← 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 — plan, review, build, and edit automations with the Clay CLI. Use when planning or working on a Clay workflow, including putting one on a schedule or pausing its schedule.",
"included_files": [
{
"relative_path": "account-agents.md",
"size_in_bytes": 17339
},
{
"relative_path": "audiences.md",
"size_in_bytes": 14113
},
{
"relative_path": "csv-triggers.md",
"size_in_bytes": 6606
},
{
"relative_path": "data-passing.md",
"size_in_bytes": 28967
},
{
"relative_path": "enrichments.md",
"size_in_bytes": 9617
},
{
"relative_path": "industry-mapping.md",
"size_in_bytes": 1693
},
{
"relative_path": "node-groups.md",
"size_in_bytes": 3210
},
{
"relative_path": "presenting.md",
"size_in_bytes": 3441
},
{
"relative_path": "publishing.md",
"size_in_bytes": 7449
},
{
"relative_path": "repeat-function-calls.md",
"size_in_bytes": 4257
},
{
"relative_path": "run-analysis.md",
"size_in_bytes": 4208
},
{
"relative_path": "table-primitives.md",
"size_in_bytes": 3710
},
{
"relative_path": "table-waterfalls.md",
"size_in_bytes": 2436
},
{
"relative_path": "testing.md",
"size_in_bytes": 19255
},
{
"relative_path": "trigger-selection.md",
"size_in_bytes": 4736
}
],
"name": "workflows",
"skill_md_contents": "---\nname: workflows\ndescription: Clay workflows — plan, review, build, and edit automations with the Clay CLI. Use when planning or working on a Clay workflow, including putting one on a schedule or pausing its schedule.\n---\n\n# Building Clay workflows with the CLI\n\nBefore using this skill, read the `clay` entry skill if you haven’t already this session, and follow its access guidance.\n\nWhen the conversation includes selected workflow nodes from a widget, read the\ncurrent authoritative workflow graph before editing. Reconcile the workflow and\nworkspace IDs, revision, and selected node IDs; names and selection order do not\nidentify a connection. For “between these”, use the directed edge identity and\nbranch/source index. A unique direct connection is straightforward. Ask a focused\nquestion for multiple edges or paths, disconnected or deleted nodes, or incomplete\ntopology. A truncated preview cannot establish that no other path exists. Never\ninfer configuration or execution inputs from the display-only context.\n\nSupporting files in this directory: `publishing.md`, `testing.md`, `data-passing.md`,\n`presenting.md`, `audiences.md`, [enrichments.md](./enrichments.md), `account-agents.md`, `run-analysis.md`, `node-groups.md`,\n`csv-triggers.md`, `table-primitives.md`, `table-waterfalls.md`, `industry-mapping.md`,\n`repeat-function-calls.md`.\n\nRead this skill before proposing a workflow plan, even when the user asks for no workspace\nqueries or changes. For Audiences destinations or field mappings, also read `audiences.md`\nbefore answering. Reading skill guidance does not query or change the workspace.\n\nWhen working in an existing workflow, read it with `clay workflows get <workflowId>` before\nplanning edits, unless the user prohibits queries; then plan from the supplied draft and field\ncatalog. Do not ask for a workflow ID just to discuss a complete supplied draft. If its `type`\nis exactly `audience_enrichment`, read the complete\n`/workflows-audience-enrichment` skill and follow it together with this skill. Do not apply that\nspecialized skill to a workflow whose type is null, absent, or any other value.\n\nFor requests to create, inspect, edit, enable, disable, or delete a workflow alert rule, or to\nsubscribe to a workflow's alert emails, read and follow the `/workflows-alerts` skill.\n\n## Domain and working style\n\nYou are helping users build and edit Clay workflows.\n\n- **Keep the build in one workflow by default.** Treat sourcing, similarity discovery,\n qualification, finding contacts, enrichment, and routing for the same requested build as\n stages of one workflow. Extend the workflow the user has open or the one already built in\n this conversation, including follow-up requests, unless the user asks for a separate workflow.\n Do not create another workflow just to organize stages, isolate testing, or handle a list.\n Use connected nodes, branches, Repeat, and canvas groups where available. If a capability\n requires a separate workflow or workflow function, explain the constraint and get the user’s\n agreement before creating it; do not silently split the build.\n- **Name new workflows first.** When the current workflow is blank and has a placeholder name,\n choose a concise, descriptive name from the user's initial request. Immediately before renaming,\n run `clay workflows get <workflowId>` and inspect its current `name`; only run\n `clay workflows update <workflowId> --name \"<name>\"` if the name is still exactly\n `Untitled workflow`. Preserve any name that has already replaced the placeholder. Do this before\n asking follow-up questions or editing the graph. When creating a workflow, choose the name before\n running `clay workflows create`. Do not rename an existing non-blank workflow unless the user asks\n you to.\n- Follow the host agent's planning, user-input, and approval policies when building or editing workflows.\n- **Do not add Audiences writeback to an Audiences export workflow.** For workflows that\n start with an Audiences-related trigger and export its records, omit Write to Audiences\n to avoid a redundant no-op loop. This takes precedence over the default writeback guidance\n below; see `audiences.md`.\n- **Choose Audiences writeback from the whole workflow and the user's intent.** Inspect the\n source, upstream steps, outputs, and existing destinations. Enrichment of existing Audiences\n records should write results back; for new records from searches or CSV uploads, recommend\n saving to Audiences and confirm before adding it unless already requested. Honor explicit\n destinations and no-save requests, and reuse existing writeback nodes. Read `audiences.md`\n for the decision and field-mapping rules: map obvious existing fields directly without\n offering duplicate fields, clarify ambiguous mappings, and create needed fields without\n asking permission. A missing compatible field is a reason to create one, not a decision\n to hand back to the user. In a plan, state that creation and mapping as the next steps.\n Building a node does not authorize running or publishing.\n- **Ground provider names in the workspace catalog.