← ClayCONTENT HISTORY
WHAT CHANGED
Update to Clay
Snapshot Nov 18, 2026 · 06:19 UTC · version 2.38.0
Collection source: downloaded plugin package.
First saved snapshot
No earlier snapshot is available to establish a change.
Save Clay and check future changes in Following.
Up to 3 plugins free · no card required.
Compare saved observations
Download comparison JSONFull technical diff · 0 changed fields
Full snapshot data
{
"description": "Clay workflows — discover available actions for workflow nodes (email lookup, company enrichment, phone finders, etc.) and inspect their input and output schemas. Use while building a workflow.",
"included_files": [
{
"relative_path": "catalog.md",
"size_in_bytes": 16629
},
{
"relative_path": "cost-and-budget.md",
"size_in_bytes": 7571
}
],
"name": "workflows-discover-actions",
"skill_md_contents": "---\nname: workflows-discover-actions\ndescription: Clay workflows — discover available actions for workflow nodes (email lookup, company enrichment, phone finders, etc.) and inspect their input and output schemas. Use while building a workflow.\nallowed-tools: Read, Grep\n---\n\n# Discover enrichments with search\n\nBefore using this skill, read the `clay` entry skill if you haven’t already this session, and follow its access guidance.\n\nUse this skill before choosing enrichments, connector actions, or reusable functions\nfor a workflow. Start with ranked search, inspect actual schemas, and use `catalog.md`\nfor unresolved requirements, connection discovery, or unavailable search.\n\nIf `clay workflows actions search` is unavailable or returns a workspace-enablement\nerror, use `catalog.md` instead. A failed search is not evidence that no matching\nenrichment exists. For tool access or authentication problems, read the `clay` entry\nskill before attempting discovery.\n\nFor pricing and spending decisions, read `cost-and-budget.md`.\n\nChoose the plan yourself from search results; do not call `actions recommend` in\nthis path.\n\nSearch returns individual actions, table functions (`type: \"function\"`), and workflow\nfunctions (`type: \"subworkflow\"`). “Workflow function” is the user-facing term for\n“subworkflow”; both refer to the same reusable workflow. Use “workflow function”\nwhen speaking with users, and recognize “subworkflow” as the same concept. Use\n“table function” to distinguish table-backed functions, and preserve technical\nidentifiers such as `subworkflow` and `subworkflow_call` in API payloads.\nOnly workflow functions can be called with a `subworkflow_call` node. Table\nfunctions use a `tool` node with a `tools` entry containing\n`toolType: \"clay_function\"` and `tableId` set to the function ID. Publishing an ordinary\nworkflow or registering it as a routine does not make it a workflow function. If node\ncreation or `clay functions get workflow:<id>` reports that a workflow is not a workflow\nfunction, tell the user it cannot be called as a workflow function instead of retrying.\nFunctions may encapsulate\nmultiple steps, including installed waterfalls. Compare both against the requested\ninputs and outputs; a suitable function can satisfy a capability without composing\nseveral actions.\n\nUse a matching Clay waterfall or Clay’s Enrich Person or Enrich Company action\nfor a lookup instead of an individual provider. A provider accepting fewer inputs\nis not a reason to bypass a matching Clay option: obtain the missing inputs first,\nand flag unresolved records if that fails. Reuse any requested information\nreturned along the way. Use individual providers for capabilities these Clay\noptions don’t cover, or when the user explicitly requests one.\n\nWhen native workflow waterfalls are enabled, search can also return `type: \"waterfall\"`.\nThese candidates contain `templateId`, its `attribute`, and ordered `providers`\n(`actionPackageId` + `actionKey`). Parent results also include `parentTemplateId` and use\ntheir default child; standalone results use their own template ID. They are template\ndescriptions, not function IDs or executable action IDs. Apply a candidate with\n`clay workflows nodes insert <workflowId> --input <json|file|->` using `step.kind: \"waterfall\"`,\n`step.name`, and `step.waterfallConfig`; read `clay workflows nodes insert --help` for its shape.\nDo not pass it to a function or action node, or expand it into separate provider nodes.\n\nBuild `waterfallConfig` from the candidate's `attribute`,\n`templateSelection: { parentTemplateId, templateId }` (omit `parentTemplateId` for standalone templates),\nand ordered `providers`. Give every\nprovider entry a unique `id` (including repeated providers), `authAccountId: null` for Clay's\nconnection or an accessible provider connection ID, and\n`inputOverrides: { inputMappingConfig: {}, inputRefs: {} }`.