Skill or system prompt?
Split a 9,000-token system prompt into what stays, what becomes a skill and what becomes a hook, with the break-even per block and the saving per day.
On this page
A system prompt that every team appends to is a shared file with a bill attached. Splitting it has an answer per block, and the arithmetic settles most of them.
The situation
An internal engineering assistant answers 5,000 requests a day, one model call each. Its system prompt has grown to 9,000 tokens:
| Block | Tokens | Share of requests that need it |
|---|---|---|
| Role, output format, tone | 2,000 | all |
| "Never commit secrets; never push to main" | 150 | any request touching git |
| Release-notes procedure | 900 | 3% |
| Migration-review procedure | 1,400 | 4% |
| Incident write-up | 1,100 | 1% |
| Code-review checklist | 1,600 | 45% |
| Dependency-upgrade procedure | 800 | 5% |
| Frontend style guide | 1,050 | 15% |
Assume a skill's name and description cost 50 tokens on every request, and that loading a body adds one round trip which re-reads about 6,000 tokens of context. P_in is $3 per million input tokens, a cache read is C_r = 0.1 × P_in, and a month is 30 days. The prompt is not cached today: its first line is You are helping {user_name}.
The question
Lay out the split — what stays in the prompt, what becomes a skill, what becomes something else — with the break-even for each candidate and the net daily saving, then say what fixing the cache does to the case.
A strong answer · Rubric · Where answers go wrong
Write your answer first, then open