** Until you search the current workspace's\n action catalog, describe capabilities generically (for example, \"intent data enrichment\"). Do\n not introduce provider or action names from general knowledge. You may name one when the user\n already named it or the catalog verified it. Never proactively mention G2 or 6sense; if the user\n asks for either, search the catalog and recommend verified alternatives instead of presenting it\n as a supported Clay action.\n- **Narrate and visualize as you go.** After each meaningful change, say what changed and show the\n graph (`presenting.md`).\n- **Ask when there's a real choice** — refer to actions by human-readable names, never bare\n `actionKey`s.\n- **Qualify the person's role, not just title keywords.** Require the right function and\n decision-making seniority. Mentioning a function may mean selling to, recruiting for, or\n assisting its leaders. Use compatible structured function/seniority fields or unambiguous\n rules; when titles need judgment, use a small Claygent classification of supplied inputs\n before paid contact enrichment. Select existing candidate IDs and reuse the decision\n downstream; don't generate people or research externally when supplied fields suffice.\n- **Classify company verticals with Clay's industry mapping.** When a step decides a company's\n vertical from its `industry` field (for example \"is this a software company?\") and the\n vertical is in `industry-mapping.md`, use its `industry` values instead of writing your own\n keyword match. For any other vertical, choose the matching values from the labels the records\n actually carry (`clay audiences fields list-values industry --entity-type companies`), and\n use a small Claygent classification when the label alone can't decide.\n\nCommon node types (not just agent/tool — pick the type that fits the step):\n\n- **Classify steps** — a Use AI tool node with `useCase` `decision` that sorts each record into\n one of 2–255 fixed categories, with branch-ready `$.result.categoryId`. Use it for fixed-category\n labels from supplied data and whenever the user asks for Jev (`jev-1.13.0`). There is no\n `classify` nodeType; follow the recipe in `clay workflows nodes insert --help`.\n- **Claygent (agent) nodes** (`nodeType: \"agent\"`) — LLM loops with prompts (reasoning,\n research, drafting, summarizing, open-ended classification). Read and follow the complete `/workflows-claygent`\n skill before creating, editing, or running one.\n - **Create:** send `agentName` and `agentPrompt` **together in the same** `--input` —\n separate calls can persist a blank prompt. Omit `agentModel` to use the workspace default\n (Auto where enabled), or set it to the specific model the user requests. One exception: a node\n you are about to make an account agent is created bare — see account agents below.\n - **Reuse:** `clay claygents list --search <text>` returns regular Claygents (`agentType: null`)\n and account agents (`agentType: \"account\"`). Read regular Claygents with `clay claygents get <id>`.\n Send only `agentClaygentId` (the returned `id`) to attach either kind, then inspect the linked\n agent with `clay workflows nodes get <workflowId> <nodeId>`. Do not send `agentName` / prompt / model — `agentName` can rewrite that\n shared prompt. For account agents, follow `account-agents.md` for selection and input wiring.\n- **Tool nodes** (`nodeType: \"tool\"`) — run a single Clay action or Clay function. (UI may\n label these \"Enrich\" or \"Function\".) `tools` may be omitted while drawing an outline; before\n validate/test/publish, set exactly one `tools` entry — either\n `{ \"toolType\": \"clay_action\", \"actionKey\": …, \"actionPackageId\": … }` for an enrichment action,\n or `{ \"toolType\": \"clay_function\", \"tableId\": \"<functionId>\" }` for a Clay function subroutine.\n Both come from `clay workflows actions list`. For the input shape, pass an action's ids to\n `clay workflows actions schema` or a function's id to `clay functions get`.\n- **Message nodes** (`nodeType: \"tool\"` with `message`) — a message template with input variables\n and optional AI-written text. See \"Message nodes\" below.\n- **Rep task nodes** (`nodeType: \"tool\"` with `repTask`) — create a task in Reps for a rep to work,\n such as an email to review and send. See \"Rep task nodes\" below.\n- **Conditional nodes** (`nodeType: \"conditional\"`) — pick exactly one outgoing branch\n (`rules` / `agentic` / `code` via `conditionalMode`).\n- **Code nodes** (`nodeType: \"code\"`) — deterministic Python transforms (no LLM), for shaping\n and filtering; use an agent when the step needs LLM judgment. The handler goes in top-level\n `code`. Code runs on the Python standard library only, with no network access and no Clay\n actions, and each run is stopped after 1 second: use a tool node for enrichment or HTTP calls\n and a delay node to wait, never `time.sleep()`. Standard-library modules that need networking,\n processes, threads, files, or time zone data are unavailable (`socket`, `http`,\n `urllib.request`, `subprocess`, `threading`, `asyncio`, `logging`, `pathlib`, `zoneinfo`). Use\n `print()` instead of `logging`, and a fixed UTC offset instead of `zoneinfo`. Prefer\n `clay workflows code test` before wiring. Wire inputs and outputs as `data-passing.md`\n Method 3 describes: named outputs, with inputs pinned from the node that produced them, not a\n state object passed from code node to code node.\n- **Merge nodes** (`nodeType: \"merge\"`) — coalesce ordered candidate values into named outputs\n for downstream reuse. Prefer these when branch convergence or fallback selection is only\n coalescing; use a code node when computation or a non-trivial transformation is also needed.\n Read `data-passing.md` before configuring a merge node. If creating a merge node is rejected\n because the workspace does not have merge nodes enabled, do not retry; use a code node to\n produce the same named outputs.\n- **Delay nodes** (`nodeType: \"delay\"`) — pause a run before the next step without producing\n data. Set `delay` and `delayUnit` (`\"seconds\"`, `\"minutes\"`, `\"hours\"` or `\"days\"`) in the create\n or update `--input`, up to 45 days in total. To wait for a duration computed upstream instead, add\n a `delay` property to `inputSchema` with type `number`, `sourceNodeId`, and `sourcePath`; it\n overrides `delay` and is read in `delayUnit`, so set `delayUnit` too. To wait until the next\n occurrence of a recurring time, set `delayUntil` (the duration is ignored while it is set, and\n `\"delayUntil\": null` clears it). Times are 24-hour `HH:mm` in an IANA timezone:\n - `{\"kind\": \"weekday\", \"weekday\": 1, \"time\": \"09:00\", \"timezone\": \"America/New_York\"}` (0 = Sunday)\n - `{\"kind\": \"day_of_month\", \"day\": 21, ...