run-workflow — Claude Code skill for Firefly workflow automation
A skill that routes agent requests to the run-workflow MCP server. It lets an agent build, run, publish, and save Adobe Firefly workflows.
Apache-2.0 250 installs ★ 335 updated 2026-11-07
npx skills add adobe/skills --skill run-workflow -a claude-codeWhat run-workflow does
The skill connects an agent to the run-workflow MCP server and keeps its safety rules loaded before any request is handled. It processes images, video, and 3D models in .usdz and .glb formats through Firefly workflows, with routing for creation, execution, and publishing.
Routing is strict. A bare "run banners at scale" goes to list_golden_workflows, a bare ID only triggers inspect_run, and compose requests always start from scratch with compose_workflow. The agent calls list_actions before touching nodes, so the catalog stays current.
This requires an enterprise Adobe account with a Firefly Services or Firefly Creative Production entitlement assigned through the Adobe Admin Console. Personal accounts will not work, and that is the main constraint: no enterprise access, no server.
What it helps with
- compose a Firefly workflow from scratch
- run image or video processing through a workflow
- save a draft and publish a workflow as an API
- inspect a failed execution by ID
- find and run a Quickstart workflow
Example prompts for the agent
- Compose a workflow that generates banners from a CSV template
- Run the
banners at scaleQuickstart workflow - Inspect what happened with
batchId=...
How to install run-workflow
1Skills CLI (Vercel)
npx skills add adobe/skills --skill run-workflow -a claude-code2Manually: put the files in place
~/.claude/skills/run-workflow/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/run-workflow/
1Skills CLI (Vercel)
npx skills add adobe/skills --skill run-workflow -a codex2Manually: put the files in place
.agents/skills/run-workflow/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add adobe/skills --skill run-workflow -a cursor2Manually: put the files in place
.cursor/skills/run-workflow/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install adobe/skills plugins/creative-cloud/run-workflow/skills/run-workflow2Manually: put the files in place
.github/skills/run-workflow/SKILL.md1Skills CLI (Vercel)
npx skills add adobe/skills --skill run-workflow -a gemini-cli2Platform CLI
gemini skills install https://github.com/adobe/skills --path plugins/creative-cloud/run-workflow/skills/run-workflowManually: put the files in place
.devin/skills/run-workflow/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add adobe/skills --skill run-workflow -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about run-workflow
How do I install run-workflow in Claude Code?
Run npx skills add adobe/skills --skill run-workflow -a claude-code, then open ~/.claude/skills/run-workflow/SKILL.md.
What does run-workflow do?
It connects an agent to the run-workflow MCP server so it can discover, compose, run, publish, and save Adobe Firefly workflows. It handles images, video, and .usdz or .glb 3D models.
What Adobe account does it require?
An enterprise account with a Firefly Services or Firefly Creative Production entitlement, assigned through the Adobe Admin Console. Individual and personal accounts are not supported.
Source text
SKILL.md · Apache-2.0
---
name: run-workflow
description: >-
Use the run-workflow MCP to discover, compose, execute, publish, and save
Firefly workflows. TRIGGER when: user asks what the MCP can do; wants
to process images/video/3D models (.usdz/.glb) via workflow; wants to build/run/save/publish a
workflow; wants to list/view saved or recent workflows, presets/interfaces,
or Quickstart workflows (e.g. 'run banners at scale'); pastes any
workflow/batch/execution ID. Create/compose requests default to
compose_workflow from scratch. Call get_featured_workflow ONLY when explicitly asked for a
featured/predefined template. Bare 'run banners at scale' = QUICKSTART
(list_golden_workflows). BARE ID = INSPECT ONLY — call inspect_run, never
run_workflow_submit. ALWAYS call list_actions first for action/node
discovery; a "what tools" question is answered directly from your visible
tool list, no call. DO NOT TRIGGER for direct Firefly API calls without
MCP (use firefly-api-specs).
compatibility: Requires an enterprise Adobe account with a Firefly Services or Firefly Creative Production entitlement, assigned via the Adobe Admin Console. Not available to individual/personal Adobe accounts.
license: Apache-2.0
metadata:
version: 1.0.0
visibility: public
---
# run-workflow MCP
Discover, compose, run, publish, and save Adobe Firefly workflows through the run-workflow MCP.
Requires an enterprise Adobe account with a Firefly Services or Firefly Creative Production
entitlement assigned via the Adobe Admin Console; individual/personal accounts are not supported.
This core routes requests and keeps safety rules always loaded. `packages/mcp-server/` is the
source of truth for `SKILL.md` and `references/`; bump `skill-version.json` with every change.