\nStart with `sharedInputs: { inputMappingConfig: {}, inputRefs: {} }` and `validation: null`.\nInspect the providers with `clay workflows actions schema <packageId> <actionKey>` to identify\ntheir required inputs and semantic types. Map shared inputs by semantic type, preserving upstream reference\nexpressions and `inputRefs`; each reference is `{ sourceNodeId, path }`, with `path`\nas a JSONPath such as `$.full_name`. Use provider parameter names for provider-specific overrides.\nPreserve the catalog's provider order and repeated entries. Use a unique node name based on\nthe current graph. The insertion creates a native waterfall without executing providers. Unmapped inputs are allowed when information is missing, but do not\nrun the workflow until the required inputs have been configured.\n\nFor a custom waterfall specified by an attribute and an ordered provider list, use the\nsame native insertion path. Omit `templateSelection` when no template was selected.\nInspect each provider's schema, preserve the requested order, and bind shared inputs\nand provider overrides from upstream nodes. Configure one waterfall node rather than\nseparate provider nodes or a newly published function. An explicit request for one\nprovider alone should use that provider's action instead.\n\nTo revise an existing native waterfall, read `clay workflows nodes get <workflowId> <nodeId>`\n(or the full graph) and modify `node.waterfallConfig`. Send the complete config using\n`clay workflows nodes update <workflowId> <nodeId> --input <json|file|->` with only the\n`waterfallConfig` field. Preserve the template selection, provider IDs and ordering, and any\nsettings not being changed. Shared inputs use semantic types; provider input overrides use\nprovider parameter names. Never update the generated call bindings or hidden child nodes directly.\n\n1. Identify the inputs available at the insertion point from the conversation and\n existing workflow inspection. For an existing workflow, use\n `clay workflows graph get <workflowId> --mode full` and inspect individual nodes\n or triggers as needed. Use only upstream outputs; fields on other branches or\n downstream nodes are not automatically available. For a blank workflow, use\n the inputs stated by the user. Unknown output shapes require inspection or\n deferred mapping; do not invent field paths.\n2. Search for each distinct capability with\n `clay workflows actions search \"<capability>\" --limit 10`. Each call searches one query.\n Keep unrelated steps in separate searches\n so a combined shortlist does not crowd out one part of the request.\n3. Save search responses to a local JSON file so large schemas do not hide later\n candidates. Use a distinct `enrichment-search-<capability>-<attempt>.json` filename\n for every search, retaining earlier files for composition and ID verification.\n Use only lowercase letters, digits, and hyphens in the capability and attempt names.\n Scan a compact view of every candidate before choosing or searching\n again:\n\n ```bash\n clay workflows actions search \"<capability>\" --limit 10 > enrichment-search-work-email-1.json\n jq '.candidates[] | {type, candidateId, name: (.displayName // .name), packageId, actionKey, functionId, workflowId, parentTemplateId, templateId, attribute, providers, description, score, creditCost, requiresApiKey, appAccountType, requiredInputNames, requiredInputCombinations, inputs: [.inputParameters[]? | {name, type: (.type // .typeSettings.type), optional}], functionInputs: ((.inputSchema.properties // {}) | keys), functionOutputs: ((.outputSchema.properties // {}) | keys), outputs: [.outputParameters[]? | {outputPath, semanticType}]}' enrichment-search-work-email-1.json\n ```\n\n The jq command is an optional compact-view convenience. If shell filtering\n requires approval, use the read-only file tool to inspect the saved response\n in bounded sections instead; do not require approval just to inspect results.\n Inspect complete schemas only for promising candidates, reusing the saved file\n or the type-specific inspection commands below rather than printing every candidate's full schema.\n\n Compare the output fields against each part of the request. Reuse the saved\n response for detailed inspection instead of repeating the same query.\n Inspect candidate inputs, outputs, credentials and limits. Compare complete\n plans, including steps to obtain missing inputs, rather than only actions that\n accept the starting fields.\n Search order is the default preference among candidates with comparable fit.\n Use this order as a relevance signal, not proof of provider accuracy.\n Choose a lower-ranked candidate for a concrete advantage in required outputs,\n evidence, freshness metadata, or coverage, and explain it.