Roadmap: AI Assembly "Describe a Goal → an Engine" (Honestly: In Development)

In short

This article is about the future, and it says so up front. The from-goal idea: you write a single sentence — "I want a CRM for a small studio" — and the engine decides on its own which roles, prisms, rules and MCP servers are needed, producing a ready spec to build from. Today that does NOT exist in the product: from-goal is a P2 roadmap item, because assembling from a free-form goal requires a server-side LLM layer with reproducible output — that's a separate engineering effort, not a wrapper around a prompt. What already works: the engine is assembled deterministically through presets, the builder and a compiler targeting three tools. Below: a "same goal, today vs tomorrow" comparison and the other roadmap items we don't pass off as finished.

This article is about the FUTURE. Everything marked roadmap has not been built yet — we don't present it as done.

The idea. The ideal flow for a beginner: you write one sentence — "I want a CRM for a small studio" — and the engine decides which roles to enable, which prisms are needed, which rules and MCP servers to wire in, and hands you a ready spec to build from. That's from-goal: AI assembly, "describe a goal → an engine".

The honest status. from-goal is a P2 roadmap item (in the project audit: "AI assembly 'describe a goal → an engine' (from-goal, server-side LLM)"). It does not exist today in the working product. While from-goal is in development, the picture is simple: you pay once for delivery of an assembled engine (see the article on buying).

Why not "tomorrow". Assembling from a free-form goal requires a server-side LLM layer that has to survive a weak target model, invent nothing extra and deliver reproducible output. That's a separate engineering effort, not a wrapper around a prompt. So it's more honest to call it a direction than a feature.

What already works TODAY (instead of from-goal)

Until AI assembly ships, the engine is assembled deterministically and predictably:

1. From a seed/preset. POST /api/quest-mods/from-seed creates an engine from a ready seed (the flagship is Quest: a team of roles, prisms, the full quest workflow, phases) or from the catalog of agent/prism/skill/MCP presets.
2. Manual assembly and editing. You enable the roles, prisms and rules you need (always / auto-by-glob / agent / manual), plus skills and MCP servers, and edit the spec.
3. Text-based elicitation (intake v2). This is NOT AI assembly, but it's already a real layer: the exported package carries an "ask like a team" recipe — up to 4 questions in one batch, each with a recommended default, routed by complexity via key substrings (migrat/billing/auth/deploy → the heavy route with a spec approval). So "I want a CRM → clarifying questions → a spec" is partly covered ON THE TARGET MODEL'S SIDE, by a literal recipe in the package.

Worked example: the same goal today and tomorrow

Input: "I want a CRM for a small studio."

Today (for real):

1. You take Quest via from-seed — you get a team of 8+ roles, prisms with veto power, and the quest workflow.
2. You install the engine into your tool. On the first task its built-in elicitation asks you ≤4 questions with recommendations (which stack, do you need multi-tenancy, where to store data) and builds an EARS spec and a plan from your answers.
3. From there the usual pipeline runs: spec → plan → step + check → prism review → delivery with proof.

Tomorrow (roadmap from-goal):

1. You type the same sentence into a "describe your goal" field.
2. A server-side LLM layer picks the roles/prisms/rules/MCP servers itself and generates a draft spec — no manual assembly.
3. You edit the draft and buy delivery.

Result. Today you get the same quality engine, you just start from a ready Quest and answer a couple of questions instead of starting from a blank sentence. The gap between "today" and "tomorrow" is in the convenience of the starting point, not the quality of the assembled engine.

Other honest roadmap items (not presented as done)

• Specialized directions. Part of this roadmap is already delivered — but not as "separate catalog engines"; as modules that extend the engine: deck-craft (presentations), design-craft (design before code), data-insight (read-only data analysis with conclusions), deep-research (verifiable research), plus red-team, iron-gate, prompt-forge and landing-copy. Each is bought separately and plugs into your build. There are no video or growth modules yet (video is currently covered only by themed gallery templates) — that stays on the roadmap; full data pipelines/ETL are also outside the base Quest engine's coverage.

We don't attach invented effectiveness numbers to from-goal and we don't call it finished — it's a direction the product is growing toward.