Read the relevant reference **before** acting; do not load unrelated references or re-read them
just to present completed outputs.
| Read before... | Reference |
| --- | --- |
| Listing workflows/runs, executing, polling, presenting outputs, saving, or publishing | [Workflow guide](references/workflow-guide.md) — full routing, taxonomy, lifecycle, key rules, and anti-patterns |
| Processing a creative/marketing brief or composing template + CSV data merges | [Intake](references/intake.md) — required inputs, inference vs. questions, one compose call, run/save offer |
| Using an explicitly requested featured template, custom INDD/PSD, or MOGRT motion graphics template | [Featured and templates](references/featured-and-templates.md) — compose/rewire, CSV mappings, DGR controls and video counts |
| Composing a graph, handling missing inputs, multi-variant outputs, or a 3D scene | [Compose](references/compose.md) |
| Inspecting IDs, rerunning, handling errors, reattachment, or filtered history | [Diagnostics](references/diagnostics.md) |
| Uploading files, resolving inline pastes, macOS permission errors, or downloading outputs | [Asset input](references/asset-input.md) |
| Alert triage/RCA (stdio CLI only, `login --dev`, Splunk + Slack) | [Alert RCA](references/alert-rca.md) |
## Bare ID = inspect only
When the user's entire input is an ID (UUID, workflowId, batchId, executionId) without context:
1. Call `inspect_run(id)`.
2. Present executions, outputs, and diagnostics.
3. **STOP. Never call `run_workflow_submit` as a follow-up.**
"run-workflow" is the product name, not an execution instruction. Submission requires explicit
intent such as "run", "execute", "process", "generate", "rerun", or "redo". Never try submission
first and fall back to inspection after a 404. Questions about required assets for a
Quickstart workflow/interface use `inspect_run({ id: workflowId, interface: true })`,
not `get_featured_workflow`.
## Discovery and routing
Ground capability answers in live tools, not training knowledge. **Tools** and **actions/nodes**
are different vocabularies: answer "what MCP tools do you have?" directly from visible tool
definitions without calling a tool; call `list_actions` first for action/node catalog discovery.
Use `get_action_schema` for uncertain action parameters.
Homepage showcase workflows are **Quickstart Workflows**, formerly "golden workflows/demos".
Both names resolve to `list_golden_workflows`, not featured templates or internal golden fixtures.
Preserve configured headings in `renderedTable` verbatim, including
`RW_MCP_QUICKSTART_WORKFLOWS_HEADING` (default: `Quickstart Workflows`).
| Intent | Route |
| --- | --- |
| Create/compose, including named use cases, banners at scale, templates, or briefs | `compose_workflow` from scratch; no featured lookup/base unless explicitly requested |
| Explicitly use a featured/predefined workflow/template | `get_featured_workflow`, then follow the featured/template reference |
| List Quickstart/golden workflows | `list_golden_workflows`; render `renderedTable` verbatim |
| Run a Quickstart workflow (e.g. "run banners at scale") | `list_golden_workflows` to resolve its workflowId; ask which if multiple matches, then submit `workflowId`, `inputs: []`, `interfaceRerun: true` unless the user requested their own assets |
| Any broad "list my workflows"/"what have I built/saved?" request, even with "all" | One `list_all_workflows` call; no separate Quickstart lookup or history call |
| Filtered/date-ranged own workflows | `list_workflow_history` |
| Interfaces/presets (definitions) | `list_interfaces` |
| Jobs/interface/preset runs | `list_interface_runs` |
| Plain workflow runs/"what have I run?" (not interface-specific) | `list_workflow_runs`, which excludes interface/preset runs |
| Broader/filtered run queries (any source, specific UI/host) | `list_batches` with appropriate filters |
| Single most recent run | `list_workflow_history(limit=1)`, then `inspect_run` |
| One named run, pasted ID, outputs, logs, or failure | `inspect_run` |
| Known-good examples to debug compose; reload edited examples | `get_workflow_examples`; `reload_examples` |
| Upload local file; abort running batch | `upload_asset`; `cancel_workflow` |
| Execute; poll | `run_workflow_submit`; `run_workflow_get_status(includeOutputs: true)` |
| Save editable draft; publish reusable API; get published curl | `save_workflow`; `publish_workflow`; `generate_curl` |
| Catalog stale | `refresh_catalog` |
| Display/download outputs | `display_asset` only outside Claude Desktop; `download_output` only after the user asks |
Read the workflow guide for complete listing semantics. Render each listing tool's `renderedTable`
verbatim. For `list_all_workflows`, it already contains Quickstart rows, `typeSummary`, then the
own-workflows table in that order; never split, summarize, reconstruct, or separately paste
`typeSummary`. Mention `goldenUnavailable` when true. Each new listing intent requires a fresh
matching tool call, not filtering an earlier result from a different catalog.
Use plain type labels: Published, Preset (same as interface), Inline, Draft, Preview, Shared.
By default hide un-run drafts; include them only when explicitly asked. A draft stays a draft
after execution. Never offer to edit/republish a Shared workflow or claim another author's work.
