← falCONTENT HISTORY

Update to fal

Snapshot Nov 14, 2026 · 00:10 UTC · version 1.3.2

Collection source: downloaded plugin package.

WHAT CHANGED

Instructions updated for fal-workflow

Skill instructions — Now says: Bundled guidance is fallback craft. For inputs, recommend_model's `input_contract` and live validation at submission are authoritative. See [the shared...

Observed in instructions or declared skills. Runtime behavior has not been tested.

Save fal and check future changes in Following.

Up to 3 plugins free · no card required.

All fields and original values.

Skill instructions

Before

Nov 1, 2026 · 00:09 UTC

- Cloud workflow JSON: use when the user wants a reusable graph asset. - MCP-orchestrated workflow: use when the client should run a sequence now through the fal.ai MCP tools. 4. Capture dependencies: which output URL feeds which next ...

After

Nov 14, 2026 · 00:10 UTC

Bundled guidance is fallback craft. For inputs, recommend_model's `input_contract` and live validation at submission are authoritative. See [the shared production contract](../shared/production-contract.md). - Read-only prepared graph: `...

Compare saved observations

Download comparison JSON

SKILL.md text changes

− removed · + added

--- before
+++ after
@@ -8,15 +8,20 @@
 
 # fal.ai Workflows
 
+Bundled guidance is fallback craft. For inputs, recommend_model's
+`input_contract` and live validation at submission are authoritative. See
+[the shared production contract](../shared/production-contract.md).
+
 Use this skill when a single model call is not enough. A workflow is a planned
 sequence of model calls with dependencies, retries, quality gates, and a final
 manifest.
 
 ## Workflow Modes
 
-- Cloud workflow JSON: use when the user wants a reusable graph asset.
-- MCP-orchestrated workflow: use when the client should run a sequence now through
-  the fal.ai MCP tools.
+- Read-only prepared graph: `prepare_media_workflow` checks at most eight
+  exact nodes and explicit typed edges. It does not run or save a cloud graph.
+- Host-orchestrated workflow: the client executes ready nodes through existing
+  MCP tools and binds observed outputs before dependent submission.
 
 Do not call genmedia CLI from this plugin. If a source skill mentions CLI
 orchestration, translate that intent into MCP tool usage.
@@ -27,17 +32,28 @@
 2. Split into nodes: input assets, generation, utility transforms, edits,
    video/audio/3D steps, and final packaging.
 3. Choose model per node with `model-routing`.
-4. Capture dependencies: which output URL feeds which next input.
-5. Inspect schema for each endpoint before building payloads.
-6. Run independent expensive nodes async.
+4. Capture dependencies with canonical raw-result JSON Pointers such as
+   `/images/1/url` and exact private job handles. Keep account routing receipts
+   in those handles; do not export them as public links.
+5. Build each node's input from its `input_contract`.
+6. Call `prepare_media_workflow` to check schemas and edge cardinality. Pending
+   bindings stay unresolved; use `resolve_model_inputs` after source completion.
+   Run independent authorized expensive nodes async.
 7. Poll async nodes without restarting completed work.
-8. Gate each step: preserve identity/product, check duration, check format.
-9. Return a manifest: node, endpoint, request id, inputs, output URLs, defects.
+8. Carry original/approved/rejected sources and 1–6 preservation requirements.
+   Read actual reference/output pixels with `inspect_images` or host vision.
+   Failed or uncertain results stop dependent work; inspect before fan-out.
+   Probe requested duration/format/audio facts with a host decoder; missing
+   capability stays `unverified`, even if the model job completed.
+9. Use `prepare_deliverables` for exact output selectors, source roles,
+   expected/observed facts and attributed inspection/save outcomes. If the user
+   asked to save, `save_job_outputs` registers selected/all completed originals.
+   Preserve partial successes and read unknown writes before any retry.
 
