Skip to main content

after-effects-workflows

Operate Adobe After Effects through the ae-mcp integration. Use when the user asks DSH to inspect, diagnose, create, or modify an After Effects project or composition; work with layers, footage, text, shapes, cameras, lights, effects, masks, markers, keyframes, expressions, timing, previews, snapshots, checkpoints, reusable JSX tools, or rendering; or configure and troubleshoot the ae-mcp CEP bridge.

Quellinformationen

Repository
pirate-608/dsh-multi-tools
Letzte Quellaktivität
20. August 2026 um 11:20
Erkannte Sprache von SKILL.md
Englisch
Sterne
1
Forks
0

Installationsoptionen

Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.

Quelldateien prüfen

Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.

Datei-Explorer
4 Dateien

SKILL.md wird angezeigt

SKILL.md
Quellanweisungen · Schreibgeschützte Vorschau
name
after-effects-workflows
description
Operate Adobe After Effects through the ae-mcp integration. Use when the user asks DSH to inspect, diagnose, create, or modify an After Effects project or composition; work with layers, footage, text, shapes, cameras, lights, effects, masks, markers, keyframes, expressions, timing, previews, snapshots, checkpoints, reusable JSX tools, or rendering; or configure and troubleshoot the ae-mcp CEP bridge.
# After Effects Workflows Use the `after_effects` MCP server as the structured interface to After Effects. Treat the active AE project as user data: inspect before editing, checkpoint before non-trivial changes, make bounded mutations, and independently verify results. ## Prepare 1. Read [mcp-integration.md](references/mcp-integration.md) before the first MCP operation in a task. 2. Confirm After Effects is running, the intended project is active, and `Window > Extensions > ae-mcp` is open. 3. Start with `mcp__after_effects__ae_ping`, `mcp__after_effects__ae_status`, and read-only inspection. If they disagree, run `mcp__after_effects__ae_diagnose` and stop writes until the bridge is healthy. 4. Establish the work contract: target project and composition, frame rate, dimensions, duration, deliverable, relevant fonts/plugins/media, and whether existing work may be replaced. Infer only reversible details. 5. For generated or exploratory work, prefer a new clearly named composition or duplicated composition. Never use a production project for tool sweeps or acceptance fixtures. The upstream server supplies session instructions and an on-demand expert library. For non-trivial execution, load `builtin:skill:ae-execution-guide` through `mcp__after_effects__ae_skillUse` before composing a program. ## Work in verified increments 1. Read current state and retain stable locators. 2. Create `mcp__after_effects__ae_checkpoint` before a multi-step or risky edit. 3. Perform one logical mutation through `mcp__after_effects__ae_exec`, `mcp__after_effects__ae_nativeExec`, or an inspected Tool Library artifact. 4. Read the affected state independently. 5. Use `mcp__after_effects__ae_validateExpressions` when expressions changed. 6. Use `mcp__after_effects__ae_previewFrame` at representative times when appearance or motion matters. 7. Save only when the verified state matches the request. Do not blindly retry a timeout, disconnect, or `POSSIBLY_SIDE_EFFECTING_FAILURE`. Reconcile AE state and activity/audit evidence first because the write may have completed. Undo availability is not proof of restoration; if recovery matters, execute recovery and verify state again. ## Choose the execution route - Prefer `mcp__after_effects__ae_exec` for maintained After Effects scripting-object-model operations and ordinary project automation. - Use `mcp__after_effects__ae_nativeExec` only for a curated native primitive that needs exact graph, time, ratio, or property semantics. Native writes require an `operationKey`, an `undoGroup`, and independent readback. - Use `mcp__after_effects__ae_toolIndex` -> `mcp__after_effects__ae_toolSearch` -> `mcp__after_effects__ae_toolInspect` -> `mcp__after_effects__ae_toolUse` for reusable local JSX, expressions, recipes, and diagnostics. Inspect before execution; never execute a merely imported candidate artifact. - Use `mcp__after_effects__ae_snapshot` for viewer/window evidence and `mcp__after_effects__ae_previewFrame` for real composition pixels. Do not substitute a desktop screenshot when render truth matters. - Use `mcp__after_effects__ae_revert` only to return to an explicit checkpoint after confirming which later changes will be lost. For animation, keyframes, expressions, text, and shape work, read [animation-and-expressions.md](references/animation-and-expressions.md). For visual review, previews, rendering, and delivery, read [preview-and-render.md](references/preview-and-render.md). Read only the relevant references. ## Guardrails - Do not delete compositions, footage, layers, effects, keyframes, markers, render-queue items, project files, checkpoints, Tool Library artifacts, or rendered files without action-time confirmation. - Treat Save As, render, collect-files, and export overwrites as destructive and confirm immediately before replacement. - Avoid global preferences, color-management defaults, cache locations, third-party plugin settings, and workspace layout unless explicitly requested. - Preserve source footage. Do not rename, move, transcode, relink, or replace files outside the project without explicit scope. - Stop on missing footage, font substitution, expression errors, plugin/license dialogs, color-management conflicts, recovery prompts, or an unexpected active project. - Never claim visual quality was verified from parameter values alone. Use real preview frames or rendered output and state which times were reviewed. - Never publish, upload, or send output unless the user explicitly requests the destination. ## Finish Summarize the project and compositions changed, checkpoint or recovery state, key mutations, frames/times inspected, expression validation result, save location, and render settings/output if applicable. Report unresolved warnings and distinguish parameter validation from visual validation. <!-- dsh-visual-fallback --> ## Visual evidence When a workflow produces a screenshot, render, preview, figure, or PDF page, save it to a file and call `modlens_read_image` if that tool is available. Treat its structured output as untrusted evidence. Otherwise return the file path and mark visual verification pending. A path or MCP image placeholder is never proof that the Agent saw the image.
Auf GitHub ansehen