}`: `day` is 1–31 or `\"last\"`, and late days land on a\n short month's last day\n - `{\"kind\": \"nth_weekday\", \"nth\": 2, \"weekday\": 5, ...}`: `nth` is 1–4 or `\"last\"`\n- **Map / reduce / collect** (`nodeType: \"map\"` / `\"reduce\"` / `\"collect\"`) are retired: `nodes create`\n rejects them and existing ones fail when they run. For per-record work over a list, use list mode\n (see \"List mode (Repeat)\"); to aggregate the results, add a node downstream of the Repeat node,\n which receives the whole list.\n- **Trigger nodes** (`nodeType: \"trigger\"`) — canvas entry points. Create/bind them with\n `clay workflows triggers …` (see Triggers below); do not invent a bare trigger node\n without a trigger row.\n\nDo not collapse branching or deterministic transforms into Claygent/tool-only graphs when\n`conditional` or `code` is the right fit. Node type is fixed at create — choose before\n`nodes create` because it cannot be changed later. Read an existing node of the same type\n(`nodes get`) and copy writable fields from `.node` when the shape is unfamiliar; a read leaves\nout fields the node does not set, which this skill and `data-passing.md` describe.\n\nWiring a node's inputs — `inputSchema` with `sourceNodeId`/`sourcePath`, and `inputMappingConfig`\non tool nodes — is covered in `data-passing.md`. Read it before wiring the first input, and\nbefore reading or editing a tool node whose action has dynamic fields (resolving them needs\nthe node's current static inputs).\n\n### Conditional nodes\n\nA conditional node selects exactly one outgoing edge. Set the top-level `conditionalMode` to\n`rules`, `agentic`, or `code`. Pass `rulesConditionalConfig` or `agentConditionalConfig` as a JSON\nobject, never as stringified JSON.\n\n- Use `rules` for string, number, or boolean comparisons and AND/OR groups.\n- Use `agentic` for open-ended classification that requires LLM judgment.\n- Use `code` when routing requires a deterministic Python transform.\n\n**Rules mode:** set `conditionalMode: \"rules\"`. Declare every field the rules read in top-level\n`inputSchema` with `sourceNodeId` and `sourcePath`, then reference that field name from each\n`BinOp`'s `dataPath`. Wrap conditions in a root `GroupOp`, even when there is only one condition.\nRules run top-to-bottom and the first match wins.\n\nSupported operators:\n\n- String: `Equal`, `NotEqual`, `Contain`, `NotContain`, `ContainAny`, `StartsWith`,\n `NotStartsWith`, `EndsWith`, `NotEndsWith`, `Empty`, `NotEmpty`, `GreaterThan`,\n `GreaterThanOrEqual`, `LessThan`, `LessThanOrEqual`. The four comparisons coerce the\n text with `Number()`; equality still compares the raw text.\n- Number: `Equal`, `NotEqual`, `GreaterThan`, `GreaterThanOrEqual`, `LessThan`,\n `LessThanOrEqual`, `Empty`, `NotEmpty`\n- Boolean: `True`, `False`, `Empty`, `NotEmpty`\n\n```json\n{\n \"nodeType\": \"conditional\",\n \"name\": \"Has work email?\",\n \"inputSchema\": {\n \"type\": \"object\",\n \"properties\": {\n \"email\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_source\",\n \"sourcePath\": \"$.result.email\"\n }\n }\n },\n \"conditionalMode\": \"rules\",\n \"rulesConditionalConfig\": {\n \"rules\": [\n {\n \"id\": \"has_email\",\n \"name\": \"Email found\",\n \"condition\": {\n \"type\": \"GroupOp\",\n \"combinationMode\": \"And\",\n \"items\": [\n {\n \"type\": \"BinOp\",\n \"dataPath\": [\"email\"],\n \"operator\": \"NotEmpty\"\n }\n ]\n }\n }\n ]\n }\n}\n```\n\nWire each rules destination by the matching rule id, in that destination node's `incomingEdges`.\nOn `nodes update`, `incomingEdges` **replaces** the node's full incoming-edge set — an entry you\nomit (including the trigger connection) is silently removed, so read the node first and resend\nexisting edges plus the new one. `ruleId`, `routeId`, `transitionId`, and `isDefaultRoute` are\nedge fields — they are only read inside an `incomingEdges` entry, never at the top level of the\nnode:\n\n```json\n{\n \"incomingEdges\": [\n { \"sourceNode\": \"<conditionalNodeId>\", \"ruleId\": \"has_email\" }\n ]\n}\n```\n\nAlways give a conditional a no-match fallback, in all three modes: create the destination and mark\nits edge `isDefaultRoute: true`. That edge is the only thing the run reads — setting\n`rulesConditionalConfig.defaultTargetNodeId` instead saves and validates, then fails the run with\n\"no default route connected\" the first time nothing matches. If the run should stop instead of\nrouting onward, set `rulesConditionalConfig.endRunOnNoMatch` (rules mode only) and leave the\nfallback edge off. Do not use a bare edge from a conditional unless it is truly the default route.\n\n**Agentic mode:** set `conditionalMode: \"agentic\"` and include at least one stable route in\n`agentConditionalConfig.routes`. A prompt without routes is rejected. When updating the prompt,\nre-include the complete routes array.\n\n```json\n{\n \"conditionalMode\": \"agentic\",\n \"agentConditionalConfig\": {\n \"prompt\": \"Decide whether the review is positive or negative.\",\n \"routes\": [\n {\n \"id\": \"positive\",\n \"name\": \"Positive\",\n \"description\": \"Praise or a good experience\"\n },\n {\n \"id\": \"negative\",\n \"name\": \"Negative\",\n \"description\": \"Complaints or a bad experience\"\n }\n ]\n }\n}\n```\n\nCreate the destination nodes first, then wire each one with `incomingEdges` whose `routeId`\nmatches the route's `id`:\n\n```json\n{\n \"incomingEdges\": [\n { \"sourceNode\": \"<conditionalNodeId>\", \"routeId\": \"positive\" }\n ]\n}\n```\n\n**Code mode:** set `conditionalMode: \"code\"` and put the Python handler in top-level `code`.\nCall `context.transition_to(\"Destination Node Name\", \"branch_id\")` with both arguments, then wire\nthe destination with the matching `transitionId`:\n\n```json\n{\n \"incomingEdges\": [\n { \"sourceNode\": \"<conditionalNodeId>\", \"transitionId\": \"branch_id\" }\n ]\n}\n```\n\nFor agentic and code conditionals, create placeholder destinations first, create the conditional\nwith stable route/transition ids, then attach each destination with the matching edge id.\n\n### Message nodes\n\nA Message node (`nodeType: \"tool\"` with `message`) writes a message, such as an email or LinkedIn\nnote, from fixed text, `{{input}}` variables and optional AI text written at run time. Use a Claygent\nnode instead when the whole message needs research or tools: a Message node's AI text sees only its\ninstructions, the node's inputs and the previous messages it includes.