 ## Useful Patterns
 
 - Anchor first: generate or upload a strong still, then drive edits/video from
-  that approved asset.
+  that asset.
 - Fan-out: create several variants from the same anchor, then select one.
 - Fan-in: combine product, logo, character, or style references into an edit.
 - Frame bridge: generate first and last frame, then use image-to-video or
@@ -65,8 +81,11 @@
 
 ## Failure Handling
 
-- 422/schema error: re-inspect schema and remove guessed fields.
-- Product/identity drift: switch to reference/edit workflow.
+- 422/schema error: fix the input from the returned field errors and remove
+  guessed fields.
+- Product/identity drift: retain original conditioning sources. Only one
+  targeted repair is permitted within explicit existing authorization/budget;
+  record repairs_used. A failed permitted repair stops dependent generation.
 - Video still running: keep polling the same request id.
 - Bad prompt alignment: reduce prompt scope and change one variable at a time.
 - Cost spike: reduce variants, duration, resolution, or use draft endpoints.
Full technical diff · 1 changed fields

changed /skill_md_contents

BEFORE
"---\nname: fal-workflow\ndescription: >\n  Design multi-step fal.ai media workflows for MCP execution. Use when the user\n  wants model chaining, batch generation, fan-out/fan-in, frame bridging,\n  narrated video, dataset creation, reusable pipeline logic, or workflow JSON.\n---\n\n# fal.ai Workflows\n\nUse this skill when a single model call is not enough. A workflow is a planned\nsequence of model calls with dependencies, retries, quality gates, and a final\nmanifest.\n\n## Workflow Modes\n\n- Cloud workflow JSON: use when the user wants a reusable graph asset.\n- MCP-orchestrated workflow: use when the client should run a sequence now through\n  the fal.ai MCP tools.\n\nDo not call genmedia CLI from this plugin. If a source skill mentions CLI\norchestration, translate that intent into MCP tool usage.\n\n## Planning Pattern\n\n1. Define final deliverable and acceptance criteria.\n2. Split into nodes: input assets, generation, utility transforms, edits,\n   video/audio/3D steps, and final packaging.\n3. Choose model per node with `model-routing`.\n4. Capture dependencies: which output URL feeds which next input.\n5. Inspect schema for each endpoint before building payloads.\n6. Run independent expensive nodes async.\n7. Poll async nodes without restarting completed work.\n8. Gate each step: preserve identity/product, check duration, check format.\n9. Return a manifest: node, endpoint, request id, inputs, output URLs, defects.\n\n## Useful Patterns\n\n- Anchor first: generate or upload a strong still, then drive edits/video from\n  that approved asset.\n- Fan-out: create several variants from the same anchor, then select one.\n- Fan-in: combine product, logo, character, or style references into an edit.\n- Frame bridge: generate first and last frame, then use image-to-video or\n  supported controls to bridge motion.\n- Audio-first: generate/upload voice first, then match visuals to duration.\n- Utility cleanup: background removal, compression, upscaling, format\n  conversion, segmentation, or OCR as deterministic nodes.\n\n## Local File Inputs\n\n- With hosted MCP and local file/HTTP access, call `upload_file` with\n  `prepare_upload: true`, the basename as `file_name`, and the actual byte count\n  as `file_size` (up to 90MB). This uses the already connected fal account.\n- PUT the actual file bytes to the returned `upload_url` with its returned\n  headers. For example, a local HTTP client or `curl --upload-file` can stream\n  the file without putting its contents into the conversation. Keep the signed\n  URL private; send no Authorization header and do not follow redirects.\n- `status: upload_required` means no bytes have been uploaded yet. Only use\n  `file_url` as model input after the PUT returns 2xx. If the PUT fails, stop\n  before generation; do not guess that a file exists at the returned URL.\n- Never pass a local path to hosted MCP, copy base64 through chat, or obtain a\n  different API key to work around a rejection. If local file/HTTP access is\n  unavailable, ask for a public file URL. Stdio supports `file_path` when the\n  server's local-file allowlist permits it.\n\n## Failure Handling\n\n- 422/schema error: re-inspect schema and remove guessed fields.\n- Product/identity drift: switch to reference/edit workflow.\n- Video still running: keep polling the same request id.\n- Bad prompt alignment: reduce prompt scope and change one variable at a time.\n- Cost spike: reduce variants, duration, resolution, or use draft endpoints.\n"
AFTER