\n Match output semantics to the request: a title field alone does not establish\n current employment, and a location field must refer to the requested entity.\n For actions, use `clay workflows actions schema <packageId> <actionKey>` for more detail.\n Inspect table functions with `clay functions get function:<functionId>`.\n For workflow function (`type: \"subworkflow\"`) candidates, use the published `inputSchema`\n and `outputSchema` already returned by search. For additional implementation context,\n use `clay functions get workflow:<workflowId>` when available; it does not require routine registration.\n Inspection is rolled out separately from search. If it returns `not_found`, continue using\n the saved search candidate's published schemas and `workflowId`; do not retry inspection\n or discard the candidate solely because inspection is unavailable. Node validation still\n determines whether the function can be added.\n `candidateId` is only a search identifier. The returned `schema` is implementation context:\n table functions render their table columns/actions as text; workflow functions render the\n published graph as JSON text. Do not pass the implementation text as a node-creation payload.\n Create a `subworkflow_call` node using the candidate’s `workflowId` as `subworkflowId`;\n wire upstream inputs with `inputSchema` sourceNodeId/sourcePath and map values with `inputMappingConfig`.\n Top-level `inputMappingConfig` currently supports only `subworkflow_call` nodes, not code or conditional nodes.\n For tool nodes, keep mappings inside `tools[].inputMappingConfig`.\n Do not redefine its schemas or convert it to a `clay_function` tool.\n For table functions (`type: \"function\"`), inspect the saved candidate’s `inputSchema` and `outputSchema`;\n use `clay functions get <functionId>` before rejecting a promising function or\n adding steps when its outputs or validation behavior are unclear. If inspection\n establishes the requested validation, use the function's result without another\n validator for the same check. Add validation only for a missing requested guarantee,\n stale results, or an explicit independent check.\n Read input headers and alternatives before treating a field as required.\n When `requiredInputCombinations` is present and nonempty, satisfy every input in\n any one combination; the combinations override individual input optional flags.\n\n4. If the candidates leave a gap, change the search rather than repeating the same\n constraints in slightly different words. Try the missing capability; if that\n returns the same options, broaden to the parent capability or entity enrichment.\n When broadening, remove the narrow constraints from the query instead of\n appending them to a broader phrase. Search for the category of information or\n the entity itself, then inspect the returned schemas for the missing field.\n A general enrichment can contain the needed field even when a specialized action\n does not. Keep the user's requirements when checking the broader results.\n5. Consider whether results already found can work together. A list or history can\n be filtered or reduced with native workflow logic. An identifier or URL can feed\n a detail action when its schema supports that input. Check limits and missing\n matches; do not claim complete coverage when the available actions are capped.\n Native processing can transform already retrieved data; it cannot bypass a\n retrieval limit or fetch unseen rows. Do not replace a missing capability with\n an unspecified native join, pagination loop or other unverified operation.\n Name the supported retrieval step and retain its limitations.\n A category enum need not cover every concept the user requested. Check whether\n documented keyword search can retrieve the missing concept, then use supported\n semantic classification or deterministic filtering as appropriate. Classification\n can assess retrieved content; it cannot recover content the retrieval excluded.\n6. Select one or more grounded results and pass them to the existing workflow\n builder. Copy each `candidateId`, `packageId`, `actionKey` or `functionId`\n verbatim from the tool response; never reconstruct an identifier from memory.\n Immediately before handing off, extract the selected identifiers from the saved\n JSON with jq, or read the selected entries with the read-only file tool. `candidateId` identifies a search result,\n not a builder payload. For search and catalog selections, pass the extracted\n `actionKey` and `packageId` as the builder’s `actionKey` and `actionPackageId`\n for `toolType: \"clay_action\"`; pass `functionId` as `tableId` for\n `toolType: \"clay_function\"`. For `type: \"subworkflow\"`, create a `subworkflow_call`\n node with `subworkflowId` and optional `inputMappingConfig`. Copy the\n workflowId from the response. Output references use `$.result.