Prompt Optimizer /prompt-optimizer

About this skill

Prompt Optimizer treats your system prompt like what it is: code that deserves tests. Give it your prompt and an eval suite, and it iteratively rewrites, A/B-tests variants, and keeps the winners — stopping when scores plateau or your budget runs out.

Every iteration is logged: what changed, why the optimizer thought it would help, and the before/after eval scores. You get the final prompt plus the full experiment history, so the improvement is auditable rather than vibes-based.

It's pre-1.0 and honest about it — best on classification, extraction, and structured-output tasks today, with multi-step agent prompts improving every release. The eval-suite requirement is deliberate: no evals, no optimization, no cargo-cult prompt tweaks.

Inputs

• system_prompt · string
  The system prompt to optimize.
• eval_suite · string
  Path to eval cases with expected outputs.
• budget · string
  Max iterations or API spend cap.

Outputs

• optimized_prompt · string
  The winning system prompt.
• score_delta · string
  Before/after eval scores.
• experiment_log · string
  Full iteration history with rationale.

Permissions

Deny-by-default: the skill can only use what it declares.

Manifest

{
  "spec": "agentskills.io/v1",
  "name": "prompt-optimizer",
  "version": "0.9.7",
  "description": "Takes your system prompt and eval suite, then iteratively rewrites and A/B-tests the prompt until scores stop improving.",
  "inputs": {
    "system_prompt": {"type": "string", "description": "The system prompt to optimize.", "required": true},
    "eval_suite": {"type": "string", "description": "Path to eval cases with expected outputs.", "required": true},
    "budget": {"type": "string", "description": "Max iterations or API spend cap.", "required": false}
  },
  "outputs": {
    "optimized_prompt": {"type": "string", "description": "The winning system prompt."},
    "score_delta": {"type": "string", "description": "Before/after eval scores."},
    "experiment_log": {"type": "string", "description": "Full iteration history with rationale."}
  },
  "permissions": ["fs:read", "fs:write", "secret:read"],
  "author": "Latent Forge",
  "category": "AI Engineering"
}

Changelog