"---\nname: fal-workflow\ndescription: >\n  Design multi-step fal.ai media workflows for MCP execution. Use when the user\n  wants model chaining, batch generation, fan-out/fan-in, frame bridging,\n  narrated video, dataset creation, reusable pipeline logic, or workflow JSON.\n---\n\n# fal.ai Workflows\n\nBundled guidance is fallback craft. For inputs, recommend_model's\n`input_contract` and live validation at submission are authoritative. See\n[the shared production contract](../shared/production-contract.md).\n\nUse this skill when a single model call is not enough. A workflow is a planned\nsequence of model calls with dependencies, retries, quality gates, and a final\nmanifest.\n\n## Workflow Modes\n\n- Read-only prepared graph: `prepare_media_workflow` checks at most eight\n  exact nodes and explicit typed edges. It does not run or save a cloud graph.\n- Host-orchestrated workflow: the client executes ready nodes through existing\n  MCP tools and binds observed outputs before dependent submission.\n\nDo not call genmedia CLI from this plugin. If a source skill mentions CLI\norchestration, translate that intent into MCP tool usage.\n\n## Planning Pattern\n\n1. Define final deliverable and acceptance criteria.\n2. Split into nodes: input assets, generation, utility transforms, edits,\n   video/audio/3D steps, and final packaging.\n3. Choose model per node with `model-routing`.\n4. Capture dependencies with canonical raw-result JSON Pointers such as\n   `/images/1/url` and exact private job handles. Keep account routing receipts\n   in those handles; do not export them as public links.\n5. Build each node's input from its `input_contract`.\n6. Call `prepare_media_workflow` to check schemas and edge cardinality. Pending\n   bindings stay unresolved; use `resolve_model_inputs` after source completion.\n   Run independent authorized expensive nodes async.\n7. Poll async nodes without restarting completed work.\n8. Carry original/approved/rejected sources and 1–6 preservation requirements.\n   Read actual reference/output pixels with `inspect_images` or host vision.\n   Failed or uncertain results stop dependent work; inspect before fan-out.\n   Probe requested duration/format/audio facts with a host decoder; missing\n   capability stays `unverified`, even if the model job completed.\n9. Use `prepare_deliverables` for exact output selectors, source roles,\n   expected/observed facts and attributed inspection/save outcomes. If the user\n   asked to save, `save_job_outputs` registers selected/all completed originals.\n   Preserve partial successes and read unknown writes before any retry.\n\n## Useful Patterns\n\n- Anchor first: generate or upload a strong still, then drive edits/video from\n  that asset.\n- Fan-out: create several variants from the same anchor, then select one.\n- Fan-in: combine product, logo, character, or style references into an edit.\n- Frame bridge: generate first and last frame, then use image-to-video or\n  supported controls to bridge motion.\n- Audio-first: generate/upload voice first, then match visuals to duration.\n- Utility cleanup: background removal, compression, upscaling, format\n  conversion, segmentation, or OCR as deterministic nodes.\n\n## Local File Inputs\n\n- With hosted MCP and local file/HTTP access, call `upload_file` with\n  `prepare_upload: true`, the basename as `file_name`, and the actual byte count\n  as `file_size` (up to 90MB). This uses the already connected fal account.\n- PUT the actual file bytes to the returned `upload_url` with its returned\n  headers. For example, a local HTTP client or `curl --upload-file` can stream\n  the file without putting its contents into the conversation. Keep the signed\n  URL private; send no Authorization header and do not follow redirects.\n- `status: upload_required` means no bytes have been uploaded yet. Only use\n  `file_url` as model input after the PUT returns 2xx. If the PUT fails, stop\n  before generation; do not guess that a file exists at the returned URL.\n- Never pass a local path to hosted MCP, copy base64 through chat, or obtain a\n  different API key to work around a rejection. If local file/HTTP access is\n  unavailable, ask for a public file URL. Stdio supports `file_path` when the\n  server's local-file allowlist permits it.\n\n## Failure Handling\n\n- 422/schema error: fix the input from the returned field errors and remove\n  guessed fields.\n- Product/identity drift: retain original conditioning sources. Only one\n  targeted repair is permitted within explicit existing authorization/budget;\n  record repairs_used. A failed permitted repair stops dependent generation.\n- Video still running: keep polling the same request id.\n- Bad prompt alignment: reduce prompt scope and change one variable at a time.\n- Cost spike: reduce variants, duration, resolution, or use draft endpoints.\n"
Full snapshot data
{
  "description": "Design multi-step fal.ai media workflows for MCP execution. Use when the user wants model chaining, batch generation, fan-out/fan-in, frame bridging, narrated video, dataset creation, reusable pipeline logic, or workflow JSON.\n",
  "included_files": [],
  "name": "fal-workflow",