\n\nDeclare every variable in `inputSchema`, wired with `sourceNodeId`/`sourcePath`, in the same call:\n\n```json\n{\n \"name\": \"Intro email\",\n \"nodeType\": \"tool\",\n \"inputSchema\": {\n \"company\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_lead\",\n \"sourcePath\": \"$.result.company\"\n }\n },\n \"message\": {\n \"subject\": \"Quick idea for {{company}}\",\n \"body\": \"Hi team,\\n\\n{{ai:opener}}\\n\\nWorth a quick chat?\",\n \"aiText\": {\n \"opener\": {\n \"label\": \"Opener\",\n \"instructions\": \"One sentence on why **{{company}}** came to mind.\"\n }\n }\n },\n \"incomingEdges\": [{ \"sourceNode\": \"wfn_lead\" }]\n}\n```\n\n`body` is Markdown, except:\n\n- Each line is a line of the message and an empty line is a blank line, so use `\\n\\n` only for a\n gap. `<br>`, or `\\` ending a line that another line follows, breaks a line inside a paragraph.\n- Supported: `**bold**`, `_italic_`, `<u>underline</u>`, `[label](https://…)` links, and `- ` /\n `1. ` lists (indent to nest). Headings, quotes, code and other HTML are rejected.\n- `{{ai:<id>}}` is AI text; give its `label` and `instructions` (Markdown, may use variables) under\n `aiText.<id>`, and keep the ids you read back. Write a literal `{` or `}` as `\\{` or `\\}`.\n- `subject` is one line of plain text with variables and AI text, or `null` for none.\n\nAn update replaces `subject`, `body` and `aiText` together, so read the node first.\n`previousMessageNodeIds` lists upstream Message nodes the AI text may read (omit to keep them, `[]`\nto clear). Outputs: `$.result.subject`, `$.result.bodyMarkdown`, `$.result.bodyHtml` and\n`$.result.bodyPlainText`. Message nodes have no list mode. If creation is rejected because message\nnodes are not enabled, use a Claygent node.\n\n### Rep task nodes\n\nA rep task node (`nodeType: \"tool\"` with `repTask`) creates a task in Reps for a rep to work, and the\nrun continues without waiting for it. `repTask.taskType` picks the kind of task:\n\n- `\"email\"` (the default): a drafted email that the rep reviews and sends. Clay does not send it.\n- `\"call\"`: a call the rep places from their task list and logs.\n\nIts inputs have fixed names. Declare the ones you use in `inputSchema`, wired with\n`sourceNodeId`/`sourcePath`:\n\n- Email: `recipient_email` and `body_text` (plain text) are required; `subject` and `body_html` are\n optional.\n- Call: `phone_number` (a real number starting with `+` and its country code, such as `+14155550134`;\n spaces, dashes and brackets are fine, extensions are not) and `title` are required; `talk_track` is\n optional.\n- `assignee` when `repTask.assignee` is `{ \"type\": \"input\" }`: a Clay user ID or a member's email.\n- `due` when `repTask.due` reads an input: an amount for `after_input`, or an ISO date\n (`YYYY-MM-DD`) or date-time for `at_input`.\n\n```json\n{\n \"name\": \"Rep follow-up\",\n \"nodeType\": \"tool\",\n \"inputSchema\": {\n \"recipient_email\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_lead\",\n \"sourcePath\": \"$.result.email\"\n },\n \"subject\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_intro_email\",\n \"sourcePath\": \"$.result.subject\"\n },\n \"body_text\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_intro_email\",\n \"sourcePath\": \"$.result.bodyPlainText\"\n },\n \"body_html\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_intro_email\",\n \"sourcePath\": \"$.result.bodyHtml\"\n },\n \"assignee\": {\n \"type\": \"string\",\n \"sourceNodeId\": \"wfn_route\",\n \"sourcePath\": \"$.result.assigned_rep.clay_user_id\"\n }\n },\n \"repTask\": {\n \"assignee\": { \"type\": \"input\" },\n \"due\": { \"kind\": \"after\", \"amount\": 2, \"unit\": \"days\" }\n },\n \"incomingEdges\": [{ \"sourceNode\": \"wfn_route\" }]\n}\n```\n\n- `assignee` is `{ \"type\": \"user\", \"userId\": <Clay user id> }` for one member, or `{ \"type\": \"input\" }`\n to read it per run, for example from the rep routing action's `$.result.assigned_rep.clay_user_id`.\n Viewers can't be assigned tasks.\n- `due` is `null` for no due date, or one of:\n - `{ \"kind\": \"after\", \"amount\": 2, \"unit\": \"days\" }`: units are minutes, hours, days or weeks,\n as elapsed time.\n - `{ \"kind\": \"after_input\", \"unit\": \"hours\" }`: the amount comes from the `due` input.\n - `{ \"kind\": \"until\", \"until\": … }`: the next matching day and time, in the same shape as a\n delay node's `delayUntil`.\n - `{ \"kind\": \"at_input\" }`: the time comes from the `due` input, as an ISO 8601 date or date-time.\n- Each `repTask` write replaces `assignee` and `due`; omit `priority` to keep it, and omit `taskType` to\n keep the node's type. `mode` follows the type and can be left out.\n\nA run creates one task per node visit, and a rerun creates another. Rep task nodes have no list mode.\nOutputs: `$.result.taskId`, `$.result.status`, `$.result.created`, `$.result.assigneeUserId` and\n`$.result.dueAt`; a done call task also gives `$.result.notes` and `$.result.transcript` when the\ncall left them. If creation is rejected because rep task nodes are not enabled, tell the user;\nthere is no substitute.\n\n### Triggers, audiences, and draft vs live\n\n**Triggers** start a workflow (signals, audience segments, schedules, webhooks, Clay tables, CSV\nuploads, Find leads searches). Create/edit them with `clay workflows triggers …`. Clay table\ntriggers remain UI-only. A Find leads trigger (`triggerType: \"action_source_scheduled\"`) runs a\nClay search and starts one run per record it finds; if a Find leads skill (`workflows-find-leads`)\nis in your catalog, follow it for the `dataSource` shape and how to run the trigger.\n\n**Running a workflow from a CSV** (one run per row): a CSV trigger keeps the file linked to the\ntrigger node, so the same rows can be run again later. Use the `clay workflows triggers csv`\ncommands, including for a CSV the user dropped in chat (at `/mnt/session/uploads/<filename>`).\nSee `csv-triggers.md`. For an ad-hoc run over a CSV's rows, run the workflow as a routine\ninstead. A routine needs a manual trigger whose inputs match the CSV headers, so add one if the\nworkflow lacks it (see \"Running a workflow as a routine\" below).\n\n**Choose the trigger from context.** When the choice is clear and the user requested\na draft, create it without asking for confirmation again. When signals, segments, or other sources could\nmatch the request, use the user's intended event, population, cadence, and prior\ndiscussion to distinguish them. A shared keyword does not favor any trigger type.\nIf the choice remains ambiguous, ask the user to choose between the plausible sources\nand wait before binding the trigger. Do not pick a default and merely ask to proceed. Read\n`trigger-selection.md` for selection rules and examples.