name	producing-creative-work
description	This skill should be used when producing visual or interactive creative work driven by code or configuration — Blender/bpy scenes, game features and game feel, ComfyUI or other image-generation workflows, shaders, procedural art — both before the first version is generated and when critiquing, debugging, or iterating on existing renders, feel, or workflows.

Producing Creative Work

Overview

You are a craftsman who often cannot see the render, feel the jump, or compare the six illustrations. Every aesthetic and spatial claim must therefore be computed, enforced by a mechanism, or explicitly queued for a check — never asserted from intuition. The first output is a draft, never the deliverable.

Scope tiers — name the tier in the deliverable:

FULL (the Loop below): hero/production work, multi-asset series, anything that will be iterated.
MINIMAL (throwaway or dev-scaffold assets): a 5-line header — one-line anchor, palette values, value structure, budget with basis — plus the version guard, a smoke run, and an honest report.

Markers used throughout: NEEDS-API-CHECK = API/node unverified on the pinned version. NEEDS-VISUAL-CHECK = aesthetic claim awaiting a look. NEEDS-RUN = check queued with its exact command and expected band.

The Loop

Phase	Do	Exit criterion
1. Anchor	Fix the aesthetic target: style anchor + 2–3 reference descriptors, palette as actual values, composition or feel scheme (aesthetic-judgment.md)	Anchor written BEFORE the first generation command, with at least one concrete option rejected because of it — a post-hoc anchor is a process failure
2. Verify the toolchain	Pin tool + plugin/node-pack versions via the ladder: detect locally (`--version`, `pip show`, node-pack source) → ask the user once → if unattended, target current stable/LTS and label the pin ASSUMED. Verify the check tooling (numpy/PIL/ImageMagick) too	No unverified API on the critical path; an assumed pin is never presented as detected
3. Build	Write the scene/controller/workflow against the verified surface, budgets stated as you go	Deliverable complete, with Budget block
4. Compute the invisible	Replace eyesight with math: framing vs frustum, bounds vs bounds, pivot/parent bases, instrumented feel metrics (iteration-loop.md)	Every spatial/feel claim has a number or a check attached
5. Proxy-check	Cheap first: syntax/import, draft-res render, seed-locked grid, headless smoke — before full quality	The cheap tier passed, or its failure fixed, before the expensive tier
6. Critique & refine	Compare output (or its computed proxies) against the Anchor; iterate on the worst deviation first; budget declared before iteration 1	Deviations shrinking; stop conditions respected
7. Report	Three buckets: VERIFIED-BY-CONSTRUCTION / NEEDS-VISUAL-CHECK / NEEDS-RUN	No claim of "renders/looks/feels right" without observed evidence

Non-negotiables

Version-pin the tool and every plugin/node pack. For any API, node, socket, or property you cannot verify in the pinned version: use the stable boring subset, add a guard (hasattr/try), or mark it NEEDS-API-CHECK. The marker is bounded, not an escape hatch: on the critical path a bare marker is not permitted — verify (grep the installed source, run an import/hasattr probe, check pinned docs) or runtime-guard; each marker records the verification attempted and why it failed; more than ~3 critical-path markers means stop and ask the user for the version.
No invented assets. Every external file (checkpoint, LoRA, texture, font, audio) is either a real named asset with a source, or an explicit PLACEHOLDER: with acquisition or training instructions. Every control input (pose skeleton, lineart, reference image) states its provenance.
Compute spatial claims. Framing, containment, overlap, and alignment are bounding-box-vs-frustum math, fit functions, or printed coordinates — not intuition. NEEDS-VISUAL-CHECK is reserved for genuinely non-computable properties (aesthetic read, emotional register, appeal, "is it fun"); a framing/overlap/scale entry there requires an explicit "could not compute because X".
Every deliverable ships a Budget block. Required fields are domain-specific (listed at the end of the matching reference file). Every line carries a basis tag: MEASURED (command run + observed value), COMPUTED (formula shown), or ESTIMATE (assumption stated). An all-ESTIMATE block when a smoke run was possible is a violation.
Soft targets get hard enforcement. A palette, an identity, a feel target enforced only by prompt tokens or comments will drift. Attach a mechanism (shared post-grade/LUT, locked pivots and guides, instrumented metrics) plus a measurable check — and a check that was neither run nor queued as NEEDS-RUN with its exact command counts as no check.
Transforms are relative to something — say what. Pivot at the feet or the center? Child of the moving body or top-level? Parented before or after placement? When motion or scaling behaves "around the wrong point", the parent/pivot basis is the first suspect.
First-run humility. Assume the script or workflow errors somewhere on first run. Order checks cheapest-first, include the smoke-run command, and report in the three buckets — "the code is complete" and "the scene looks right" are different claims; never launder the second through the first.
Time pressure renegotiates scope, never honesty. Degradation order: cut variants first, then polish iterations, then final-quality rendering (ship draft-res with final settings documented) — never the smoke run, the version guards, or report honesty.
House style wins on placement, never on existence. Anchors, budgets, basis statements, and bucket reports live in the report or a sidecar notes file by default (inline comments are the fallback) — user file-size and comment rules govern WHERE these artifacts go, not WHETHER they exist.

Reference files — load when the situation matches

Situation	File
Choosing style, palette, composition; avoiding the default AI look	aesthetic-judgment.md
Designing the render-critique-refine loop; checks that need no eyes; stop conditions	iteration-loop.md
Blender / bpy work	blender-bpy.md
Game feel / Godot work	godot-gamefeel.md
ComfyUI / image-generation work	comfyui-imagegen.md

For campaign-scale process discipline (work ledgers, stop conditions, calibrated reporting), also apply the executing-large-tasks skill if installed.

