forgeplan generate

Generates a fully-drafted artifact from a natural language description using an LLM (Gemini / OpenAI / Anthropic, configured in .forgeplan/config.yaml). Unlike forgeplan new, which produces an empty template you fill manually, generate asks the model to write Problem, Goals, Non-Goals, Functional Requirements, and the rest of the MUST sections in one pass. The result is a draft you still review — but the 30-minute cold-start of staring at an empty template disappears.

When to use

You know what you want and would rather review an LLM draft than type the boilerplate yourself.
Prototype artifact for discussion: generate a throw-away PRD to explore a half-formed idea with the team before committing to the real one.
Brown-field discovery: describe an existing subsystem and ask the model to reverse-engineer a PRD / Spec / RFC for documentation.
You have a clear one-liner (“add rate limiting to /api/v1/”) and want all the sections filled consistently.
Fast iteration during Shape phase when you want to try 2–3 framings quickly.

When NOT to use

You don’t have LLM credentials configured — run forgeplan config set llm.provider ... first, or fall back to forgeplan new.
The decision is tactical and doesn’t need an artifact at all — just commit.
You want full manual control over wording — use forgeplan new and fill sections yourself.
You’re logging a decision from a conversation — use forgeplan capture.
The content already exists in a saved memory — use forgeplan promote.

Usage

forgeplan generate <KIND> <DESCRIPTION>

Arguments

  <KIND>         Artifact kind: prd, epic, spec, rfc, adr, problem, solution, evidence
  <DESCRIPTION>  Natural language description of what to generate

Options

  -h, --help     Print help
  -V, --version  Print version

generate reads your active LLM provider from .forgeplan/config.yaml. If no provider is set or the API key is missing, the command fails early with a config hint instead of falling back to an empty stub.

Examples

Example 1: Generate a PRD from a one-liner

forgeplan generate prd "add rate limiting to /api/v1/ endpoints"

Produces .forgeplan/prds/prd-NNN-add-rate-limiting.md with Problem (“public endpoints are exposed to abuse…”), Goals, Non-Goals, Target Users, and an FR list drafted by the model. Open the file, read critically, tighten the wording, then run forgeplan validate PRD-NNN to confirm MUST rules pass.

Example 2: Reverse-engineer an RFC for an existing subsystem

forgeplan generate rfc "current embedding pipeline: fastembed BGE-M3, 1024 dims, batch 32, cached in .forgeplan/.fastembed_cache/"

Good for brown-field documentation — the LLM drafts an RFC describing the as-built architecture with Implementation Phases already checked off. Use this as a starting point, then correct any hallucinated details against the real code.

Example 3: Draft a ProblemCard for a fresh signal

forgeplan generate problem "users report search returns stale results after renaming artifacts"

Generates Signal / Context / Goals / Anti-Goodhart indicators sections. ProblemCards don’t require a validation gate, so you can activate as soon as the card is coherent and link follow-up Evidence or Solutions.

How it fits the workflow

route  →  generate  →  (review + edit)  →  validate  →  reason  →  code  →  evidence  →  review  →  activate

generate slots into the same Shape phase as new, but collapses “create stub” and “fill MUST sections” into one LLM call. The rest of the pipeline is unchanged: you still need validate to pass, reason for Deep/Critical depth, and Evidence before activate. Treat the generated text as a first draft — the model will not catch domain-specific constraints the way a human operator will.