Blender spec

blender-spec — decide what "right" looks like before making anything

A model can only be verified against a spec, and only be excellent against a direction. "Make a chair" has neither a failure condition nor a taste target — this protocol gives it both. Output: a spec block posted to the user (and into the state file where long-task-state is installed).

1. Extract the spec — fill every row, mark guesses assumed

Field	Question	Default when unstated
Subject & identity features	What is it? Which 3-5 features make it read as this subject and not something else?	name them explicitly (a mug is: cylinder body, one handle, open top, flat bottom)
Art direction	3-5 style words + what they imply (e.g. "mid-century, warm, worn" → tapered legs, wood + brass, edge wear)	"clean contemporary" — and say so, it is a choice
Quality tier	draft / production / hero	hero — this pack is quality-first; downgrade only if the user says quick
Style	realistic / stylized-clean / low-poly faceted	stylized-clean unless art direction implies realism
Target use	still render / game asset / 3D print / scene prop	still render
Dimensions	real-world size of the whole and of key parts	archetype table below, or look it up; write the NUMBERS into the spec
Poly budget	tri limit, evaluated (modifiers applied)?	render ≤500k · game prop 2–20k · print: watertight matters, count doesn't
Topology	quads-for-subdiv? n-gons acceptable?	per blender-topology policy: quads on curvature, n-gons on flats
Materials	PBR realism or flat colors? palette?	PBR per blender-materials at production/hero; flat color-block at draft
Deliverables	what files count as done?	4-view probe set + lit 3/4 beauty; hero adds turntable (8 views) + 2-3 close-ups; plus .blend or build script

Target use drives everything downstream: game → budget + quads + applied transforms; print → manifold watertight, wall thickness, exact units; render → silhouette, shading, and presentation above all. If target use is genuinely unknowable from the request, it is the one question worth asking.

2. Reference before geometry — numbers and looks

• Web/image access available → collect 2-3 reference views of the subject AND 1-2 material/lighting references (what does "worn brass" actually look like?). Extract proportions as NUMBERS.
• No reference access → write the proportion sheet from knowledge anyway, numeric, before modeling. A wrong number visible in the spec gets corrected in review; a wrong shape discovered at the beauty render gets rebuilt.
• Specific named subjects (a brand product, a known character) you cannot look at → say you are interpreting freely, or ask for an image. Never imply likeness you cannot check.

Archetype proportions (starting numbers when the user gives none — override with reference):

Archetype	Numbers
Human (scale anchor)	1.7 m tall, ~7.5 heads; eye level 1.6 m
Chair	seat 0.45 h × 0.45 d, back top 0.8–1.0, legs 30–60 mm thick
Table / desk	top 0.74 h, 25–40 mm thick; dining 1.6×0.9
Door	2.0 × 0.9; handle at 1.0
Mug	95 h × 80 ⌀, wall 4 mm, handle clears 25–30 mm
Sedan car	4.5 × 1.8 × 1.45; wheels ⌀ 0.65, at corners not under body
Room	ceiling 2.4–2.7; counter 0.9; walkway ≥0.8

3. Assumptions and the one batch of questions

Ask only what changes the build and cannot be defaulted (target use for ambiguous requests; exact dimensions for prints; art direction when the subject is style-defining). One batch, not a drip. Everything else: pick the table default and record it marked assumed — the user vetoes the spec, not a questionnaire.

4. Done-criteria — measurable or it isn't a criterion

Write 5-10, each checkable by a specific render or probe. Shape:

• Silhouette reads as <subject> in all 4 canonical views (probe-render set)
• Overall dims within ±5% of <W×D×H> m; <key part> within ±10% (dimension probe)
• ≤ <N> evaluated tris; 0 non-manifold edges (0 boundary edges too, if printing)
• Every object semantically named; scale applied; one material per part-list entry
• <identity feature 1..n> present and visible in the 3/4 view
• <art-direction cue 1..2> visible in the beauty render (e.g. edge wear on handle corners)
• Rubric floor for the tier: production ≥3 all axes / hero ≥4 all axes (blender-verify)

"Looks nice" is not a criterion. "Handle clears the body by ≥ 25 mm so a finger fits" is. For hero, also state the excellence exit: two consecutive critique passes with only cosmetic findings.

5. Post it

Post the spec block (subject, art direction, tier, use, dimension sheet, budget, materials plan, deliverables, assumptions, done-criteria). If the user is present, give them one chance to veto; silence = the spec stands, and deviations from it are now deviations, not preferences.