Plant the intent. Grow the system.
SEED is a DSL-based brief compactor that compresses prose project intentions into dense token chains for LLMs — and generates .dna files for 3D visualization in FORGE.
A typical project brief sent to an LLM is 200+ words of prose, which:
- Wastes tokens (cost + latency)
- Carries ambiguity (synonyms, fluff, filler)
- Is hard to version (diff a paragraph vs diff a structured chain)
SEED transforms:
"j'aimerais un programme qui filtre mes mails par ordre de pertinence et les enregistre, puis m'alerte sur Slack" (~30 words)
into:
filtrer <mail> <pertinence> → enregistrer <db> & alerter <slack>
(~12 tokens — ~10× compression)
🌱 v0.1.0 — Engine ready (2026-04-25)
The engine library is complete and tested. Six units (TokenDB, Composer, Parser, Transpiler, Compressor, Storage) compile a CLI-grammar project intent into:
- A compressed DSL string (sent to LLMs — measured ≥ 2× compression on real prose)
- A canonical
.dnaJSON file (consumed by FORGE for 3D viz, separate plan)
Test coverage: 47 tests, all green (dotnet test).
Build:
dotnet build # builds Seed.Engine class library (.NET 8)
dotnet test # runs full test suiteNext: FORGE integration plan (Godot panel UI + 3D renderer adapter that consumes .dna files).
- Design :
docs/superpowers/specs/2026-04-25-seed-design.md - Plan :
docs/superpowers/plans/2026-04-25-seed-engine-plan.md - Visual library catalog :
docs/visual-library/icon-catalog.md
- Engine (
src/Seed.Engine/) — pure C# .NET 8 class library, no UI/host deps :TokenDb— embedded JSON registry, verb/modifier validationComposer— UI choices → DSL stringParser— DSL string → typed AST + errorsTranspiler— AST → canonical.dnaJSONCompressor—.dna→ minimal DSL for LLM (strips comments, normalizes whitespace)Storage— CRUD on.dnafiles (JSON file storage; SQLite alternative later if needed)
- Host A (palier 1) : module inside FORGE — composer UI + 3D rendering of
.dna(separate plan) - Host B (palier 2) : standalone web SPA via Blazor Wasm — same engine, browser host
Private repo for now. Open-source release planned for v2 under Apache 2.0 (open-core model — engine + spec free, FORGE integration + SaaS commercial).
bui1 (Simon Cantin) — designed in collaboration with Claude (Anthropic).
Architecture-first, AI-paired. Designed and shipped in a single April 2026 sprint with Claude (Opus 4.6) as paired implementation and audit partner. Each commit cross-audited: code review, dependency check, test-coverage gate. v0.1.0 ships with 47 passing tests across the six engine units (TokenDB, Composer, Parser, Transpiler, Compressor, Storage).