<publishedOutput>`.\n Never select the caller itself. Set `subworkflowId` when creating the call; changing its target afterward is not supported.\n Check every handed-off identifier against this extracted list; a plausible-looking\n identifier is not sufficient.\n Explain the fit and any remaining gap. The user's multi-part request already\n permits a multi-step plan; ask only for genuinely missing facts or consequential\n choices that prevent proceeding.\n\nStop when the plan covers the request or changed searches stop adding useful\noptions. Do not exhaustively inspect every result or repeat searches just to fill a\nshortlist. Before declaring a capability unavailable, try a broader search as well\nas the specific one. A search error is a technical failure, not evidence that no\nmatching enrichment exists; report it separately from any catalog recovery.\nNo matches establish only that those searches found none, not that the workspace\nlacks the capability or has it disabled. State the discovery limit unless explicit\ncapability or enablement evidence supports a stronger conclusion.\n\nIf a specific and a broader search still leave a requirement unsupported, use\n`clay workflows actions list` to look for a complementary action, unless the session\nexplicitly restricts discovery to search only. Save and filter the catalog locally. Include output field names and paths in\nthe filter, not just action names or descriptions: a generically named enrichment\nmay already expose the missing facts. Inspect the matched fields and credit costs,\nthen verify promising candidates with the type-specific inspection commands above. Use this recovery for a\nconcrete gap. Subsequent searches should resolve missing capabilities, ambiguous\noutput semantics, or coverage limitations. Do not make an extra search or catalog\npass solely to find a cheaper or simpler plan. Listed credits may be reported,\nbut should not override stronger fit or drive exploration. When the ranked results\nsupport the requirements and no material ambiguity remains, stop. Explain any\nrequirement that remains unsupported.\n\nCheck requiresApiKey before requiring a provider connection (for catalog selections,\nread actionLabels.requiresApiKey on the selected action). A false value supports\nsaying no customer-supplied provider credentials are required; null or missing is\nunknown, not false. A function's managed source identifies who maintains it, not\nits credential requirements. Keep unknown prerequisites qualified rather than\nclaiming a connection is required or unnecessary without evidence.\nWhen a connection is required,\nor the user requests their own credentials, use the action's appAccountType with\n`clay app-accounts list --filter type=<appAccountType>`.\nReuse the returned accounts for other selected actions with the same connection type.\nIf multiple accounts match and the user's intent or existing workflow does not identify\nwhich to use, ask. For the workspace's own credentials, exclude isClayManaged accounts.\nPass the chosen account's id explicitly as appAccountId, even when passing a toolId.\nIf no accessible accounts are returned, explain that a connection or access is needed.\nDo not fetch the full catalog for credential checks.\n\nUnknown schemas do not prove\nan action cannot do the work: distinguish a promising candidate from verified field\nmappings. For table lookups and research actions, an empty declared output schema means\nthe builder must resolve dynamic outputs before mapping; it does not make the\naction unavailable. Clay's table-lookup actions are valid built-in operations,\nnot unnecessary external enrichments. When one supports the requested cross-table\nread, select it and defer only its dynamic field mapping; do not substitute an\nunspecified native join.\nDo not execute actions merely to rank them or invent output paths.\n\nUse the existing workflow-building instructions for node construction and data mapping. Resolve dynamic\nfields when needed, preserve workflow capability and credential checks, and keep\nuncertain results distinct from confirmed outputs. Search discovers candidates;\nthe builder remains responsible for creating and validating the workflow.\n\nWhen the remaining gap is evidence rather than a missing field, search for a\nresearch or source-reading action that can perform the verification. Do not stop\nat an unverified numeric field plus an unspecified “verify later” step. Select\na supported verification capability, require appropriate source evidence, and\nkeep unverifiable results unknown. This does not guarantee universal coverage.\nDo not add research to ordinary lookups that do not have this evidence gap.\n"
}SHA-256 of public snapshot: 10d9fd84dcead1d225ee3aac74dca27c5149e85d0bd29343eec312edd510557d