Transform novice prompts into expert-framed requests with logical rigor.
An AI agent skill that bridges the gap between how non-specialists phrase requests and how domain experts would frame the same request. Unlike simple prompt rewriters, prompt-forge adds:
- Pre-transformation audit — catches flawed assumptions, loaded questions, and false dichotomies before rewriting
- Prompt classification router — applies the right expert patterns based on prompt type (informational, creative, analytical, action, decision)
- Hybrid mode selection — automatically chooses between single-rewrite (simple prompts) and interactive (complex prompts with ambiguity)
- Post-transformation verification — validates intent preservation, no invention, recognition, and proportionality
- Reasoning integration — embeds structured reasoning cues into decision/analysis prompts
sequenceDiagram
accTitle: Prompt Forge Pipeline
accDescr: Sequence diagram showing the six-step transformation pipeline from novice prompt to expert-framed request, with mode selection branching and verification loop.
participant U as 👤 User
participant A as 🔍 Audit
participant R as 🗂️ Router
participant M as ⚖️ Mode
participant T as 🔧 Transform
participant V as ✅ Verify
U->>A: Submit novice prompt
A->>A: Run 4 checks<br/>(assumptions, intent,<br/>framing, scope)
A->>R: Pass audit results
R->>R: Classify prompt type<br/>(informational, creative,<br/>analytical, action, decision)
R->>M: Pass classification
M->>M: Evaluate 3 signals<br/>(classification, audit, structure)
alt < 2 complex signals
M->>T: Automatic mode<br/>(single rewrite)
else ≥ 2 complex signals
M-->>U: Interactive mode<br/>2-3 options + rationale
U->>T: Select option
end
T->>T: Apply expert patterns<br/>+ reasoning cues (if decision/analysis)
T->>V: Submit rewrite
loop Verification
V->>V: Run 4 tests<br/>(intent, invention,<br/>recognition, proportionality)
alt Fail
V->>T: Return for revision
else Pass
V->>U: Deliver expert rewrite
end
end
| Feature | prompt-enhancer | prompt-forge |
|---|---|---|
| Mode | Always automatic | Hybrid (auto + interactive) |
| Logical audit | None | Pre-transformation checklist |
| Fallacy detection | None | 6 common prompt fallacies |
| Classification | None | 5-type router |
| Verification | None | 4-test post-check |
| Reasoning cues | None | Conditional (decision prompts) |
| Token footprint | ~169 lines | ~196 lines (more features, less text) |
npx skills add hyuce/prompt-forgeThis installs the skill to your agent's skills directory (Claude Code, Cursor, Gemini CLI, GitHub Copilot, OpenCode, etc.).
Global install (all projects):
npx skills add hyuce/prompt-forge --globalSpecific agent:
npx skills add hyuce/prompt-forge --agent cursorThis skill contains only five files: SKILL.md (the methodology), examples.md (transformation examples), test-results.md (RED/GREEN/REFACTOR evidence), README.md (this file), and LICENSE (MIT). All are plain markdown or text. No executable code, no scripts, no network calls, no obfuscation, no data exfiltration, no prompt-injection patterns. The npx skills add command copies these files to your agent's skills directory; the agent reads them as instructions, nothing more.
Independent audits:
- Gen Agent Trust Hub — Pass / Safe. Verified the skill is a prompt-based methodology with no executable content.
- Snyk — Pass / Low Risk (last audited Jun 20, 2026). The initial E005 flag (triggered by the bare
npx skills addinstall command pattern) was downgraded to no issues detected after the README gained a Security section — Snyk's content analysis found sufficient evidence that the skill contains no executable code.
Invoke when:
- User asks to improve, refine, or rewrite a prompt
- User needs help framing a request for an AI system
- A prompt feels vague, overloaded, or logically unsound
- Load the skill:
@prompt-forgeor readSKILL.md - Provide the prompt: paste the novice prompt you want to improve
- Let the pipeline run:
- Pre-audit checks for logical flaws (false dichotomies, loaded questions, scope creep)
- Classification router identifies the prompt type (informational, creative, analytical, action, decision)
- Mode selection decides: automatic (single rewrite) or interactive (2-3 options)
- Receive the output:
- Automatic mode: single expert rewrite + "what changed" explanation
- Interactive mode: rationale for interactive mode, 2-3 framing options with trade-offs, you select one, then receive the full rewrite
You: "@prompt-forge improve this prompt: 'Should I use PostgreSQL or MongoDB?'"
Skill:
Processing in interactive mode because: false dichotomy detected (only two options presented when more may exist) + decision-type prompt.
Option A: Compare PostgreSQL vs MongoDB directly — faster if you've already eliminated other options. Option B: Evaluate PostgreSQL, MongoDB, and other viable alternatives — broader but may surface options you haven't considered. Option C: Start from your requirements first, then match to databases — most thorough, avoids anchoring on any specific technology.
Which approach fits your intent?
You: "B"
Skill:
Expert rewrite: "Evaluate database options for my use case. Assess PostgreSQL, MongoDB, and any other viable alternatives against: data model fit (relational vs document), scalability requirements, team expertise, operational complexity, and total cost of ownership. For each option, list what supports it and what contradicts it. Recommend based on my specific constraints — I'll provide those."
What changed: False dichotomy expanded. Decision framework named. Reasoning cues embedded. Constraints deferred to user.
If the pre-audit flags your prompt as flawed, that's the skill working. The friction prevents the kind of amplification that produces expert-sounding wrong answers. See the Rationalization Table in SKILL.md for the excuses the skill is designed to resist.
prompt-forge/
├── SKILL.md # Core skill logic (~196 lines)
├── examples.md # 12 transformation examples (cross-reference)
├── test-results.md # RED/GREEN/REFACTOR evidence
├── README.md # This file
└── LICENSE # MIT
MIT — see LICENSE.