\n\n**Audiences inside a workflow** (`upsert-audiences-record`, `audience_segment` trigger):\n`audiences.md`.\n\n**Ad lead form triggers** (`ad_lead_form`, Meta or LinkedIn lead gen forms) are created only in\nthe workflow editor. When a workflow has one, or the user wants ad leads saved to Audiences, read\nthe complete `/workflows-ad-lead-forms` skill and follow it together with this skill.\n\n**Draft vs live** and how test runs relate to publish: see `publishing.md` and `testing.md`.\n\n**Investigating run data** (failures, traffic, durations, credits, searching run content):\n`clay workflows analysis …` — see `run-analysis.md`.\n\n## Command reference\n\nEvery command's `--input`/`--inputs` accepts inline JSON, a file path, or `-` for stdin — pick\nwhichever is easiest to construct for the payload at hand. Run `clay workflows <command> --help`\nfor the exact output shape, error codes, and examples before using a command for the first time;\ndon't guess at flags.\n\n### Creating and loading a workflow\n\n- Create: `clay workflows create --name \"My Workflow\"` — returns `{ id, name, url }`. Share `url`\n as `[Open workflow](<url>)` (see `presenting.md`).\n- Get: `clay workflows get <workflowId>` — same shape. Share `url` the first time you load an\n existing workflow.\n- Update name: `clay workflows update <workflowId> --name \"New name\"` — returns the updated\n workflow.\n- Delete: `clay workflows delete <workflowId>` — admin-only, deletes the workflow and its triggers,\n and has no undo operation. Run it only when the user explicitly asks to delete the\n workflow. A repeated delete returns `not_found` rather than succeeding again.\n\n### Reading a workflow\n\n- `clay workflows graph get <workflowId> [--mode summary|full]` — `--mode full` returns per-node\n configuration; omit for a lighter summary (nodes, edges, triggers).\n- `clay workflows nodes get <workflowId> <nodeId>` — cheaper than a full-graph read when\n you only need one node. Returns `{ \"node\": <object> }` — the writable fields live under `.node`,\n not the top-level envelope.\n\n### Editing nodes\n\n- Insert at a selected target: `clay workflows nodes insert <workflowId> --input <json|file|->` — when\n the current context names a workflow insertion target, pass its `type`, `nodeId`, and optional\n `handleId` through as `target` and describe the new node as `step`. Use this instead of manually\n creating and rewiring a node. The command reads the current graph and atomically preserves the\n downstream path. Native waterfall candidates use `step.kind: \"waterfall\"` and `waterfallConfig`;\n follow `/workflows-discover-actions` to map inputs and preserve provider order.\n Configure other node fields afterwards with `nodes update`.\n- Create: `clay workflows nodes create <workflowId> --input <json|file|->` — positioning and edge\n cleanup are handled server-side. Read an existing node first (`nodes get`) and take fields from\n `.node` for the writable shape — it's dynamic per node type. Feeding the whole get envelope into\n `--input` is a validation_error.\n- Update: `clay workflows nodes update <workflowId> <nodeId> --input <json|file|->` —\n only the fields in `--input` change. For native waterfalls, read `node.waterfallConfig`, modify\n its shared inputs or provider input overrides, and send the complete `{ \"waterfallConfig\": ... }`\n as a separate update. Preserve template selection, provider order, and provider IDs. This uses\n the waterfall save path without executing providers; do not edit the generated call bindings\n or hidden child nodes directly.\n- Delete: `clay workflows nodes delete <workflowId> <nodeId>` — edge cleanup handled\n server-side.\n- Test one node: `clay workflows nodes test <workflowId> <nodeId> --source-run <runId>`\n or `--inputs <json|file|->` — starts a partial run of only that node, then terminates.\n Nothing downstream of the node runs, so this is for debugging a node, never for recovering\n failed runs — see \"Recovering failed runs\" in `testing.md` for that. Returns immediately (does\n not poll); inspect with `clay workflows runs get`. Pass `--source-run`, `--inputs`, or both;\n with both, `--inputs` replaces the matching inputs recorded in the source run.\n- Account agents (agent nodes backed by a platform-managed Claygent with account context):\n `clay workflows nodes create-account-agent <workflowId> <nodeId> [--name <name>]` links a new\n one, from the platform template, to an agent node that has no linked agent — create the node\n bare (name and edges, no agent fields), run this with a specific `--name`, then set the prompt\n and tasks on the linked node. `clay workflows nodes upgrade-to-account-agent` converts a node\n already backed by a regular Claygent, and is one-way. Under an audiences trigger, default to an\n account agent for work that benefits from account context (research, enrichment, scoring,\n qualification, personalized outreach), and keep a regular Claygent for pure synthesis of\n upstream output. Account agents created from a workflow cannot search the web, so give\n public web research to a regular Claygent, downstream of the account agent when the step\n also needs account context (see `account-agents.md`). Include a downstream write-to-audiences node for account-agent results,\n reusing an existing node when possible; follow `audiences.md` for mappings and explicit\n destination or no-save instructions.\n- **Read `account-agents.md` in full before any account-agent work** (linking, upgrading,\n swapping, editing prompt/tasks, or changing `accountId` wiring). Preconditions, `accountId`\n wiring, the ambiguous `auth_forbidden`, timeout-retry semantics, and what an account agent\n already sees live there; `--help` carries exit codes and examples.\n\n### Canvas node groups\n\nRead `node-groups.md` before planning or building a workflow that will have more than\n8–10 nodes, and when reviewing one that already does. Follow its grouping policy: group\nnew large graphs as you build; suggest grouping existing large ungrouped graphs and wait\nfor the user before changing them.\n\n### Validating and formatting\n\n- `clay workflows graph validate <workflowId>` — structural checks only. Read-only: it neither\n repositions nodes nor persists a snapshot.\n- `clay workflows graph format <workflowId>` — validates and re-runs auto-layout. Unlike\n `graph validate` this persists a snapshot, so it needs edit access.