Skip to main content

You were on git main. Claude wasn't.

Map visible context surfaces before your next expensive session. One command. Read-only. Instant report.

389,169
tokens burned in one incident
1
command to diagnose
< 10s
to generate report
0
files modified (read-only)

How It Works

Three steps. Instant visibility.

Step 1

Run

Point the script at any project directory. It scans your visible Claude Code context surfaces.

Step 2

Report

Get approximate token-equivalent rankings for every context surface: instruction files, rules, skills, hooks, transcripts.

Step 3

Act

Use the generated Runtime Diet Card, next-session rule, and Opus Passport to start your next session clean.

bash scripts/claude-runtime-diet-report.sh /path/to/project

What You Get

A complete diagnostic toolkit extracted from a real 389K-token incident investigation.

One-command Runtime Diet Report generator
Read-only Claude Code baseline script
Approximate token-equivalent rankings
10-minute START-HERE workflow
Auto-filled top-five risk worksheet
Auto-filled Runtime Diet Card
Auto-drafted Opus Passport pre-session gate
Context surfaces checklist
Session hygiene playbook
Pre-session checklist
Hook context-risk guide
Move-procedures-to-skills guide
Troubleshooting guide
Sanitized 389K-token case study

Your Runtime Diet Card

The report generator creates a share-safe summary of your visible context surfaces. No secrets. No file contents. Just the numbers that matter.

Largest visible context surface ranked
Session branch vs. git branch status
Hooks, skills, and rules surface area
Generated next-session rule
Example Runtime Diet Card showing visible context surfaces
Honest disclaimer

Do not buy this if you only use Claude Code casually.

Buy it if at least two are true:

You resume long sessions across unrelated work
Your CLAUDE.md or rules have grown over months
You use hooks, skills, MCP, or custom memory
You hit usage limits faster than the visible prompt explains
You are about to start an expensive Opus / Max session

Built From a Real Incident

Not theory. Not a template pack. A diagnostic method extracted from a 389,169-token startup burn investigation.

Read-Only by Default

The baseline script never modifies your files. Safe to run on any project.

ZIP Integrity

Verify your download matches exactly.

SHA256: a614ba1b8f5c714cab7429278291972542b9ec7645afc201ec77989847173105

Battle-Tested

Extracted from a real long-running agent harness — not an academic exercise.

Before and after Runtime Diet — visible context surfaces comparison

Choose Your Path

The kit shows you the surfaces. The report tells you what to keep.

Most Popular

Self-Serve Kit

$49
  • One-command report generator
  • Read-only baseline script
  • Auto-filled Runtime Diet Card
  • Opus Passport draft
  • Session hygiene playbook
  • Pre-session checklist
  • 389K-token case study
  • Troubleshooting guide
5 Spots Only

Founding Audit

$199
  • Everything in Self-Serve Kit
  • Founder-led 3-bullet teardown
  • Redacted baseline review
  • One prioritized diet action
  • 72-hour turnaround

Async Deep Report

$297
  • Everything in Founding Audit
  • Full async Runtime Diet Report
  • Keep-vs-cut analysis
  • Prioritized cleanup plan
  • Before/after projections
  • Skills/hooks migration advice

FAQ

Does this guarantee lower Claude usage?

No. It helps you identify visible context surfaces and apply safer hygiene. It cannot guarantee account-specific savings.

Does this show exact billed tokens?

No. The script prints rough byte_count / 4 estimates so you can rank local surfaces. It does not access Anthropic billing, private system prompts, tool schemas, or exact model-side context state.

Does the script change my Claude settings?

No. The report generator and baseline script are read-only by default.

Is this official Anthropic tooling?

No. This is an independent practitioner diagnostic extracted from a real long-running agent harness after a 389,169-token incident review.

Can I use it with Codex or other agent CLIs?

The method transfers, but v1 is written around Claude Code because that is where the case study and buyer pain are sharpest.

Stop guessing. Start measuring.

Your repo branch is clean. Your Claude session branch might not be.

No magic. No fake savings claims. Just a practical diagnostic from a real 389K-token incident.