← ClayCONTENT HISTORY

Update to Clay

Snapshot Nov 18, 2026 · 06:19 UTC · version 2.38.0

Collection source: downloaded plugin package.

WHAT CHANGED

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 JSON
Full technical diff · 0 changed fields
Full snapshot data
{
  "description": "Build and run Claygent agent nodes in Clay workflows. Use when the user asks for custom LLM research, reasoning, open-ended classification that needs research, drafting, or summarization over records; when no existing routine fits a Claygent task; or when creating or editing a workflow `agent` node.",
  "included_files": [],
  "name": "workflows-claygent",
  "skill_md_contents": "---\nname: workflows-claygent\ndescription: Build and run Claygent agent nodes in Clay workflows. Use when the user asks for custom LLM research, reasoning, open-ended classification that needs research, drafting, or summarization over records; when no existing routine fits a Claygent task; or when creating or editing a workflow `agent` node.\n---\n\n# Claygents in workflows\n\nBefore using this skill, read the `clay` entry skill if you haven’t already this session, and follow its access guidance.\n\nA Claygent in a workflow is a dedicated `nodeType: \"agent\"` node. It is not a Clay action\nor function, so its absence from `clay workflows actions list` does not make Claygent\nunavailable. Use `/workflows-discover-actions` only for tool nodes.\n\nTo sort each record into one of a fixed set of categories from data the workflow already has, or\nwhen the user asks for Jev, use a Classify step instead: a Use AI tool node whose inputs are listed\nin `clay workflows nodes insert --help`. Jev is a Classify model, not a Claygent model.\n\nIf an existing routine already performs the requested task, use the `/routines` skill.\nOtherwise build the Claygent workflow here. Read the main `/workflows` skill for graph,\npublishing, testing, and presentation rules.\n\n## Create or reuse the Claygent\n\n- To reuse a regular Claygent, find it with `clay claygents list --search <text>` and read it\n  with `clay claygents get <id>` to check its inputs, outputs, and tools. Then set only\n  `agentClaygentId` on the node. Do not also send its name, prompt, model, or outputs.\n- To create one, send `agentName` and `agentPrompt` together in the agent node's first\n  `clay workflows nodes create` call. Omit `agentModel` to use the workspace default (Auto\n  where enabled). Set `agentModel` only when the user requests a specific model, and do not\n  silently substitute a different model. When editing an existing node, omit `agentModel`\n  unless the user asks to change it so the saved model is preserved.\n- Put every prompt variable in top-level `inputSchema.properties`. Wire upstream values with\n  `sourceNodeId` and `sourcePath`, then use the property name as `{{variable}}` in the prompt.\n- Set `outputSchema` to a flat field map with a useful description on every field. A\n  `select` field's `options` value is a JSON-encoded string containing a non-empty array of\n  objects with non-empty `text`, for example:\n\n```json\n{\n  \"signal\": {\n    \"type\": \"select\",\n    \"description\": \"Strength of the evidence\",\n    \"options\": \"[{\\\"text\\\":\\\"Strong\\\"},{\\\"text\\\":\\\"Weak\\\"},{\\\"text\\\":\\\"None\\\"}]\"\n  }\n}\n```\n\nIf a write rejects an output field or schema, use the validation message to correct it\nbefore testing. Do not retry the same shape.\n\nIf the user asks for a numeric score within a bounded set of ranges (for example \"80–100\nstrong, 35–65 mixed, 0–25 weak\"), the Claygent prompt must say that the score has to fall\ninside one of those ranges.\n\n## Connected tools\n\nA Claygent's tools — MCP servers such as Linear or Notion, routines, document sources, and\nweb search — are configured on the Claygent, not on the workflow node. An agent node's\n`tools` field is always empty, so it says nothing about what the Claygent can use. Read\n`clay claygents get <id>` and check `tools`, `internetSearchEnabled`, and\n`peopleAndCompaniesSearchEnabled`.\n\nAn agent node's `agentInternetSearchEnabled` shows whether its Claygent can search the web.\nSet it with `clay workflows nodes update` to turn web search on or off for a regular\nClaygent; the change applies everywhere that Claygent runs. Clay models always search the\nweb, so web search cannot be turned off for them.\n\nWhen the agent needs a resource behind authentication, such as a private Linear attachment,\nfirst look for a connected tool that already holds that account's credentials, and tell the\nClaygent in its prompt to use it. Do not build a separate fetch-and-rehost chain of workflow\nnodes for something a connected tool can reach. Connecting a new tool is done in the\nClaygent's settings in Clay.\n\n## Choose the source\n\n- For a bounded Clay Search population, use a Find leads trigger and follow the trigger\n  guidance in the main `/workflows` skill. Preserve the latest search query and its explicit\n  overall limit; never use an estimated result count as the run limit or turn search results\n  into workflow inputs by hand.\n- For Audiences records, use the trigger/backfill guidance in the main `/workflows` skill.\n- For explicit inputs supplied by the caller, use a manual trigger and wire its input schema.\n\nEach Find Leads or Audiences source record starts one workflow run, so the downstream\nClaygent reads one person's or company's fields directly from the trigger node.\n\n## Research responsibility\n\nWhen the task requires public evidence, put the research criteria and source expectations in\nthe Claygent prompt. The Claygent performs the per-record web research during its run; the\nhost agent should not manually research every record before launching it.\n\nWhen the user needs an exact criterion checked by the research — a funding event's date, a\nlaunch in the last 12 months — have the Claygent return it as an output field, left empty when\nthe evidence does not establish it rather than guessed. Add a conditional that stops\nnon-matching records only when the user asked to keep just the matches (\"only companies\nthat…\", \"skip anyone without…\"); otherwise every record continues with the field as returned.\nDo not claim the Claygent's future research has already verified it.\n\n## Test, run, and deliver\n\n1. Validate the graph.\n2. Test the Claygent on one representative record before launching the batch. Inspect the\n   actual node output and correct schema, model, prompt, or tool failures first.\n3. Bound and preflight the full run using the source query limit and the cost policy from\n   `workflows-discover-actions/cost-and-budget.md`.\n4. Publish when the source trigger requires the live graph, start the source, and record the\n   launch time.\n5. List the runs created by that trigger, follow pagination, and wait for terminal states.\n   Read actual Claygent outputs with `clay workflows runs get`; use `--node-id` or\n   `--verbose` when needed.\n6. Build requested CSVs or reports from completed run outputs, not Search preview rows.\n   Include completed/failed counts and identify failed records in partial deliveries.\n"
}

SHA-256 of public snapshot: e7266affec0d761337a1a0db17a3ada8404ce0d1afff9fe1b357b88bdcc6fb5d