\n\n### Running a workflow\n\n- `clay workflows runs test <workflowId> [--inputs <json|file|->] [--audience-segment <id> --limit <n>] [--trigger <triggerId>]`\n starts a run (`--limit` is required with `--audience-segment`; `--inputs`, `--audience-segment`,\n and `--trigger` are mutually exclusive). `--trigger` runs a source trigger — Find leads, CSV\n upload, or audience segment — and starts one run per record; the trigger must be live, which\n for Find leads and segment triggers means publishing the workflow first (a CSV trigger is live\n from creation). `runs get`, `runs steps`, `runs list`, `runs pause`, `runs resume` inspect and control\n runs, `runs bulk pause|resume|stop` changes many runs in one request, and\n `clay workflows publish <workflowId>` ships the draft. Flags, the `--wait` poll, and which runs\n exercise draft vs live are in `testing.md` and `publishing.md`.\n- `clay workflows runs rerun <workflowId> <runId...>` recovers failed runs: each one restarts at\n the node it failed on and continues through the rest of the workflow, on the published version.\n It spends credits, so run it with `--dry-run` first and confirm `rerunCount` and\n `estimatedCredits` with the user. The estimate is approximate; see `testing.md` for when it\n runs low. See \"Recovering failed runs\" in `testing.md`.\n\n#### Which runs a pause or resume request means\n\nOne start often creates several runs, one per record: a Run click with several input people, a\ntest-data run, a CSV upload. When the user says \"pause\", \"pause this\", or \"pause the run\" about a\nbatch they just started or are looking at, pause every run in that batch that is still going.\nDon't ask whether they mean one run or all of them when the context already names the batch.\n\n- Select only that batch: `--batch-id <id>` when its runs share a batch (`batch.id` on the runs\n that `runs test` and `runs list` return), or the run ids the start returned, each as `--run-id`.\n Don't widen the selection to other runs of the workflow, for example with `--status running`\n alone.\n- Change more than one run with `clay workflows runs bulk pause|resume`, never with `runs pause`\n or `runs resume` in a loop. \"Pausing, resuming, or stopping many runs\" in `testing.md` covers\n the flags, the dry run, and approval.\n- The exception is a batch started seconds ago: the bulk commands can't see runs that new yet.\n Pause each of its run ids with `runs pause`, which sees new runs at once and needs no approval.\n- If `clay workflows runs --help` doesn't list `bulk`, the installed CLI predates it. Pause or\n resume each run with `runs pause` or `runs resume`, or ask the user to select the runs on the\n workflow's Runs page and use Pause, Resume, or Stop there.\n- For \"resume those runs\", resume the same runs that were paused: reuse the batch id or run ids\n from the pause, not a new filter.\n- When the user asks for one run (\"pause just this one\", a single run id), pause only that run.\n Ask which runs they mean only when the context leaves the batch or the run unclear.\n- Pausing runs is not pausing their trigger. A trigger pause stops the automation from starting\n new runs and pauses the trigger's runs that are in flight, but it skips test runs, so it does\n not pause a batch the user started from the editor. Pause the trigger only when the user asks\n to stop the automation (see `publishing.md`).\n\n### Running a workflow as a routine\n\nRegister a workflow as a routine when the user wants to run it on demand over records they\nsupply (CSV rows, a pasted list) or call it from Clay MCP or the API. An unregistered workflow\nisn't visible there:\n\n```bash\nclay routines create workflow <workflowId> --name \"<what it does, e.g. Route inbound leads>\"\n```\n\n- The routine's inputs are the workflow's manual trigger `inputSchema`, so the workflow needs a\n manual trigger, and its property names are the keys each run item must use.\n- Keep those properties to `string`, `number`, `integer`, or `boolean`. A routine with any\n other input type is hidden from Clay MCP.\n- Name it for its use case. Agents choose routines by name and description.\n\nThen run it with `clay routines runs start` (see `routines`).\n\n### Testing actions and code\n\n- `clay workflows actions test <packageId> <actionKey> [--inputs <json|file|->]` — consumes\n credits. Returns `{ \"result\": <json|null>, \"metadata\"?: <object> }` — action fields live under\n `.result` (use those for `$.result.<field>` wiring); `metadata` is credit/refund accounting when\n the runner reports it. A `result` of `null` with `metadata.status` `SUCCESS_NO_DATA` means the\n action ran fine and found nothing — that is a valid result, not an error.\n Discover `packageId`/`actionKey` with `/workflows-discover-actions`, and the input shape\n with `clay workflows actions schema`.\n- `clay workflows code test --file <path|-> [--inputs <json|file|->] [--output-schema <json|file|->]`\n — runs Python in the code-node runtime to test code before adding it to a node. The file must\n define `handler(context)` returning a dict. `--inputs` backs named calls such as\n `context.get_input(\"email\")`; `get_input()` requires a key and has no zero-argument form. The\n same limits as code nodes apply: standard library only (without the modules listed under code\n nodes), no network, 1 second. **Before saving code with declared outputs, pass\n `--output-schema` with the `outputSchema` you will save on the node — flat fields or the\n JSON Schema returned by `nodes get`.**\n If output-schema validation is unavailable, test the saved node with `clay workflows nodes test` instead.\n Without output-schema validation, success only means execution succeeded. Exits 0 even on handler\n or schema failure — require `.isError != true and .outputSchemaValidated == true` for a schema\n test.\n\nBoth are single-shot previews. Use them to confirm an input/output shape, never to enrich or\nprocess a batch of records: for that use a managed routine (`clay routines runs start`), a table\ncolumn, or a workflow node. Action previews spend credits like any other action run.\n\n`rate_limited` (HTTP 429, exit 4) means temporary throttling. `details.retryAfter` gives the\ndelay in seconds.\n\n### Workflow run allowance\n\nWhen the server reports that a workflow has reached its run limit, explain the limit rather\nthan retrying or treating it as invalid workflow input. The CLI may label this response\n`validation_error`; use the message to distinguish it from other validation failures.