First drafts that are credible.
Forge turns a one-line business idea into a typed, sectioned plan — then runs a separate critic agent over it to catch what an investor would catch.
Blank-page anxiety, not writing time.
Founders waste hours staring at section headings before they start writing. Existing templates are too long to use or too thin to be credible. Forge is built around a different assumption: the cost being attacked is the blank page, not the words.
“Generic LLM prompts produce ten sections that read like the model averaging across every business plan it has ever seen.”
Three engineering decisions, three visible features.
Schema-first, not prompt-first
Output is a typed BusinessPlan with 10 industry-standard sections — driven by OpenAI structured outputs with strict mode. No prose parsing, no formatting drift, no markdown surgery.
Two separate LLM calls
A drafter generates the plan. A separate critic — different prompt, different context, no knowledge of the drafter's reasoning — reviews it. The model can't defend what another conversation just wrote.
Tone via structure, not prompt
Every critique is tagged high / medium / low. The UI shows medium-and-above by default; low collapses under "+N nits". Critic stays sharp without being mean.
A typed plan, plus a sharp second opinion.
What's deliberately not here.
Each of these was decided once and locked. Worth seeing them written down — the gaps are part of the design.
Two LLM calls, not self-critique
A model defends what it just wrote. Self-critique is shallow. Two calls, two contexts.
No persistence, by design
No accounts, no DB, no saved drafts. Closing the tab loses the plan. Privacy via absence.
Severity field, not tone prompt
Tone-of-voice prompts produce either too-mean or encouraging-but-useless output. Structured severity does the job.
Schema before critic
Critic before structured outputs means rebuilding both. Build order matters.
FastAPI for orchestration
Frontend stays a UI, not an agent host. All LLM calls and schema validation live behind a single backend.
"Two agents" is a pipeline
Sequential LLM calls with different prompts and contexts. Not a tool-use agent loop. Honest framing matters.