Luthier

A luthier builds the instrument you then play.

Luthier takes a product requirements document and produces a scaffolded, compilable Go project wired with axon components. You describe what you want to build. Luthier figures out the architecture, selects modules, writes the glue code, and produces a step-by-step implementation plan.

Install

just install

Usage

luthier my-prd.md

Luthier reads your PRD, calls an LLM to analyse the requirements, asks clarifying questions if anything is ambiguous, then writes a project directory with:

• cmd/<name>/main.go : glue code wiring selected axon modules together
• go.mod : dependencies with local replace directives
• AGENTS.md : architecture decisions, module selections, det/non-det boundaries
• CLAUDE.md : working instructions for Claude Code
• plans/<date>-initial-build.md : commit-sized implementation steps
• justfile : build, test, install targets

The generated project compiles immediately. Business logic is stubbed; the plan tells you (or Claude Code via /iterate) what to build next.

How it works

1. Analyse: an LLM reads your PRD alongside the axon module catalog and calls tools (select_module, define_boundary, add_plan_step, raise_gap, finalize) to build a scaffold spec incrementally
2. Resolve gaps: if the spec contains ambiguities, luthier asks you to clarify
3. Compose: snippets for each selected module are topologically sorted by dependency and composed into a working main.go
4. Write: templates render the scaffold deterministically from the spec
5. Verify: go build ./... confirms the scaffold compiles

Module catalog

Each axon module owns a luthier.yaml manifest declaring its purpose, when to use it, and its scaffold snippet. Run just sync-catalog to regenerate the catalog from the module manifests in the lamina workspace.

just sync-catalog    # regenerate system prompt + snippets from luthier.yaml files

Eval

Luthier includes an eval tool that runs the analysis N times against a PRD and scores consistency across runs.

just build
bin/luthier-eval fixtures/coding-agent-prd.md 5           # Sonnet (default)
LUTHIER_MODEL=claude-opus-4-6 bin/luthier-eval fixtures/coding-agent-prd.md 5  # Opus
LUTHIER_PROVIDER=local bin/luthier-eval fixtures/coding-agent-prd.md 3         # local llama-server

Scores: module selection consistency, boundary count stability, gap detection stability, plan step count stability.

Guiding principle

If the outcome can be determined from the set of inputs, write code to get that outcome. Only establish a conversational loop when the output is genuinely non-deterministic.

The module catalog and wiring patterns are known inputs, so snippet composition is deterministic. The LLM is only involved where genuine reasoning is needed: which modules fit this PRD, where the det/non-det boundaries are, and what the implementation plan should be.