Back to projects
Active Started Feb 2026

Prompt Builder

Prompt Policy Engine. Classify, diagnose, rewrite, and score prompts calibrated to model tier and deployment context. Callable by agents, tools, humans.

Claude Code Skill 6-Part Stack

The Answer

Feed it a raw prompt and get back a structured rewrite, a 5-dimension score, the top three issues, tagged assumptions, and a temperature hint, all calibrated to the model tier and deployment you name.

The Problem It Solves

Prompts are rewritten by vibes. A prompt tuned for Opus backend use gets pasted into a Haiku mobile app and underperforms, but nobody can point to which dimension regressed. “Improve this prompt” returns a longer version that is not measurably better.

How It Works

Every call returns a fixed contract:

  • A 6-Part-Stack prompt: Role, Task, Constraints, Context, Output Format, Acceptance Criteria
  • A 5-dimension score: Accuracy, Clarity, Constraint Strength, Output Determinism, Completeness
  • Top 3 diagnosed issues in the original
  • Tagged assumptions ([ASSUMED: ...]) the caller should confirm
  • A runtime TEMPERATURE_HINT
  • Regression detection when iterating on a prior version

Calibration covers model tier (T1 frontier, T2 mid, T3 small or fast) and deployment (interactive, backend, rag_pipeline, agent, plugin, eval_judge, personal_mobile). The same prompt emerges differently for each pairing.

What Makes It Different

Scoring runs alongside rewriting, so iterations are measurable rather than vibes-based. Tagged assumptions keep the rewrite honest about what it invented. The shared contract makes the skill callable by other plugins, agents, or humans through the same interface.