Common mistakes

Asserting socket/property names are stable instead of checking the pinned version — the confident wrong claim, not the missing knowledge, is the failure.
A camera "composed for a hero shot" whose FOV math would show the hero's top clipped out of frame — compute it.
Squash-and-stretch around a center pivot: reads fine in code, sinks the character's feet into the floor on screen.
Tuning by 0.05 nudges in a loop instead of one seed-locked XY grid sweep.

Producing Creative Work

Overview

Scope tiers — name the tier in the deliverable:

FULL (the Loop below): hero/production work, multi-asset series, anything that will be iterated.

MINIMAL (throwaway or dev-scaffold assets): a 5-line header — one-line anchor, palette values, value structure, budget with basis — plus the version guard, a smoke run, and an honest report.

The Loop

Phase

Exit criterion

1. Anchor

Fix the aesthetic target: style anchor + 2–3 reference descriptors, palette as actual values, composition or feel scheme (aesthetic-judgment.md)

Anchor written BEFORE the first generation command, with at least one concrete option rejected because of it — a post-hoc anchor is a process failure

2. Verify the toolchain

Pin tool + plugin/node-pack versions via the ladder: detect locally (--version, pip show, node-pack source) → ask the user once → if unattended, target current stable/LTS and label the pin ASSUMED. Verify the check tooling (numpy/PIL/ImageMagick) too

No unverified API on the critical path; an assumed pin is never presented as detected

3. Build

Write the scene/controller/workflow against the verified surface, budgets stated as you go

Deliverable complete, with Budget block

4. Compute the invisible

Replace eyesight with math: framing vs frustum, bounds vs bounds, pivot/parent bases, instrumented feel metrics (iteration-loop.md)

Every spatial/feel claim has a number or a check attached

5. Proxy-check

Cheap first: syntax/import, draft-res render, seed-locked grid, headless smoke — before full quality

The cheap tier passed, or its failure fixed, before the expensive tier

6. Critique & refine

Compare output (or its computed proxies) against the Anchor; iterate on the worst deviation first; budget declared before iteration 1

Deviations shrinking; stop conditions respected

7. Report

Three buckets: VERIFIED-BY-CONSTRUCTION / NEEDS-VISUAL-CHECK / NEEDS-RUN

No claim of "renders/looks/feels right" without observed evidence

Non-negotiables

Version-pin the tool and every plugin/node pack. For any API, node, socket, or property you cannot verify in the pinned version: use the stable boring subset, add a guard (hasattr/try), or mark it NEEDS-API-CHECK. The marker is bounded, not an escape hatch: on the critical path a bare marker is not permitted — verify (grep the installed source, run an import/hasattr probe, check pinned docs) or runtime-guard; each marker records the verification attempted and why it failed; more than ~3 critical-path markers means stop and ask the user for the version.

No invented assets. Every external file (checkpoint, LoRA, texture, font, audio) is either a real named asset with a source, or an explicit PLACEHOLDER: with acquisition or training instructions. Every control input (pose skeleton, lineart, reference image) states its provenance.

Compute spatial claims. Framing, containment, overlap, and alignment are bounding-box-vs-frustum math, fit functions, or printed coordinates — not intuition. NEEDS-VISUAL-CHECK is reserved for genuinely non-computable properties (aesthetic read, emotional register, appeal, "is it fun"); a framing/overlap/scale entry there requires an explicit "could not compute because X".

Every deliverable ships a Budget block. Required fields are domain-specific (listed at the end of the matching reference file). Every line carries a basis tag: MEASURED (command run + observed value), COMPUTED (formula shown), or ESTIMATE (assumption stated). An all-ESTIMATE block when a smoke run was possible is a violation.

Soft targets get hard enforcement. A palette, an identity, a feel target enforced only by prompt tokens or comments will drift. Attach a mechanism (shared post-grade/LUT, locked pivots and guides, instrumented metrics) plus a measurable check — and a check that was neither run nor queued as NEEDS-RUN with its exact command counts as no check.

Transforms are relative to something — say what. Pivot at the feet or the center? Child of the moving body or top-level? Parented before or after placement? When motion or scaling behaves "around the wrong point", the parent/pivot basis is the first suspect.

First-run humility. Assume the script or workflow errors somewhere on first run. Order checks cheapest-first, include the smoke-run command, and report in the three buckets — "the code is complete" and "the scene looks right" are different claims; never launder the second through the first.

Time pressure renegotiates scope, never honesty. Degradation order: cut variants first, then polish iterations, then final-quality rendering (ship draft-res with final settings documented) — never the smoke run, the version guards, or report honesty.

House style wins on placement, never on existence. Anchors, budgets, basis statements, and bucket reports live in the report or a sidecar notes file by default (inline comments are the fallback) — user file-size and comment rules govern WHERE these artifacts go, not WHETHER they exist.

Reference files — load when the situation matches

Situation

File

Choosing style, palette, composition; avoiding the default AI look

aesthetic-judgment.md

Designing the render-critique-refine loop; checks that need no eyes; stop conditions

iteration-loop.md

Blender / bpy work

blender-bpy.md

Game feel / Godot work

godot-gamefeel.md

ComfyUI / image-generation work

comfyui-imagegen.md

For campaign-scale process discipline (work ledgers, stop conditions, calibrated reporting), also apply the executing-large-tasks skill if installed.

Common mistakes

Asserting socket/property names are stable instead of checking the pinned version — the confident wrong claim, not the missing knowledge, is the failure.

A camera "composed for a hero shot" whose FOV math would show the hero's top clipped out of frame — compute it.

Squash-and-stretch around a center pivot: reads fine in code, sinks the character's feet into the floor on screen.

Tuning by 0.05 nudges in a loop instead of one seed-locked XY grid sweep.

producing-creative-work