Creative references and demonstrated workflows

Use a project brief to preserve visual intent, and an executable project to preserve the ability to change it. The workbench examples demonstrate working tools; their visual choices are proposals, not evidence of personal taste.

Work from annotated references

Keep a short CREATIVE-BRIEF.md beside the project. The workbenches provide starting templates:

Work	Template or entrypoint	Editable handoff
Browser UI	desktop-workbench/assets/creative-brief.md, then browser-workbench	App source, assets/design tokens, run command, interaction tests, and viewport screenshots
Blender scene or game asset	blender-workbench/assets/creative-brief.md	.blend, project scripts, textures/fonts, named asset collection, and requested engine export
Motion graphics and video	media-workbench scaffold includes CREATIVE-BRIEF.md	React/SVG composition, source audio, lockfile, .kdenlive timeline, and rendered output

These paths are under ~/.claude/skills/, with shared harness links managed by the existing skill setup. To start a browser/design brief without copying an entire demo:

cp ~/.claude/skills/desktop-workbench/assets/creative-brief.md ./CREATIVE-BRIEF.md

For every useful reference, identify its source, a view or timecode, the specific aspect it should guide, and its status. A user-supplied image is a reference; it becomes an approved direction only when feedback establishes that scope. Approval of its lighting does not imply approval of its geometry or palette. Mark agent-found references and generated alternatives as candidates. Keep rejected and superseded directions annotated when that prevents repeating a failed choice.

Resolve consequential visual uncertainty with something inspectable: a viewport comparison, a material/camera preview, a short animated passage, or a working interaction. Preserve the selected version and the user’s actual feedback in the brief. Do not require a formal brief or approval meeting for an ordinary small edit; use the context already provided.

Concrete examples

For Blender, the orbital-compute example under ~/Documents/Codex/2026-09-07/ai-setup-upgrade/blender/orbital-compute contains native scenes, renders, animation samples, and a GLB. Its validation record covers native save/reopen, imported geometry/materials/animation, and resumed frames. A real brief could retain the asset-collection/export structure while replacing the orbital subject, orthographic camera, or materials. Those choices were agent-produced, not recorded user preferences.

For video, the Creative System example under ~/Documents/Codex/2026-09-07/ai-setup-upgrade/media/film keeps motion typography in src/index.jsx, generated sound in scripts/score.mjs, and assembly edits in out/creative-system.kdenlive. It has a verified MP4 and a separate MLT export. Native playback and a saved/reopened project-note edit were exercised. The dark palette and instrumental score are replaceable examples. Changing text inside React is different from trimming the rendered chapter in Kdenlive; the handoff explains which source owns each change.

For browser UI, use the browser workbench’s assets/interaction-lab for an editable create/complete/filter/reopen interaction and screenshots across browser engines. A design reference might govern the form hierarchy while the interaction test preserves behavior. Keep screenshots of the selected layout and the actual working source together. A passing browser test does not approve the visual direction.

Teach a workflow by demonstration

Record & Replay can turn a user demonstration into skill instructions. The desktop workbench’s record-and-replay.md contains the current macOS entrypoint and refinement procedure, sourced from OpenAI’s documentation. Recording is optional; no demonstration was captured during this setup work.

Use desktop-workbench/assets/workflow-template.md to name the recurring outcome, changing inputs, fresh starting observations, semantic actions, and result checks. Keep the provenance: user recording, user description, or agent-produced example. Refinement should remove accidental coordinates and one-off filenames while preserving meaningful user choices. Replay must observe the current project and controls, recover from an already-open file or in-progress export, and reopen the result before calling the task complete.

The media workbench includes kdenlive-replay.md, an agent-produced review/save/export/reopen example based on the exercised project. It is usable now with explicit project and output paths. A later user demonstration can teach preferred review gestures, naming, or export decisions without making recording a prerequisite for productive work.

Keep a project-specific procedure with that project. Extend an existing skill when it learns a broadly useful method. A distinct recurring workflow may earn its own skill after replay with changed inputs shows that it works beyond the original demonstration.