This is the third post in a series that started with vibe guiding and continued with an efficiency audit. Two months ago I laid out the four-skill loop: /start-issue writes a /scratchpad, /tackle-scratchpad-block executes one step at a time, status flags transition from pending to done, and /finish-issue reads what shipped and writes the PR title + description.
Why I changed my mind
The driver was token consumption. My employer gave each team dashboards of our AI tool usage and I was always the top consumer on mine — fairly high on the department chart too. That's when I started to suspect it was not what I was building but how I was working with the tools.
/scratchpad's JSON step block and /tackle-scratchpad-block chain meant every task paid the same ceremony tax: status flags to transition, JSON to validate, a fenced block to parse. Useful scaffolding for multi-hour work. Overhead I was paying on every session. If my process was the reason I topped the dashboard, revisiting it was the cheapest experiment I could run.
Once I started looking, the redundancy wasn't only scratchpad-vs-note. The templates the composite skills were generating — what /start-issue, /start-side-quest, /tackle-pr-comment, and /finish-issue wrote into their output documents — carried weight they didn't need. A few examples from the audit that led to PR 130:
