A MoreSalamander · StudioLabs production

my-AI-script

The writer of the my-AI suite. An LLM interviews you about what you want, then keeps writing until it can pass a pure-Python score — and only then emits a validated ProductionSpec (for my-AI-scene) or SongSpec (for my-AI-beats).

Brief in. Spec out.

What it is

It turns a loose idea into something a machine can check.

A vague brief is useless to a renderer. my-AI-script does the hard part — pinning a loose idea down into a complete, structured spec — and it refuses to emit one until that spec would actually pass the downstream validator. The interview manufactures the rigor up front; a deterministic scorer holds the line. The model proposes; Python disposes.

Watch it run

The model proposes a spec. The scorer sends it back until it's good.

my-AI-script · live pipeline attempt 1

interview

→

llm_generate

→

spec_validate

gate ●

→

spec_score

gate ●

→

spec_ready

score

— / 100 · need ≥ 75

Every stage emits the suite's shared NDJSON events. Two blocking gates ● in sequence: spec_validate checks structure, spec_score checks quality. Below 75, the breakdown feeds back and the model rewrites. The scorer is never an LLM.

The three moves

Explain · Synthesize · Verify — the whole methodology, in one tool.

01 · Explain

The interview is the constraint-gathering phase. The chosen target's schema defines every field that must be filled — the LLM's questions serve the schema. The scoring rubric is in the system prompt before the model writes a word, so it knows the bar it's aiming for.

02 · Synthesize

Once the model is confident it can score well, it generates the full spec from the conversation — every beat and narration line for a video, or every section, prompt, and energy value for a song. Reference-quality detail, on demand.

03 · Verify

A target-specific pure-Python judge scores out of 100 — video on narration/footage/grade, music on prompt specificity and the energy arc. Below 75 the breakdown feeds back and the model revises. The scorer is never an LLM.

Best at — and the honest limit

One engine, two targets.

my-AI-script is the universal front door to the suite: it routes a brief to VIDEO → a ProductionSpec (my-AI-scene) or MUSIC → a SongSpec (my-AI-beats), via one Target abstraction. Proven end to end: a synthwave interview drove a local model to a 98/100 SongSpec, validated and rendered.

What it can't do: the scorer checks that a spec is complete and renderable — not that the finished video or song is good. It manufactures rigor about structure, and is honest that taste lives downstream. A complete spec is a floor, not a guarantee of art.