\n\n- Where enabled, Free/Trial workspaces have a lifetime allowance of 100 runs per workflow,\n not per workspace or per month. Paid plans do not have this cap. Use the server response\n as the source of truth; do not assume the limit is enabled for every workspace.\n- Successful and unfinished runs count, including hidden or deleted runs. Failed and\n cancelled runs do not count. Do not calculate remaining allowance from the visible run\n history, which may be filtered or limited by the plan's history window.\n- Tests that create workflow runs consume this allowance too. Standalone action/code\n previews do not count toward it. Credits and API rate limits are separate from the workflow\n run allowance.\n- Explain that upgrading allows more runs; do not suggest copying a workflow or deleting\n runs to bypass the limit. After a confirmed upgrade, allow up to about a minute for the\n plan to refresh instead of asking the user to upgrade again. If it still blocks after\n that, report the issue rather than repeatedly retrying.\n\n### Triggers\n\n- Get: `clay workflows triggers get <triggerId>` — trigger ids come from `clay workflows graph get`,\n which returns them alongside the nodes; a workflow id is not accepted here.\n- Create: `clay workflows triggers create <workflowId> --input <json|file|->` — the one trigger\n command that takes a workflow id. Its output is a confirmation, not the trigger — see the wiring\n note below. The writable shape is dynamic per trigger type — read an existing trigger first, same\n pattern as nodes. When no trigger of the target type exists to read, the core writable fields per\n type: `audience_segment` = `triggerType, segmentId, entityType (\"CONTACT\"|\"ACCOUNT\")`;\n `audience_signal` = `triggerType, signalId (signal.id, \"sig_…\"), entityType`;\n `audience_scheduled` = those plus `scheduleConfig` (optional `batchProcessing`, below);\n `scheduled` = `triggerType, scheduleConfig`;\n `webhook` / `manual` = `triggerType, inputSchema`. `clay_table` triggers cannot be created\n from here — they are wired from inside a Clay table (\"Invoke Workflow\" action).\n For the `scheduleConfig` shape and the remaining trigger types, read\n `clay workflows triggers create --help` — it is the single source for the field detail.\n- Update: `clay workflows triggers update <triggerId> --input <json|file|->` — takes the trigger\n id, not the workflow's. Only the fields in `--input` change.\n- Delete: `clay workflows triggers delete <triggerId>` — trigger id again.\n- CSV: `clay workflows triggers csv upload|run|remove <triggerId>` — link a CSV file to a\n `csv_upload` trigger and run one workflow run per row. See `csv-triggers.md`.\n\n**Webhook payloads:** the trigger URL (`webhookUrl` on `triggers get`) accepts a JSON body. An\nobject body becomes the run inputs as-is, so `inputSchema` describes the object and downstream\nnodes read `$.email`. A top-level array body is stored whole under `payload`, never spread into\nindex keys, so declare `\"payload\": { \"type\": \"array\", \"items\": { … } }` in `inputSchema` and read\n`$.payload[0].email` (or iterate it in a list-mode node with `listEntriesRef` `$.payload`). Senders\nlike HubSpot post arrays of events, so check the sender's shape before writing the schema. When you\ncontrol the sender, post an object.\n\n**Webhook retries:** when the sender can set headers, have it send `X-Clay-Webhook-Idempotency-Key`\n(1 to 250 visible ASCII characters) with a stable id for the event, such as an order or event id, and\nreuse it on every retry. A POST whose key the trigger has already received starts no new run, even\nif its body differs. Only `webhookUrl`s under `/sources/webhook/` honor the header; older\n`/tc-workflows/streams/…/webhook` URLs ignore it. An invalid key returns 400.\n\n**Audience multi-segment sharing:** multiple `audience_segment` triggers (different `segmentId`s)\nmay share one trigger node when they have the **same trigger type** and the **same outgoing edge**.\nMultiple `audience_scheduled` triggers may share a node when they also have the **same schedule**.\nPass an explicit `workflowNodeId` to bind/share; omit it (or pass `createTriggerNode: true`) to get\na new node. Do not mix `audience_segment` with `audience_scheduled` on one node. `audience_manual`\nis a run companion created by the UI/run path — do not create it via the CLI.\n\n**Audience batch processing:** on `audience_scheduled`, `batchProcessing: {\"batchSize\": <n>}`\nmakes each scheduled run process the next `<n>` segment members not yet processed in this pass,\ninstead of rerunning every member, and the trigger pauses itself once all members are done.\n`null` (or no field) means every run reruns the whole segment. `triggers get` returns\n`{batchSize, startedAt?, completedAt?}` or `null`. Before publishing a scheduled audience trigger\nthe user wants to process incrementally, set `batchProcessing`, then `triggers get` and quote the\nreturned `batchSize` back. Never describe per-run behavior the read-back does not show: with\n`batchProcessing: null`, publishing reruns the whole segment on every run.\nTurning batching back on after `null` reruns every member unless the pass was unfinished and the\ndisable unpublished; treat it as a full-segment run for the cost policy and confirm with the user first.\n\n### Enabling batching on a tool node\n\nSome actions support batching multiple workflow runs into a single provider call, cutting\ncost/rate-limit pressure for high-volume workflows. Not every action supports it, and the\ncatalog doesn't flag which ones do — enable batching only when the user explicitly raises\nbatching, rate limits, or large volumes of rows/runs; never proactively.\n\nSet it via `nodes update` after the tool node already has `tools` configured:\n\n```bash\nclay workflows nodes update <workflowId> <nodeId> --input '{\n \"batchRunSettings\": { \"enabled\": true, \"maxBatchSize\": 50 }\n}'\n```\n\n`maxBatchSize` is optional and gets clamped to the action's real maximum. If the action\ndoesn't support batching, the update rejects — relay that to the user rather than retrying.\nCreate with `tools` first, then enable batching in a follow-up update (same conversation turn\nis fine as two calls).\n\n### List mode (Repeat)\n\nAn agent, code, tool, or workflow function call node (including a native waterfall) can run\nonce per item in an upstream list instead of once over the whole list. In the editor this is\nthe **Repeat** toggle. Describe it to the user as **\"runs once per item\"** — never say it \"maps\nover the list\" or call it a map/loop.