AI-assisted idea-to-system pipeline tool — walks a rough idea through Hidden Alchemy's own stages (spec drafting, feasibility, architecture sketching, system scaffolding), with a human decision at every transition.
Status: concept — the pipeline Promise early research is done and the go
decision is recorded; no runnable software exists yet. This status is earned
upgrade, not assumed. Last updated: 2026-09-07.
A flagship project of the Hidden Alchemy
laboratory — the org pipeline (IDEA → CONCEPT → ARCHITECTURE → SYSTEM → AUTOMATION → REALITY) made concrete end-to-end. The org builds, runs, and
demonstrates the very workflow this tool automates.
- Pipeline stage: currently CONCEPT — it has passed Research (go/no-go recorded in ideas#3) and is awaiting its Architecture note (project board: Idea Lifecycle).
The work today is honest and small: a project repository, a written architecture, and an evaluation harness — not a mocked-up demo. The repository name and identity are fixed; the scope below is the working contract.
This is:
- A CLI-first staged pipeline: each org stage produces a concrete artifact (spec draft, feasibility note, architecture sketch, scaffolded project).
- An offline evaluation harness that scores generated artifacts for quality, so the tool's output is graded with receipts, not vibes.
- Provider-agnostic prompt and template files — no hard dependency on any single model vendor.
- A public, auditable paper trail: every artifact it produces is a text file the tool leaves for a human to inspect.
This is NOT:
- A GUI-first product in v0. If a web UI arrives it is after the CLI proves the pipeline — not before.
- An auto-shipping code generator that deploys without a human. Every go/no-go decision in the pipeline stays human (org rule §21).
- A way to replace maintainers, a chatbot wrapper course, or a storehouse for API keys and org secrets (no secrets live in this repository).
Four layers, each one artifact-shaped:
- Stages (
idea.md → concept.md → architecture.md → system/) — a directory walk where each stage reads the previous artifact and emits the next via a prompt chain. - Templates (
templates/*.jinja,templates/*.md) — spec and feasibility skeletons matching the org's own §15 design system. - Eval harness (
eval/) — offline scoring of generated artifacts against rubrics; raw outputs published, never hidden. - Human gate — a review step between every stage transition; nothing moves forward without a recorded decision (mirrors §23's go/no-go).
The prompt chain is the core product; the eval harness is what keeps it honest.
No runnable software exists yet (status: concept). The first command a user
will run once v0 ships is:
idea-forge scaffold ./my-idea.md # emits ./stages/concept.md + feasibility noteUntil then, "how it works" is documented above and tracked on the project board; nothing here pretends to be installable.
Follow the org-wide CONTRIBUTING.md.
Small documentation fixes need no pre-approval; code changes start by
commenting intent on a good first issue or opening a draft PR. Strong
first-contribution surfaces now:
- Eval-harness fixtures — sample idea→concept pairs with hand-scored labels for grading tests.
- Template drafts — spec/feasibility skeletons written against the org's §15 design system.
- Prompt-engineering evals — grading chains against the rubric.
Live work is tracked on the Idea Lifecycle board and the originating idea issue.
- Research complete (GO recorded) — held in ideas#3.
- Next: Architecture note for v0 (CLI, staged artifacts, eval harness).
- Then: v0 CLI + template set; eval-harness fixtures with human-scored labels.
- Then: Prototype status, evaluated against the rubric before anything is
called
active. No dates are promised; progress is visible on the board, honestly.
- @Hassan0703 — primary accountable maintainer
MIT — see LICENSE. Back to the organization profile.