grill-me

/grill-me — interactive design session → validated design plan

You interview the user to turn a fuzzy idea into a complete, validated design plan (design.json) — the source of truth for the model. Then you hand off to generation and a vision-grounded fidelity loop. The bundled CLI studio3d (on PATH) grounds your questions and builds the plan.

Request: $ARGUMENTS (plus any attached images).

Capabilities load on demand (registry). Only grill-me and model3d are native; the reference skills (cad-authoring, print-readiness, printer-setup, 3d-modeling-foundations) and specialist agents (cad-author, design-critic, spec-analyst, mesh-validator) live in the 3d-studio-registry MCP server — load a skill with load_skill(id="…"), get an agent's prompt with load_agent(id="…") then spawn it with the Agent tool.

(a) Ground yourself before asking anything

1. Read every attached reference image — note silhouette, proportions, colors, parts.
2. Identify the subject (owl, cat, dog, fox, rabbit, bear, dragon, trex, fish, bird, frog, robot, human_bust, snowman, ghost, mushroom, tree, rocket, car, generic_figurine) and a likely style. Look them up so questions are grounded, not generic:

   studio3d reference <subject> --style <style>   # BRIEF: silhouette cues, numeric
                                                  # proportions by head-unit H, CSG recipe,
                                                  # eye_rule, print_constraints
   studio3d styles                                # list the 8 styles
   studio3d styles <style>                        # one style's numeric params
   

   studio3d reference (no arg) lists subjects. If nothing matches, use generic_figurine.

(b) GRILL — ask only what's genuinely ambiguous (AskUserQuestion)

Use the reference + images to propose defaults; ask focused questions only where intent is truly underspecified. Pin down these essentials:

• Artistic style — offer the style list (clean, realistic, cartoonish, chibi, anime, low-poly, geometric, stylized); default from the look of the images.
• Overall size — height in mm (default from category; figurines ~60–90mm).
• Key characteristics / unique details — the cues that make it this thing (ear tufts? scarf? open mouth? a specific pose?).
• Colors — hex per part; which parts are which color; AMS/multicolor or single.
• Function / constraints — must it hold water (watertight vessel)? stand unaided? mate with/clip onto something? be purely decorative?
• Target printer — confirm the active profile (studio3d profile show) or which printer.

Batch related questions; don't interrogate. Where the reference gives a confident answer, state the default and move on rather than asking.

(c) Produce a complete, validated design plan

Create the base plan, then enrich and re-validate:

studio3d plan new --subject <s> --name "<n>" --style <st> --category <c> \
  --height <mm> --color "#hex" --out <slug>.design.json

Then edit <slug>.design.json to fill in the gathered intent — characteristics[], unique_details[], colors[{name,hex,part}], parts[{name,approach,size_mm,position_mm, purpose}], constraints{min_feature_mm,must_stand,watertight_vessel,functional}, reference_images[], notes — then:

studio3d plan validate <slug>.design.json     # must pass
studio3d plan show <slug>.design.json          # confirm the plan reads right

(d) Hand off to generation + the fidelity loop

The design.json is the base design — author FROM it, never around it:

studio3d plan brief <slug>.design.json         # grounded brief (proportions by head-unit H)

1. Author a studio3d DSL script by proportion from the brief (head-unit H and the CSG recipe — not ad-hoc mm). See the cad-authoring skill; delegate to the cad-author subagent for non-trivial geometry.
2. Fabricate from the plan (it sets style/size/color/category and persists design.json into the bundle):

   studio3d gen-script --script model.py --plan <slug>.design.json --out output
   

3. Vision-grounded fidelity loop (≤4 passes, keep best): render and compare to BOTH the reference's silhouette cues and the user's images.

   studio3d render output/<slug>/model.stl --color "#hex" --views front,right,top,iso
   

   Read the views; score silhouette / proportion / feature / style (each 0–100) vs the reference + images. Revise the script and regenerate (same --plan) until it matches, or delegate a blind judgement to the design-critic subagent. See the model3d skill for the full loop.

Tweaks edit the plan, then regenerate in place

Every later change ("taller", "add a scarf", "make the eyes bigger") edits a field in design.json and regenerates from it — the plan stays the single source of truth. Bump revision. Only branch to a new design.json when the user wants a separate variant.