The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →In Sungwoo Lee’s report, a custom Claude Code /his command produced 138,701 output tokens across six runs in one project; 132,721 were thinking tokens, or 96% of that output. The figure describes Lee’s command and measurement setup—not a typical result for Claude Code skills or commands. Lee later moved repeatable bookkeeping into Python scripts, but did not repeat the measurement rigorously, so the redesign’s effect on token use is unknown.
What Lee’s 96% figure measures
Lee’s /his command appends a short record to a project’s HISTORY.md. The record captures decisions and their reasons, approaches tried and abandoned, and a useful starting point for the next session. Across six invocations in one project, Lee reports 138,701 total output tokens, including 132,721 thinking tokens. He describes the written history entry as about 1,000 tokens per run.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
Those numbers come from Lee’s account in a DEV Community article dated October 1, 2026. They are not independently verified, and the six-run sample does not establish how other commands, projects, model versions, or usage patterns behave.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallHow he counted tokens
Lee says his Claude Code session transcripts were stored as JSONL and that assistant turns included a usage block. His script identified each /his invocation and aggregated usage from the command through the next user turn.
#1 Best Overall
A complication, in his setup, was that a transcript could record the same API request more than once. Lee says he deduplicated entries by requestId before adding token counts, to avoid counting a request two or three times. This is his description of the transcripts and script he used, not a guarantee about every Claude Code transcript format or version.
What changed in the command workflow
As rules accumulated in response to mistakes, the command file grew to 16.6 KB. Lee split out mechanical work into two Python scripts, reducing the command file to 5.3 KB. The revised workflow assigns scripts and the model different kinds of work:
| Component | Lee’s described responsibility |
|---|---|
his_prep.py <slug> |
Collects facts such as changed files, commits, and Git state, then creates an entry skeleton with four empty sections. It stops near the compaction threshold or when the session is effectively empty just after /clear. |
his_finish.py <slug> |
Refuses to finish if any of the four sections is blank; inserts and reorganizes entries; and checks links, file size, and uncommitted changes. |
| Claude Code command | Writes the reasoning that the scripts cannot reliably derive from facts: decisions and why; rejected alternatives and why; failed approaches (or “none”); and what the next session should do first. |
The revised command no longer asks the model to copy file lists or commit hashes already collected in the facts file. Lee had tried pre-populating plausible reasoning from a diff, but found that a seemingly complete entry could prevent the actual reasoning from being written, so he removed that behavior.
Free tools Windows power users keep installed
One-click scans. No signup required.
Why the redesign is not proof of token savings
Lee says he has not repeated the original six-run measurement with the same rigor after changing the workflow. He therefore does not report an “after” percentage. The evidence supports a description of what changed—scripts take on fact collection, checks, and bookkeeping, while the model writes the reasons—but not a measured reduction in token use.
There is also no controlled comparison, independent replication, or confidence estimate in the report. The 96% figure should be read as a result from six runs of one custom command in one project, not as a general benchmark or a prediction for another user’s setup.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What to automate—and what to leave to the model
Lee’s practical distinction is between deterministic execution and interpretation. File lists, commit identifiers, and validation checks can be gathered or enforced mechanically. The reasons a developer chose one approach over another may not be safely inferable from a diff, so the person or model that did the work needs to record them.
Rank #4
As Lee puts it, “A long skill isn’t mainly a context cost. It’s the reasoning cost of making the same decisions again on every run.” The design lesson is not that every lengthy command wastes tokens; it is to look for repeated decisions and bookkeeping, then separate mechanical tasks from judgment that needs context.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
