What a skill costs you in context, every single turn
An active skill is re-read on every turn. Here is how to find what yours is costing, and how to split it so the reference material loads only when it is needed.
Skills are cheap to add and the cost is invisible, which is a bad combination. Nothing warns you. The session just gets shorter and the answers get vaguer, and by then you have four skills active and no idea which one is responsible.
The cost is per turn, not per session
An inactive skill costs you its description and nothing else. Once it activates, its body becomes part of what the model reads, and it stays there. Every turn. A six hundred line skill is six hundred lines re-read on your tenth message and your fiftieth.
That is the number worth knowing, because it is the one that competes with your actual work for room.
Most of a long skill is reference, not instruction
Open a skill that has grown past a few hundred lines and the split is usually obvious. A short section that tells the model what to do, and a long tail of things it might need to look up: option tables, error codes, provider lists, examples of formats.
The instructions have to be present. The lookup tables do not, until something asks for them.
| Stays in SKILL.md | Moves to references/ |
|---|---|
| What the skill does and when | Option and flag tables |
| The vocabulary that makes it trigger | Error code meanings |
| The order of operations | Long worked examples |
| What to refuse and why | Provider or platform specifics |
Do the measurement before the edit
Record what is loaded now, make one change, record it again. One change at a time, because two changes give you one number and no way to attribute it. A split that halves the always-loaded portion is a good day and it is entirely ordinary for a skill nobody has pruned.
- Measure what each active skill contributes before touching anything.
- Move the largest block of pure reference out first. It is usually most of the file.
- Keep every word the description relies on in the main file.
- Re-run your trigger prompts. A diet that breaks triggering has cost you more than it saved.
A saving you cannot show a number for is a rewrite, not an improvement.
What this buys you
Longer sessions before the context fills, and answers that stay specific because there is still room for the files you are actually working on. Neither shows up as a feature. Both are obvious the first afternoon you do not have to start a fresh session halfway through a task.
If you would rather not do the pruning by hand, Token Diet is the skill that does it, and it reports the saving per file rather than one total so a bad cut is visible.
Questions
Does an inactive skill cost me anything?
Only its description, which is a line or two. The body is loaded once the skill becomes active, and from then on it is part of what the model reads each turn.
What is a reasonable size for SKILL.md?
Small enough that you would not mind reading it on every turn, because that is what happens. A few hundred lines of instruction is ordinary. Several hundred lines of reference material is not, and it belongs in files the skill loads when it needs them.
Will splitting the file stop the skill working?
Only if you move something the description depends on. Split the reference material, keep the instructions and the triggering vocabulary in the main file, and re-run your trigger prompts afterwards to confirm nothing moved.
Keep reading
Or skip the afternoonand drop one in.
Every skill in the catalogue is a folder you drag into your skills directory. Tested on a clean machine, documented, with the trigger accuracy published.