Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Sungwoo Lee reports that thinking tokens made up 96% of the output from his custom Claude Code /his command across six runs in one project. That is a result from his own transcript analysis—not a general measure of Claude Code skills, and not evidence that his later redesign saved tokens. His redesign instead changed which work the model handled: scripts gathered and checked facts, while the model wrote down decisions and reasoning.
What Lee’s 96% figure measures
In an article dated October 1, 2026, Lee says his custom /his command produced 138,701 total output tokens over six runs in one project. Of those, 132,721 were thinking tokens, which he reports as 96% of the output. The command added about 1,000 tokens of history text 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)
The command appends a short session record to a project’s HISTORY.md. It is intended to preserve decisions and their reasons, approaches that were tried and abandoned, and a useful starting point for the next session. Lee’s figures describe that command in that project; they do not establish the typical token use of Claude Code commands or skills. The measurements are his report and were not independently verified.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Read Sungwoo Lee’s article on DEV Community.
How he counted tokens
Lee says his method relied on Claude Code session transcripts stored as JSONL. In his setup, assistant turns included a usage block. His script located each /his invocation and added usage from that command through the next user turn.
#1 Best Overall
He also found that a transcript could record the same API request more than once. To avoid double- or triple-counting, he deduplicated entries by requestId before summing them. That describes the method he says he used; it is not independently verified documentation of how every Claude Code transcript version records usage.
What changed in the redesigned command
Lee says the original command file had grown to 16.6 KB as rules were added after mistakes. He reduced it to 5.3 KB by moving predictable fact gathering and bookkeeping into two Python scripts.
his_prep.py <slug>: gather facts and prepare an entry
The preparation script collects facts such as changed files, commits, and Git state, then creates an entry skeleton with four empty sections. Lee says it stops near the compaction threshold or when a session is effectively empty just after /clear.
his_finish.py <slug>: validate and finish
The finishing script refuses to proceed if any of the four sections is blank. It then inserts and reorganizes entries and checks links, file size, and uncommitted changes.
Keep the model’s work focused on reasoning
After the change, the model’s writing task was to fill in four sections:
- Key decisions and why they were made.
- Alternatives rejected and why.
- Approaches that failed, or “none.”
- What the next session should do first.
The scripts captured file lists and commit hashes in a facts file, so the model no longer had to copy them into the history entry. Lee says he had previously tried to prefill plausible reasoning from a diff. Because that made an entry look complete without recording the actual reasoning, he rolled the change back. His division of labor is to let scripts supply facts and enforce bookkeeping, while the model that did the work explains its reasons.
Rank #4
Does the redesign prove Claude Code uses fewer tokens?
No. Lee says he has not repeated the six-run measurement after the redesign with the same rigor, so he does not report an “after” percentage. The supported result is a structural change: the old command asked the model to make many formatting and bookkeeping decisions; the revised workflow gives scripts those mechanical tasks and leaves the model four writing decisions.
There was no controlled comparison, independent replication, or confidence estimate in the reported measurement. It covers six invocations in one project, and does not establish whether the same pattern holds across other projects, commands, model versions, or usage patterns. The 96% figure should not be treated as a benchmark or as a measured saving from the redesign.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.When should a script do the work instead?
Lee’s example suggests a practical distinction for command and skill design: automate work that is deterministic and verifiable; reserve the model’s effort for interpretation that cannot safely be inferred from those facts. Collecting changed-file names or checking whether required sections are blank are mechanical tasks. Explaining why a decision was made or what trade-off mattered requires context from the work.
That distinction is a design principle, not a promise that moving a task into a script will reduce tokens by a particular amount. Lee summarizes the idea this way: “A long skill isn’t mainly a context cost. It’s the reasoning cost of making the same decisions again on every run.”
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.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
