilldown on the most
expensive sessions. Read when scope=project, mode=detailed.
### Instance scope
- [`references/instance-quick-audit.md`](references/instance-quick-audit.md) —
Per-project cost breakdown, cache health, weekly cost trend. Used for both
quick and detailed modes (no per-turn data available at instance scope).
© centminmod · MIT · centminmod/my-claude-code-setup