  "skill_md_contents": "---\nname: fal-workflow\ndescription: >\n  Design multi-step fal.ai media workflows for MCP execution. Use when the user\n  wants model chaining, batch generation, fan-out/fan-in, frame bridging,\n  narrated video, dataset creation, reusable pipeline logic, or workflow JSON.\n---\n\n# fal.ai Workflows\n\nBundled guidance is fallback craft. For inputs, recommend_model's\n`input_contract` and live validation at submission are authoritative. See\n[the shared production contract](../shared/production-contract.md).\n\nUse this skill when a single model call is not enough. A workflow is a planned\nsequence of model calls with dependencies, retries, quality gates, and a final\nmanifest.\n\n## Workflow Modes\n\n- Read-only prepared graph: `prepare_media_workflow` checks at most eight\n  exact nodes and explicit typed edges. It does not run or save a cloud graph.\n- Host-orchestrated workflow: the client executes ready nodes through existing\n  MCP tools and binds observed outputs before dependent submission.\n\nDo not call genmedia CLI from this plugin. If a source skill mentions CLI\norchestration, translate that intent into MCP tool usage.\n\n## Planning Pattern\n\n1. Define final deliverable and acceptance criteria.\n2. Split into nodes: input assets, generation, utility transforms, edits,\n   video/audio/3D steps, and final packaging.\n3. Choose model per node with `model-routing`.\n4. Capture dependencies with canonical raw-result JSON Pointers such as\n   `/images/1/url` and exact private job handles. Keep account routing receipts\n   in those handles; do not export them as public links.\n5. Build each node's input from its `input_contract`.\n6. Call `prepare_media_workflow` to check schemas and edge cardinality. Pending\n   bindings stay unresolved; use `resolve_model_inputs` after source completion.\n   Run independent authorized expensive nodes async.\n7. Poll async nodes without restarting completed work.\n8. Carry original/approved/rejected sources and 1–6 preservation requirements.\n   Read actual reference/output pixels with `inspect_images` or host vision.\n   Failed or uncertain results stop dependent work; inspect before fan-out.\n   Probe requested duration/format/audio facts with a host decoder; missing\n   capability stays `unverified`, even if the model job completed.\n9. Use `prepare_deliverables` for exact output selectors, source roles,\n   expected/observed facts and attributed inspection/save outcomes. If the user\n   asked to save, `save_job_outputs` registers selected/all completed originals.\n   Preserve partial successes and read unknown writes before any retry.\n\n## Useful Patterns\n\n- Anchor first: generate or upload a strong still, then drive edits/video from\n  that asset.\n- Fan-out: create several variants from the same anchor, then select one.\n- Fan-in: combine product, logo, character, or style references into an edit.\n- Frame bridge: generate first and last frame, then use image-to-video or\n  supported controls to bridge motion.\n- Audio-first: generate/upload voice first, then match visuals to duration.\n- Utility cleanup: background removal, compression, upscaling, format\n  conversion, segmentation, or OCR as deterministic nodes.\n\n## Local File Inputs\n\n- With hosted MCP and local file/HTTP access, call `upload_file` with\n  `prepare_upload: true`, the basename as `file_name`, and the actual byte count\n  as `file_size` (up to 90MB). This uses the already connected fal account.\n- PUT the actual file bytes to the returned `upload_url` with its returned\n  headers. For example, a local HTTP client or `curl --upload-file` can stream\n  the file without putting its contents into the conversation. Keep the signed\n  URL private; send no Authorization header and do not follow redirects.\n- `status: upload_required` means no bytes have been uploaded yet. Only use\n  `file_url` as model input after the PUT returns 2xx. If the PUT fails, stop\n  before generation; do not guess that a file exists at the returned URL.\n- Never pass a local path to hosted MCP, copy base64 through chat, or obtain a\n  different API key to work around a rejection. If local file/HTTP access is\n  unavailable, ask for a public file URL. Stdio supports `file_path` when the\n  server's local-file allowlist permits it.\n\n## Failure Handling\n\n- 422/schema error: fix the input from the returned field errors and remove\n  guessed fields.\n- Product/identity drift: retain original conditioning sources. Only one\n  targeted repair is permitted within explicit existing authorization/budget;\n  record repairs_used. A failed permitted repair stops dependent generation.\n- Video still running: keep polling the same request id.\n- Bad prompt alignment: reduce prompt scope and change one variable at a time.\n- Cost spike: reduce variants, duration, resolution, or use draft endpoints.\n"
}

SHA-256 of public snapshot: 0af5d0de9124fd891dec4199d6724399e5ff7fc9e2aaae2c7bb69f426ca6e10f