prompt-crafter

Say what the picture must contain. Check that it does. Refuse when it doesn't.

A generative image pipeline will happily hand you a hero with the wrong face, the wrong palette and none of the faction's markings — and report success, because nothing looked. prompt-crafter replaces the opaque prose prompt with a typed contract of depictable claims, uses that same list twice — once to write the prompt, once to check the pixels — and blocks the asset when a required claim is not there.

The one idea: the contract's atom list is the same list used twice. Writing the prompt and checking the result read from one source, so the thing you asked for is the thing that gets verified. That is what closes the loop an opaque prompt leaves open.

Install

pip install prompt-crafter
pcraft --help

The distribution is prompt-crafter; the import package and the command are pcraft. Requires Python 3.11+. CI runs 3.11 and 3.13 on the core + [dev]. The [image] extra is not claimed on 3.11. The core's only runtime dependency is pydantic.

pcraft demo              # the whole loop end-to-end, no GPU, deterministic stubs
pcraft list              # contract ids in the store
pcraft validate          # resolve + compile the question DAG, no generate
pcraft gate <image>      # check an image against a contract
pcraft recipe            # emit the Cloud Kontext + fist-only Fill graph
pcraft replay <record>   # re-read a bound asset's provenance receipt

What a contract looks like

Not a prose prompt. A list of atomic, depictable, individually checkable claims:

• must_have — a garment, a palette, a silhouette, a sigil. Each names which gate tier verifies it, a severity, and optionally a depends_on edge, so there is no point verifying the colour of an axe that is not there.
• must_not — anti-constraints, verified as absence on the pixels. Not a negative prompt: negative prompts leave residual features and fall to paraphrase.
• identity_ref — a reference plate. Identity is conditioning, not tokens. Anatomical text makes a diffusion model render a specimen; a reference image binds the specific face.

Contracts inherit — a character extends a faction — and inheritance is fail-closed: a child may raise a requirement, never relax or silently drop one it inherited.

The exit code is the point

The 2 / 4 split is the whole design. "I could not check" and "I checked and it is bad" are different facts. Merging them is why browsers soft-fail certificate revocation, and why monitoring standards have carried a distinct unknown verdict since the 1990s. Every gate transcript also reports how many required tiers actually executed, so a gate that quietly stopped checking cannot read as a pass.

The verifier is always a different model family from the generator, enforced by a guard that refuses to run otherwise. CLIPScore is not used as the gate metric — it behaves as a bag of concepts, blind to which attribute belongs to which object.

Honest status

v0.3.0 — the core is real. SDXL conditioning is assembled in code. A local 5090 generate() has now been run. One Cloud recipe has been run live.

• 338 tests passing (counted 2026-08-18), GPU-free and deterministic. The whole suite runs against a mock generator and verifier, which is what proves the plugin boundary holds.
• Flux Fill inpaint is wired. method=reference writes the Cloud recipe (GATE_CLOUD_SUBMIT). ControlNet pose and IP-Adapter stay refused on Flux.
• The eleven compound decision points in the core are mutation-tested — 20 of 21 mutants killed, and the survivor is named rather than hidden.
• SDXL ControlNet OpenPose, IP-Adapter, LoRA, InstantID, and regional inpaint are wired and covered by fake-torch tests. InstantID and IP-Adapter cannot share one generate. Two IP-Adapter plates stay on one adapter (all images; scale is the strongest lock). Local generate() ran on the 5090 (2026-08-18, seed 169405236028824). The frame is orcish; grip, sigil, and bracer did not land. Flux refuses those identity methods.
• pcraft recipe emits the Cloud Kontext stitch + left crop + fist-only Fill graph. A live Cloud submit (2026-08-18) produced a single-panel crop and kept the bracer.
• Tier-2 is a real DSG expansion. Escalation is a contrastive checkpoint. A live GEPA compile ran 2026-08-18 on local Ollama hermes3:8b (600B was not up). Pinned sprite.synth.v1-gepa.json. The seed artifact is untouched. The per-asset loop still uses TemplateSynthesizer.
• The identity sub-gate is not wired. Its thresholds have no holdout.
• Pre-1.0 deliberately, and a test enforces it. Promotion should follow evidence, not a version bump.

Trust

No credentials are read, stored or transmitted. No telemetry — there is no opt-out because there is nothing to opt out of. The core imports no networking library at all; the optional extras reach a model host by their nature, and installing them is a choice. File operations are not sandboxed: --records-dir and --db write where you point them, deliberately, for a local-first tool.

Deliberate refusals carry a code, a message and a hint, and raise rather than assert — so -O cannot delete them, and the suite runs a second time under -O to prove it.

Documentation and handbook → · Source · Changelog

MIT