## Compose and execute safely
- Describe intent, not node versions, ports, tags, or wiring. Do not manually design the graph.
For template/CSV data merges, gather the required template + CSV first (unless explicitly told
to "just compose"); infer from files and ask for optional fonts and output/run/save intent.
Pass `customTemplateUrl`, `csvUrl`, and `fonts` in one compose call. Do not pre-call tag
extraction/catalog tools unless the user asks to inspect them.
- For 3D products, only `.usdz`/`.glb` scenes are supported; pass `sceneUrl` to compose.
For Composite images (3D), call `describe_scene` once, use an exact user-provided name,
a single candidate, or `main_camera`; otherwise ask using returned names. Pass exact
`sceneCamera`/`sceneHeroAsset`. For Render 3D/360 spin, only select a camera if the user names one.
- Never silently substitute a simpler action, generate placeholder inputs, or use scripted
workarounds. Surface missing inputs and compose errors. Re-compose at most once after the
cause is fixed with the same intent-only message; never drop CSV/templates or request legacy
wiring to bypass failure.
- Retain `session_id` from compose/featured lookup for submit, save, and publish (2-hour lifetime).
Offer Run now / Save to Workflow Builder / both after compose. If the graph was edited, validate
it with `validate_workflow` and submit explicit edited actions/connections/inputs; the cached
session alone still points to the unedited graph.
- Submit all assets in **one** call, not one per image. Use `prepared.inputNodes` for featured
inputs. Copy presigned URLs verbatim; changing any character invalidates the signature.
- **ASK-FIRST for draft/published assets:** ask existing/prior vs. user-supplied assets unless
already established by the prompt; prefer checking prior asset resolvability and collect fresh
files for missing assets. This also applies to never-run published workflows; do not rely on
submit's automatic confirmation guard. Check `durabilityCoverage`, not just `resolvedVia`.
- **Quickstart/interfaces/presets are the opposite default:** submit `workflowId` + `inputs: []`
directly, with no asset confirmation or pre-verification; baked defaults are curated/durable.
Preserve an explicitly selected interface definition with `interfaceRerun: true` (interfaces
share their source workflow's ID). Only collect overrides when the user's prompt asks for them.
Do not reuse a prior run's custom assets silently.
- For same-assets reruns, use `workflowId` alone with `inputs: []` when available, not an inline
graph. Read diagnostics first for `rerun` object routing, first-run cases, `inline_exact`,
`interfaceRerun`, `needs_reattach`, `needs_rewire`, and `no_editable_graph`.
- Never force past `invalid_workflow`, `needs_rewire`, or `no_editable_graph`. A corrupt published
definition needs correct republication; do not compose a replacement for a user who asked to run it.
## Poll, present, and persist
Read the workflow guide before submission for the detailed lifecycle and output protocol.
Always poll with `includeOutputs: true` to a terminal state; use returned `elapsedSeconds`
and `percentage` for progress, not estimates. Wait 15s below 60s elapsed, 30s from 60–180s,
60s above 180s. Only preset/interface runs offer completion email instead of continued polling.
Show `downloadedPreviewOutputs` immediately as in-progress. On failure, inspect, summarize,
and stop; never auto-retry/rewire.
On completion, present every `VIEW YOUR ASSETS:` folder link first, then all `OUTPUT URLS:`
lines verbatim (not the trailing JSON). For oversized responses, extract those text blocks
once from the temp file; do not re-poll or repeatedly parse JSON. No placeholders/truncated URLs.
Then offer one combined choice: download locally, keep links, or next steps (save/publish/new
workflow). In stdio, download to the user's chosen `saveTo` folder; hosted download returns
`downloadUrl`/`downloadToken` for a GET with `x-download-token`, not a local save path.
Do not call `display_asset` in Claude Desktop (1MB response cap).
Save is for editable Workflow Builder drafts; publish is for API reuse. Ask if ambiguous.
Session saves use the latest accepted submission; pass both `session_id` and `batch_id` for
the exact run. Expired/missing snapshots are errors, never a reason to re-compose a replacement.
Draft inspection/reruns re-sign stored files with the user's access; report `draftAssetRefresh`
failures and `assetPersistence` reattachments. Photoshop action files are stored too but remain
original links in Builder and are not `inputs[]` slots; follow `needs_reattach` hints. Save a
corrected submission again after running, since earlier drafts are not updated automatically.
A rewired run's `postRun` requires offering save/publish/create_preset after completion;
ordinary input swaps do not.
## Further reading
- [Workflow Builder API docs](https://developer.adobe.com/firefly-services/docs/workflow-builder/)
- For direct Firefly REST calls without MCP, use the `firefly-api-specs` skill instead.
© adobe · Apache-2.0 · adobe/skills