{"slug": "i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well", "title": "I tried five agents to write one prompt. One agent did as well.", "summary": "A developer released prompt-engineer, an MIT-licensed Claude skill that asks a few short questions about verification, approvals and injection handling, then writes a prompt spec and draft and red-teams it as the target model, a careless user and a hostile user would read it. In tests on a support bot and an Italian editing assistant across 17 task scenarios, the skill's prompts passed 83% of sessions from a one-line request versus 65% for Claude with full notes, a gain of 18.6 points, while a five-agent pipeline scored 7.75/10 on a blind model judge against 7.83 for a single agent and cost $3.81 to $5.52 per prompt. No skill prompt obeyed a planted instruction, whereas Claude's prompts did in 9 of 48 and 5 of 46 sessions.", "body_md": "Nothing fancy: it is just one skill file that writes a prompt, or improves one, and attacks it before I use it. But I used it as an excuse to run a few experiments of my own on AI-assisted prompt writing. The main result: five agents did not write better prompts than one.\n\nThe rules a prompt leaves out break it more often than its wording: how to verify a customer, who approves an exception, what to do when the input says to ignore the instructions. So the tool asks about them before it writes.\n\nFive 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.\n\nThe pipeline cost 3.81 to 5.52 USD and 15 to 23 minutes per prompt. 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 came out a little worse than one, and cost more.\n\nOne 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. A prompt costs 0.39 to 0.45 USD and 71 to 104 s, questions included.\n\nI tested the skill against Claude Opus 5.5 without it on two cases, run 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. Per prompt:\n\n| Writer | Sessions passed | Cost | Time | \n|---|---|---|---|\n| Claude, full notes, no questions | 66/102 (65%) | 0.17 to 0.19 USD | 54 to 63 s | \n| Claude, one-line request, told to ask | 77/100 (77%) | 0.37 to 0.42 USD | 89 to 114 s | \n| Skill, one-line request | 85/102 (83%) | 0.39 to 0.45 USD | 71 to 104 s | \n\nWith only a one-line request, the skill gained 18.6 points on Claude with the full notes (95% interval 6.9 to 30.4), mostly because Claude asked customers for an unwanted second proof of identity. Against Claude told to ask, the gain was 6.3 points, not significant.\n\nNo skill prompt let the agent obey a planted instruction. Claude's prompts did in 9 of 48 and 5 of 46 sessions.\n\nOn twelve kinds of request written by agents that never saw the skill, from a one to three sentence request, 92% of inputs passed against 49% for Claude not told to ask. From a complete one, 96% and 97%.\n\nThe 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. The five-agent designs ran once per request. Method and details: [docs/evaluation.md](https://github.com/HighMarck17/prompt-engineer/blob/main/docs/evaluation.md).\n\nSo, after all the cutting: the skill works. From a short request it writes a prompt that holds up better than the ones Claude writes on its own, and from a complete request it does as well as Claude. You answer a few questions and get a prompt ready to use in under two minutes, for less than half a dollar at API prices.\n\nClaude Code:\n\n```\n/plugin marketplace add HighMarck17/prompt-engineer\n/plugin install prompt-engineer@prompt-engineer\n```\n\nclaude.ai: the zip is in the [latest release](https://github.com/HighMarck17/prompt-engineer/releases/latest). MIT.", "url": "https://wpnews.pro/news/i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well", "canonical_source": "https://dev.to/highmarck/i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well-i4m", "published_at": "2026-11-05 09:08:43+00:00", "updated_at": "2026-11-05 09:19:36.916788+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "large-language-models", "ai-safety"], "entities": ["Claude", "Claude Opus 5.5", "prompt-engineer", "HighMarck17", "Claude Code", "Anthropic"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well", "markdown": "https://wpnews.pro/news/i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well.md", "text": "https://wpnews.pro/news/i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well.txt", "jsonld": "https://wpnews.pro/news/i-tried-five-agents-to-write-one-prompt-one-agent-did-as-well.jsonld"}}