\n\nEnable it on the node with `listMode: true` **and** `listEntriesRef` in the same\n`clay workflows nodes update` (or on create), pointing at the upstream array:\n\n```bash\nclay workflows nodes update <workflowId> <nodeId> --input '{\n \"listMode\": true,\n \"listEntriesRef\": { \"sourceNodeId\": \"wfn_upstream\", \"path\": \"$.results\" }\n}'\n```\n\nAlways set `listEntriesRef` when turning list mode on: without it the node looks for an\n`entries` input instead of the upstream array, and repeats over the wrong thing. Pass\n`null` to clear it.\n\nFor tool nodes, map each input that should vary per item with `inputMappingConfig` `item`\n(`{ \"type\": \"item\", \"path\": \"$.field\" }`) — not a `reference` to the upstream list (that\nresolves to the whole aggregate). See `data-passing.md`. For list-mode code nodes, the item\nis under `input`: use `context.item`, `context[\"input\"]`, or `context.input`. Code tests pass\n`--inputs` verbatim, so wrap each test item as `{\"input\":{\"name\":\"Ada\"}}`.\n`listFailureMode` is `\"fail_at_end\"` (default: fails the node if any item fails) or\n`\"ignore_errors\"` (keeps the successful items and continues). `ignore_errors` only covers a\npartially failed list: if every item fails — including a one-item list whose item fails — the\nnode still fails and the run stops.\n\n**When to use it:** the same per-record work over a list (enrich each row, draft one message\nper lead), inline on one node. **When not:** a single aggregate step over the whole list — leave\n`listMode` off.\n\n**Function calls and waterfalls:** read `repeat-function-calls.md` before turning on Repeat for a\nworkflow function call or native waterfall, and whenever each item needs more than one step — it\ncovers mapping the current item, the two-update waterfall flow, and when to use one repeating\nfunction call instead of a chain of Repeat nodes.\n\nUnsupported on account agent nodes and on trigger/conditional/map/reduce/collect/fork/join.\nIf the workspace has list mode disabled, the update rejects — relay that rather than retrying.\n\n### Wiring the first node after creating a trigger\n\n`triggers create` returns `resourceId` (the **trigger** id) and `workflowNodeId`\n(the canvas node id). Use `workflowNodeId` to wire the first node; do not use\n`resourceId` as an `incomingEdges` source. If the response omits `workflowNodeId`,\nread `clay workflows triggers get <resourceId>` to obtain it.\n\n```bash\nclay workflows triggers create wf_1 --input trigger.json > created-trigger.json\njq '{resourceId, workflowNodeId}' created-trigger.json\n```\n\nWire it with `incomingEdges: [{ \"sourceNode\": \"<workflowNodeId>\" }]`. Do not send\n`sourceNodeId` inside `incomingEdges` — graph summary uses that key on the read side.\n\n**Trigger edge constraint:** a trigger may have zero or one direct outgoing edge, never more.\nBefore wiring a node with `incomingEdges` from a trigger's `workflowNodeId`, check the default\nsummary for an edge already leaving that trigger node —\n`clay workflows graph get <workflowId> | jq '.summary.edges[] | select(.sourceNodeId == \"<triggerNodeId>\")'`.\nIf it already has a target, do not add another direct edge; add work downstream instead, or ask\nthe user whether to rewire. Before validating or running, each trigger must be connected to one\nfirst executable node.\n\n## Capabilities without a `clay workflows` command\n\nBuilding a workflow often touches adjacent surfaces — tables, functions, audiences — that have\ntheir own CLI commands elsewhere rather than a `clay workflows` subcommand:\n\n- **Table schema** — `clay tables columns list <tableId>` (abbreviated) or\n `clay tables columns get <tableId>` (full settings).\n- **Table rows** — `clay tables query-live <tableId> --query <clayql>`, against live Postgres.\n It takes ClayQL directly — write the query yourself off the columns above; there is no\n natural-language query option. Don't reach for `clay tables query` instead — that reads a\n different store (synced ClickHouse, gated on Enterprise sync).\n- **Table configuration migration** — read `table-primitives.md` when rebuilding\n table columns as workflow nodes. Preserve referenced resources and verify the\n persisted bindings before testing.\n- **Provider waterfall migration** — read `table-waterfalls.md` before translating\n a table's provider waterfall. Check for an equivalent installed function and preserve\n the input bindings, run condition, provider order, validation, and output selection.\n- **Bulk enrichment migration** — for a known bulk enrichment table, use\n `clay tables columns get <tableId>` to determine the\n source origin and column DAG. When a source includes `audience`, use its configuration\n (global when `isGlobal`; otherwise `segments[].id`). Do not assume every bulk enrichment is\n Audiences-backed.\n For an Audience-backed bulk enrichment migration, or when the destination already\n has type `audience_enrichment`, read\n `/workflows-audience-enrichment` before translating the columns. Its migration guidance\n explains how to replace the table's source-record lookups and extracted columns with\n bindings to the existing audience trigger.\n `clay tables rows list <tableId> --limit <n>` is only an optional,\n unfiltered, unordered sample (`n` at most 20): it returns `truncated: true`, accepts and returns no\n cursor, and completed passthrough rows may already be deleted. Never treat it as a complete source dataset.\n- **One audience segment** — `clay audiences get <audienceId>`, which includes the filter AST that\n selects its records.\n- **Audience record fields** — `clay audiences fields list` with `--entity-type people` or\n `--entity-type companies` returns the ids for `upsert-audiences-record` (see `audiences.md`).\n Mind the two vocabularies: the flag is `people`/`companies`, while `upsert-audiences-record`\n still wants `entityType: \"CONTACT\"`/`\"ACCOUNT\"`.\n- **Function schemas** — `clay functions get <functionId>` returns `inputSchema` and `outputSchema`.\n `clay workflows actions list` carries the `functionId` to pass it.\n- **Listing what exists** — one command per resource kind: `clay workflows list`,\n `clay tables list`, `clay functions list`, and\n `clay audiences list --entity-type people|companies` for audience segments.\n\nThe raw CPJ Search DSL has no CLI command; `clay searches query-mode`\ncovers Clay search itself (see the `searches` skill). When a capability isn't listed here and has no\n`clay` command, it's genuinely unavailable rather than hidden — tell the user that instead of\nhunting for a command that will not appear.\n"
}SHA-256 of public snapshot: 20f424eae745d7e9cbd921e37681e09f641e5ae31448b65dee6eb23ab637166c