Skip to content

README.md

Latest commit

 

History

History
145 lines (88 loc) · 5.65 KB

README.md

File metadata and controls

145 lines (88 loc) · 5.65 KB
Raw

Promptor v3 — A Claude Code Skill

A prompt architect, not a prompt suggestion box. Promptor runs every prompt request through a 5-circle validation pipeline fused with 18 optimization hacks, and delivers an auditable, copy-paste-ready result.

Stop hand-writing prompts by feel. Start generating them with a trace.

The Problem

Most "write me a prompt" requests get answered with something plausible-sounding but unverified — no domain risks identified, no success criteria defined, no check that the result actually does what you needed.

That's fine for a one-off. It's expensive when the prompt is going into production, gets reused across a team, or needs to survive contact with edge cases you haven't thought of yet.

How It Works

Ask Promptor to build, optimize, or audit a prompt and it runs a 3-phase pipeline:

Phase 1 — Five Circles (validation, with a structured trace)

Each circle emits a JSON trace block ({"circle": "C1", "status": "pass|fail", "evidence": "...", "hacks_applied": [...]}) before moving to the next:

Circle Name What it does
C1 STOP Validates the request, auto-detects domain and your experience level, identifies 3 domain-specific risks
C2 RESEARCH Cites 2–3 recognized patterns per risk — facts only, marked [UNVERIFIED] if unsourced
C3 GRID Generates a binary pass/fail success checklist — no subjective terms like "good" or "modern"
C4 TRIBUNAL Applies the grid strictly to your request — pass/fail table, zero free commentary, zero global score
C5 FIX Targeted corrections for every FAIL, up to 3 iterations, then delivers best-effort with an explicit [BLOCKED: reason] if still unresolved

Phase 2 — 18-Hack Filter

A library of 18 named optimization techniques (covering token cost, context hygiene, model routing, sub-agent limits, and more) gets prioritized based on what you're optimizing for — tokens, quality, speed, security, or collaboration — and woven into the generated prompt's instructions.

Phase 3 — Delivery (A-B-C-D)

  • A — Calibration: 3 bullets max on what was detected and applied
  • B — The prompt itself: copy-paste ready, with {{VARIABLE}} placeholders for reuse across contexts
  • C — Self-critique: a 0–5 score with one concrete improvement if it's not a 5
  • D — Follow-up: 2–3 questions to iterate further

Install

Option 1 — Git clone (recommended)

git clone https://github.com/valorisa/Claude-Skills ~/.claude/skills/

Then open Claude Code (claude in Terminal).

Option 2 — Manual

  1. Create folder ~/.claude/skills/promptor/
  2. Drop SKILL.md inside it
  3. Restart Claude Code

Use

Mention any of these triggers:

  • create a prompt
  • optimize this prompt
  • promptor
  • /promptor
  • generate a system prompt
  • prompt engineering
  • build me a prompt for...
  • reverse prompt engineer this
  • improve this prompt
  • prompt audit

Example:

promptor: I need a prompt for summarizing legal contracts into plain-language bullet points.

Promptor will:

  1. Ask two questions — what the prompt should do, and which AI tool it's for
  2. Run C1–C5: detect domain (legal/contracts), flag risks (e.g. liability language being oversimplified), research relevant summarization standards, build a pass/fail grid, evaluate against it, fix any failures
  3. Filter through the 18 hacks relevant to your focus area
  4. Deliver calibration notes, the finished copy-paste prompt, a self-critique score, and follow-up questions to refine further

Special Modes

[MODE:API] — add this to your request for strict JSON output (no markdown, no conversational delivery), useful for programmatic integration.

[?word] — drop this anywhere in your request to get an immediate inline explanation of that term before Promptor resumes the pipeline.

[COLLAB:MODE] — co-construct the prompt step by step instead of receiving it all at once.

When To Use It

Good Promptor candidates:

  • Building a reusable system prompt for a specific AI tool or workflow
  • Auditing an existing prompt that isn't performing as expected
  • Reverse-engineering what a prompt is actually doing so you can improve it
  • Any prompt going into production or shared across a team — where "looks about right" isn't enough

Skip Promptor for:

  • A single throwaway question you're about to ask directly
  • Quick one-off phrasing tweaks where the validation overhead isn't worth it

What Makes It Different

Every claim has a trace

The five circles don't just produce a prompt — they produce a JSON audit trail showing what was checked, what passed, what failed, and which hacks were applied at each stage. You can see why the prompt looks the way it does, not just the final text.

No subjective success criteria

Phase 1's grid (C3) explicitly bans vague terms like "good" or "modern" as pass/fail criteria. Every criterion has to be checkable, which forces the evaluation in C4 to be a real test rather than a vibe check.

Self-critique is mandatory, not optional

Every delivery includes a 0–5 self-assessment. A score under 5 comes with a named, concrete improvement — not a vague "could be better."

Built-in instruction-injection sanitization

Before processing your request, Promptor checks it (and any supplied context) for injection patterns and flags them rather than executing blindly.

Credit

  • Implemented as a Claude Code skill by @valorisa
  • See also promptor-council for a multi-perspective deliberation variant of this same methodology

License

MIT — do whatever you want with it.