prompt-engineer is one skill file that writes a prompt, or improves yours. It asks for the rules a prompt usually leaves out, such as how to verify a customer or who approves an exception, each with a default, then attacks its draft and fixes what breaks. I tested it against plain Claude, Opus 5.5 without the skill: it helps most when your request is short or the prompt will read input you do not trust. Its design comes from a few experiments, starting with five agents against one.
Repo: https://github.com/HighMarck17/prompt-engineer
How the design was chosen
Five agents to write one prompt is overkill, I know. Nobody would run that for real. It was a test: would more agents write a better prompt, or a worse one? I had a guess, but I wanted numbers.
The pipeline cost 3.81 to 5.52 USD and 15 to 23 minutes per prompt. On two requests, one run each, a blind model judge scored its prompts 7.75 out of 10. A single agent scored 7.83, a writer with a separate red-teamer 8.12. So five agents did no better than one, at a higher cost.
What I cut
- The red-teamer scored higher, but only on two requests. On 18 runs over prompts written by another agent, with 108 planted defects in all (30 distinct), an earlier development version without the red-teamer found every one, as the red-teamer did (the released skill: 107 of 108), and a prompt cost 0.42 to 0.50 USD without it instead of 0.73 to 0.78, and 92 to 96 s instead of 165 to 170. So it went.
- A researcher agent that searched the web for current guidance showed no gain: without it 35 of 51 sessions passed (69%), with it 65 of 101 (64%), not significant, at half the cost or less, and in less time.
Each cut made a prompt cheaper and faster, and none of these tests showed that what I cut was worth its cost.
What ships
One skill file. It asks a few short questions, each with a default, writes a short spec and a draft, then attacks the draft as the target model, a careless user and a hostile user would read it, and fixes what it finds. On average a prompt costs 0.39 to 0.45 USD and takes 71 to 104 s, questions included.
How it did
I ran two cases as agent sessions with a simulated user: a support bot with order tools and an Italian editing assistant. 17 task scenarios, 6 prompts per writer and case. Cost and time per prompt, on average:
| Writer | Sessions passed | Cost | Time |
|---|---|---|---|
| Claude, full notes, no questions | 66/102 (65%) | 0.17 to 0.19 USD | 54 to 63 s |
| Claude, one-line request, told to ask | 77/100 (77%) | 0.37 to 0.42 USD | 89 to 114 s |
| Skill, one-line request | 85/102 (83%) | 0.39 to 0.45 USD | 71 to 104 s |
With only a one-line request, the skill gained 18.6 points on Claude with the full notes, mostly because Claude asked customers for an unwanted second proof of identity, in 3 of the 17 scenarios. By scenario the skill did better in 8, worse in 4 and the same in 5: not significant.
Against Claude told to ask, the gain was 6.3 points, not significant, at almost the same cost: telling Claude to ask gets you close, though the skill's prompts had 29 to 38% fewer words.
On twelve kinds of request written by agents that never saw the skill, 83 inputs ran on two prompts per writer. From a one to three sentence request, 92% of the 166 runs passed against 49% for Claude not told to ask. From a complete one, 96% and 97%.
Limits
The judge is a model, the two agent cases also helped build the skill, and the owner answering the questions was simulated: a real one may answer less fully. Method and details: docs/evaluation.md.
When it helps
When your request is short and you do not know yet what to specify: 92% against 49% on the twelve kinds, better in 8 of 12 and worse in none. When the prompt reads input you do not trust: a planted instruction was obeyed in 0 of 41 sessions with its prompts, against 9 of 42 with Claude given the full notes and 5 of 40 with Claude told to ask, all on the editing assistant. A prompt cost less than 0.50 USD and took less than two minutes on average, plus the time you spend answering.
It adds little when your request is already complete (96% against 97%), or, on the task, when you tell Claude to ask first (not significant).
Claude Code:
/plugin marketplace add HighMarck17/prompt-engineer
/plugin install prompt-engineer@prompt-engineer
claude.ai: the zip is in the latest release. MIT.
Top comments (0)