Prompt Forge

Turn vague assistant ideas into structured, high-performance system prompts — portable across Claude, GPT, and Gemini. [EXPLICIT]

Operation Modes

Mode	Trigger	Action
Create	"Create a prompt for...", "I need an assistant that..."	Full forge cycle: capture, draft, evaluate, refine, deliver
Review	"Review this prompt", "Is this any good?"	Evaluate against rubric, deliver scorecard + prioritized fixes
Evolve	"Make this better", "Optimize this prompt"	Identify weaknesses, apply targeted improvements
Repair	"This isn't working", "The AI keeps doing X"	Diagnose failure pattern, apply surgical fix
Port	"Convert this for Claude/GPT/Gemini"	Adapt format, constraints, and features for target platform

Default behavior: Generate first, confirm after. Produce a strong v1, then iterate on feedback. Do not ask 20 questions before writing.

Core Design Principles

1. Hybrid Role — Composite expert identity (domain + methodology + communication style). Not "You are a marketing expert" but a specific archetype with boundaries. [EXPLICIT]
2. Conversational Flow — Structure interaction in phases (discovery, execution, delivery) with entry/exit criteria. [EXPLICIT]
3. Deliverable-Oriented — Define what the user walks away with. Concrete output format, not vague "help with X". [EXPLICIT]
4. Dynamic Intelligence — Build reasoning techniques (chain-of-thought, structured analysis) into the prompt natively. [EXPLICIT]
5. Executive Tone — Professional, decisive. The AI sounds like a senior consultant, not a cautious intern. [EXPLICIT]
6. Complete Structure — Use the Playbook format as canonical output. [EXPLICIT]
7. Iterative Co-Design — Build self-correction triggers into the prompt. [EXPLICIT]

For deep explanations and examples, read references/design-principles.md. [EXPLICIT]

The Playbook Format

Universal output template for generated prompts:

# [Assistant Name] — v[X.Y]

## Role & Archetype
[Composite expert identity: domain + methodology + communication style]

## Objective
[What the user achieves — 1-2 sentences]

## Parameters
- Model: [Target model(s)]
- Temperature: [Recommended setting]
- Context window usage: [Strategy]

## Interaction Flow
### Phase 1: Discovery
[How the assistant gathers context]

### Phase 2: Execution
[How it processes and produces output]

### Phase 3: Delivery
[How it formats and presents results]

## Constraints
[Hard boundaries — what the assistant must NOT do]

## Key Questions
[3-5 questions for ambiguous context]

## Output Template
[Exact format with placeholders]

## Self-Correction Triggers
[Patterns that signal recalibration]

For per-section guidance, read references/playbook-template.md. [EXPLICIT]

Evaluation Rubric

Every prompt scored 1-10 on each dimension. Target: 8+ on all for production quality. [EXPLICIT]

#	Criterion	Measures
1	Foundation	Clear archetype, objective, and constraints?
2	Accuracy	All claims, frameworks, and techniques correct?
3	Quality	Writing professional, precise, filler-free?
4	Density	Maximum value per token?
5	Simplicity	Non-expert could understand structure?
6	Clarity	Instructions unambiguous? One interpretation?
7	Precision	Constraints specific enough to enforce?
8	Depth	Handles edge cases, failures, advanced scenarios?
9	Coherence	All sections reinforce each other?
10	Value	User gets meaningfully better results?

For detailed scoring and repair protocols, read references/evaluation-rubric.md. [EXPLICIT]

Context Engineering

Modern prompt design manages everything the model sees, not just instruction text. [EXPLICIT]

Layer	Scope	Example
L1: Hot	System prompt + current turn	The Playbook itself
L2: Warm	Uploaded docs, knowledge base	Reference PDFs, style guides
L3: Cold	RAG retrieval, tool outputs	Dynamic data from APIs

Design for all three layers. A great L1 with poor L2 design underperforms. [EXPLICIT]

For context hierarchy patterns and token optimization, read references/context-engineering.md. [EXPLICIT]

Multi-Platform Deployment

Platform	System Prompt	Knowledge Base	Tools
Claude Projects	Project instructions	Project knowledge files	MCP servers
ChatGPT Custom GPTs	Instructions field	Uploaded files	Actions (API)
Gemini Gems	System instructions	Google Docs/Drive	Extensions
API / Code	system parameter	RAG pipeline	Function calling

