A two-stage pipeline.
Senior-engineer output.
The first model finds the gaps in your task. The second model writes the enterprise meta prompt. Both are tuned to behave like a senior prompt engineer who has shipped thousands of production prompts to FAANG-tier AI teams.
Prompts are infrastructure, not chat.
Most prompts are first drafts pretending to be production code.
A vague sentence handed to a frontier model returns a confident answer that drifts the moment your inputs change. There's no role, no schema, no constraints, no eval rubric — nothing you can diff, version, or hand to a teammate. The result is unshippable.
Senior PE teams treat prompts like code, not chat.
Anthropic, OpenAI, Google, IBM, and Stripe all publish the same recipe: structure the prompt against a known framework (CO-STAR, RISEN, XML, Developer Message, CIDI), separate role from context from constraints from output schema, and version the artifact. The Forge automates that recipe.
One sentence in. A 3,000-word, framework-structured prompt out.
You write what you want. The Forge picks the right framework for the model, asks the surgical questions only a senior prompt engineer would ask, then ships a complete meta prompt — role, context, instructions, examples, constraints, output schema, edge cases, and a self-evaluation rubric. Drop it into git. Diff it. Own it.
Four stages. Each runs at senior quality.
Analyzer
A senior-prompt-engineer model reads your one-sentence task, maps it onto the slots of the framework you chose, and identifies which slots are empty or ambiguous. It returns up to six closed-form clarifying questions — never busywork, only the questions that materially change the final prompt.
Clarifier
You answer the questions in a clean form. Multiple-choice where the answer space is small, free-text where it's not. Each answer becomes a locked slot — the Generator no longer has to guess.
Generator
A second senior-prompt-engineer model produces the full enterprise meta prompt: framework-structured, dense, with role definition, context, instructions, few-shot examples, hard constraints, soft preferences, an output schema, edge-case rules, a self-evaluation rubric, and a refusal policy.
Output
Copy, download, or hand the prompt directly to GPT-5, Claude Sonnet 4.6, or Gemini 3 Pro. The 'How to use' block at the bottom tells you the recommended model, temperature, and call pattern.
The four rules behind every Forge output.
Frameworks are not optional.
Every output is structured to one of CO-STAR, RISEN, Anthropic XML, OpenAI Developer Message, or Google CIDI. No free-form 'best effort' prompts.
The model knows the model.
Anthropic XML for Claude. Developer Message for GPT-5. CIDI for Gemini. The Forge picks the framework that the chosen model was trained to honor.
Two stages beat one.
A single LLM call can't both interrogate your task and produce a 3,000-word artifact. Splitting into Analyzer + Generator lets each model do one job at senior-engineer quality.
Audit-ready by default.
Role, constraints, output schema, and an eval rubric ship in every prompt. Drop into git. Diff it. Review it like code.