Artifacts · image

The image post.

A brief and a brand kit go in. A finished post comes out: caption, hashtags, true 4:5 and 9:16 masters rendered by code, the judge's score, and a spec you can revise in plain words. The first artifact the v2 engine mastered.

Straight from the engine

unretouched masters
Post the engine rendered for Kiln House Ceramics
Kiln House Ceramicssplit-blockArchivo Black
Post the engine rendered for Driftwood Roasters
Driftwood Roasterssplit-blockOswald
Post the engine rendered for Vela Studio
Vela Studiophoto-fullPlayfair Display
Post the engine rendered for Driftwood Roasters
Driftwood Roastersphoto-fullOswald
Post the engine rendered for Vela Studio
Vela Studioposter-gridPlayfair Display
Post the engine rendered for Kiln House Ceramics
Kiln House Ceramicstext-panelArchivo Black

Fictional brands, live output. The chips quote the spec the dealer dealt: layout archetype and display face.

How a static earns its pixels

the image lane of the v2 engine

The shared pipeline decides the words; these are the stages that decide the pixels of a single post. The full walk, gates and all, is on how it works.

  1. 01FormatcodeContent picks one of five formats: a number wants a stat card, a testimonial wants a quote card, a how-to wants a tip card, news wants an announcement, a real moment wants a photo hook.
  2. 02ImagerymodelReal brand photos lead, ranked and placed first. When the pool is spent, the image model paints a name-free, text-free atmospheric scene. No usable imagery at all still ships: a programmatic card.
  3. 03The dealcodeOne of five layout archetypes and a typography voice, dealt deterministically per brand and post. The same post re-renders identically; the next one refuses to repeat the shape.
  4. 04RendercodeA real layout engine composites the text over the scene, in the Worker. True 4:5 and 9:16 masters, separate renders, not crops.
  5. 05Judge + revisemodelFour hard gates, eleven weighted dimensions, a design tier from TEMPLATE to GREAT, and a bounded surgical revise before anything returns.

An artifact you can edit

spec + layers · edits are diffs

Not a JPEG. A document.

Every master ships with two structures. The spec is the recipe: every decision the engine made, replayable to identical pixels. The composition is the design file: typed layers (background, scrim, kicker, headline, accent, logo) with resolution-independent geometry, the same document Taka's editor manipulates on a canvas.

So a revision is a diff, not a re-roll. Ask for a shorter headline and the photo, frame, and palette hold verbatim. And when you dictate exact words, your words win: the judge still grades, and its dissent is recorded instead of obeyed.

~/socialkit · curl
bash
# Revise in plain words
curl https://api.socialkit.sh/v1/revise-visual-v2 \
  -H "Authorization: Bearer $SOCIALKIT_KEY" \
  -d '{
    "businessContext": { "businessName": "Driftwood Roasters" },
    "feedback": "shorter headline, keep the photo",
    "spec": { ...from the generate }
  }'

# The photo, frame, and palette hold.
# The headline moves. Re-judged, re-scored.

For agents

REST + MCP · one credit on success

POST /v1/generate-visual-v2 and POST /v1/revise-visual-v2 over REST, or the generate_visual and revise_visual MCP tools, annotated credit-bearing and slow so a well-behaved agent knows what it is spending. Always send an idempotency key; these calls take tens of seconds to a few minutes, and a retry must never double-spend. Blocked results cost nothing.

Visual quickstartAgents that ship pixelsNext artifact: the deck