For platform-specific limits and deployment guides, read references/platform-guides.md. [EXPLICIT]

Workflow by Mode

Create (most common)

1. Capture — domain, user outcome, target platform. [EXPLICIT]
2. Draft — fill every Playbook section. Generate first, confirm after. [EXPLICIT]
3. Evaluate — score against rubric. Fix anything below 8. [EXPLICIT]
4. Deliver — format for target platform with deployment instructions. [EXPLICIT]

Review

1. Read the prompt completely. [EXPLICIT]
2. Score against rubric. [EXPLICIT]
3. Deliver scorecard + prioritized fix list. [EXPLICIT]

Evolve

1. Identify weakest rubric dimensions. [EXPLICIT]
2. Apply targeted improvements (density compression, constraint sharpening, flow restructuring). [EXPLICIT]
3. Re-evaluate to confirm improvement. [EXPLICIT]

Repair

1. Diagnose the failure pattern. [EXPLICIT]
2. Apply surgical fix to the specific section causing issues. [EXPLICIT]
3. Add a self-correction trigger to prevent recurrence. [EXPLICIT]

Port

1. Identify source platform features and constraints. [EXPLICIT]
2. Map to target platform equivalents. [EXPLICIT]
3. Adapt format, adjust for token limits, replace unsupported features. [EXPLICIT]

Antipatterns

Problem	Bad Pattern	Fix
Wall of text	2000-word flat instruction	Break into Playbook sections with flow
Vague role	"You are helpful"	Composite archetype: domain + method + style
No output format	"Help users with X"	Define exact deliverable template
Over-constraining	50 rules in ALL CAPS	Explain why behind each constraint
Platform-blind	Same text everywhere	Adapt to each platform's affordances
No feedback loop	Static, never improved	Build self-correction triggers + versioning

Assumptions & Limits

• Prompt quality is platform-dependent. A perfect Claude Project prompt may underperform on GPT without adaptation.
• The Playbook format is a starting point, not a cage. Some domains need additional sections; others need fewer.
• Evaluation scores are relative to the prompt's purpose. A chatbot prompt and a data analysis prompt have different ceilings on "Depth."
• Context window limits vary by platform and model. Always check current limits before optimizing token usage.
• Self-correction triggers work best in models with strong instruction-following. On weaker models, constraints may be ignored.

Edge Cases

• Prompt for a non-conversational use case (batch processing, API-only): Omit Interaction Flow. Focus on Constraints and Output Template.
• User provides a prompt in a language other than English: Write the Playbook in the user's language. Criterion names in the rubric remain universal.
• Extremely long knowledge base (100+ pages): Design for L2/L3 context. The system prompt should reference, not contain, the knowledge.
• User wants a prompt that breaks platform rules: Decline. Note the specific platform policy it would violate.
• Evolve mode finds no weaknesses: Possible for well-crafted prompts. Report that all criteria meet threshold, suggest testing with adversarial inputs.
• Porting between very different platforms (e.g., Claude to a fine-tuned open model): Some features (tool use, retrieval) may not translate. Document what was lost in porting.

Validation Gate

Before delivering a prompt, confirm:

• Playbook format is complete (all mandatory sections present)
• Role is a composite archetype, not a generic label
• At least one constraint defines what the assistant must NOT do
• Output template includes concrete format with placeholders
• All rubric criteria score 8+ (or deficits are documented)
• Platform-specific formatting applied if a target was specified
• Self-correction triggers are present and testable

Reference Files

• references/design-principles.md — Deep dive into the 7 principles with examples and antipatterns
• references/evaluation-rubric.md — Full scoring parameters, failure criteria, repair protocols
• references/playbook-template.md — Complete template with per-section guidance
• references/platform-guides.md — Platform-specific formatting, limits, deployment for Claude, GPT, Gemini, API
• references/context-engineering.md — Context hierarchy, token optimization, RAG integration

---

Author: Javier Montaño | Last updated: March 12, 2026

Usage

Example invocations:

• "/prompt-forge" — Run the full prompt forge workflow
• "prompt forge on this project" — Apply to current context