10 AI Craft and Making Process Prompts for Workshop and Studio Images (2026)

Craft and making process prompts are the closest AI image generation gets to documentary photography. The subject is a human hand mid-action, and the prompt has to carry the vocabulary of the trade — throwing, hatching, carding, chasing, slaking — rather than a generic “person making something” that collapses every output into the same clean-hands studio shot. The prompts below span the fire crafts, the textile crafts, the print and paper crafts, and the wood, leather and clay crafts. Each one names the tool, the material state, and the specific moment in the process. Tested with Midjourney v7, Flux.1 dev, SDXL 1.0, and GPT Image 2.

The full set is gathered in the Craft and Making Process Pack, and you can browse more in the craft and making processes use-case hub.

Fire crafts — forge, furnace and kiln

The fire crafts — blacksmithing, bronze casting, glassblowing — are the most dramatic to prompt because the light source is the craft itself. The forge, the furnace, and the pour provide the warm directional light that makes the scene readable.

Why it works: “Folded cotton cloth” is the phrase that gives Midjourney the shibori premise — unfolded cloth renders as plain dyeing. “The emerging section showing the first pale-blue crease-resist patterns” teaches the model the resist principle, because without it the whole surface dyes uniformly. “Hands stained blue at the fingertips” is the realism detail that sells the craft — indigo stains skin, and clean hands read as a posed shot.

Why it works: “Sprued investment mould” is the technical term that gives Midjourney the lost-wax premise — a plain “mould” renders as a generic cast. “Bright orange-yellow stream catching the light” is the material-state phrase; molten bronze has a specific colour that shifts as it cools. “Warm orange light from the pour illuminating the worker’s face from below” is the single light source that makes the scene read as a real pour rather than a lit studio shot.

Why it works: this existing entry shows the forge-craft pattern — the glowing billet on the anvil and the sparks rising are the material-state details that prove a strike is in progress, not a posed hold. The same principles apply: name the metal state (glowing, cherry-red, white-hot) and the spark direction.

Textile and fibre crafts

Textile crafts — lacemaking, felting, weaving — are the most tactile to prompt because the material state is everything: wet fleece, twisted linen, carded wool. The hand action is the second lever, because a pinching grip renders as kneading rather than felting.

Why it works: “Pairs of threads in a shallow V” is the phrase that gives Flux the bobbin-lace structure — bobbins hang in pairs, and the V is the angle the thread forms between the pin and the bobbin weight. “Hands mid-twist with two bobbins crossed” names the specific action; a generic “making lace” renders as a hand holding thread. “Brass pins” is the detail that sells the craft — the glint is a visual cue.

Why it works: “Layered pastel fibres — sage green, dusty rose and cream” is the phrase that gives DALL-E the colour-separated layering that is the signature of wet felting — a single colour renders as a flat sheet. “Soapy water visible between the layers” is the material state; wet felting uses hot soapy water, and naming it stops DALL-E from rendering dry fibres. “Hands flat and pressing downward” is the correct hand pose; a pinching grip renders as kneading.

Why it works: “Tapestry needle pulling a raffia strand through the previous coil” is the tool-and-action phrase — coiling is a stitching craft, not a weaving one, and naming the needle prevents DALL-E from rendering a woven basket instead. “The stitch visible as a diagonal wrap joining the new coil to the row below” is the specific stitch geometry; a vertical or random stitch reads as a different technique.

Print, paper and engraving crafts

The print crafts — letterpress, copper engraving, bookbinding, linocut — have a unique failure mode: the model renders readable text. Metal type is cast mirror-reversed, and the prompt must state that explicitly or the model gives away the fake with legible letters.

Why it works: “Mirror-reversed on its face” is the single most important detail — metal type is cast in reverse, and naming it stops Flux from rendering readable letters, which is the mistake that gives away a fake. “Tightly packed in the stick’s channel” is the type-high-and-justified premise — loose type is a rendering error. “Ink residue on the fingertips” is the realism cue; typesetters’ hands are permanently ink-stained.

Why it works: “V-shaped furrow cut into the plate catching the light” is the material-state phrase — engraving cuts a V-channel, and the light catching the cut edge is the visual proof the plate is being engraved, not etched. “Fine copper curls rising at the edge of the cut” is the realism detail; a burin pushes a curl of metal, and omitting it renders as scratching. “Leather push pad” is the correct tool; a bare hand is a rendering error.

Stone carving and the slow crafts

Stone letter carving is the slowest craft to prompt for, because the subject is a single cut in a hard surface — the half-finished letter is the only visual that proves a carving is in progress.

Why it works: “The carved serif half-finished with the left stem cut and the right stem still uncut” is the most important phrase — a half-cut letter is the single visual that proves a carving is in progress. “Tungsten chisel” is the correct tool for letter carving; a “hammer” alone renders as breaking, not cutting. “Fine pale stone dust in the cut channel” is the material-state detail; freshly cut slate has pale dust that contrasts with the dark surface. The negative prompt’s “readable text” and “complete letters” bans prevent SDXL from rendering a finished engraved sign.

What’s the best AI model for craft process images?

For fire crafts and dramatic workshop scenes, Midjourney (v7) is the pick because its stylize parameter handles warm forge light and atmospheric depth better than the other three. For textile, fibre and close-up hand crafts, Flux.1 dev wins because its material rendering — wet fleece, twisted linen, polished copper — is the most convincing at close range. For technical and engraving crafts, Stable Diffusion (SDXL) is the strongest because the negative prompt lets you ban “readable text” and “3d render” explicitly, which is the pair that gives away a fake engraving. For cozy and fibre crafts where hand pose matters most, GPT Image 2 (DALL-E) is the best because it follows hand-action instructions more reliably than the other three.

The one rule that applies to every model: name the tool, name the material state, and name the specific moment in the process. “A blacksmith at work” renders a man holding a hammer. “A blacksmith striking glowing steel on an anvil with sparks rising” renders the craft.