A CampaignA developer's data-production project and its settings can collect original data from people or combine human work with automated checks. Choose the workflow around the result your application needs.
These examples are recipes to configure, not shipped presets. Use the available taskA specific item a contributor claims or enters types, input routes, and Agents in your organizationThe Developer-side unit that holds Campaigns, data access, and funds.
Collect original human-sourced data
Need: examples that people create or observe, such as original sentences, voice samples, local observations, or research evidence.
Example: build a set of original questions and statements for a language application.
Path: Entry GateThe workflow step that holds a Campaign's joining requirements → Supply (source) → Accept → Sink.
Ask for a sentence and category. Define originality and acceptance criteria in the brief, bind the form to Supply, and review the submissionsThe record created when work is submitted. Start with a few items so you can check the instructions and delivered fields before expanding the Campaign.
Result: accepted records that your authorized application can retrieve once delivery conditions are satisfied.
Source research with an agent and human verification
Need: source-backed facts that require both automated checking and human judgment.
Example: assemble a research dataset of public project announcements. Contributors submit a source URL, publication date, claim, and supporting excerpt. A configured Agent checks relevance to the research brief. A human verification step checks the source and corrects or rejects unsupported claims.
Path: Entry Gate → Supply (source) → Agent Task → Supply (validate) → Accept → ETL → Sink.
Configure the Agent and its routing explicitly. Bind validation to the earlier source work, specify reviewer capacity and the decision rule, and preserve the evidence needed for acceptance. Use ETL only where field mapping is needed, for example to deliver source_url, published_at, claim, and evidence consistently.
Human contribution: finding material, interpreting context, and resolving ambiguous claims.
Automated contribution: checking candidate material against explicit instructions. An Agent Task requires a wired Agent; adding the node alone does not configure a crawler or research service.
Result: reviewed, structured records with evidence your application can inspect.
An external agent may also discover candidates, but ingestion must use a supported route. The documented dataset import is limited to eligible same-organization datasets and a bounded batch; do not describe it as arbitrary external-record ingestion. See CLI dataset imports before planning that integration.
Label and assess existing records
Need: consistent categories, human preferences, or explanations of model behavior.
Example: ask reviewers to categorize a support response and explain whether it answers the customer's question.
Path: supported record input → Supply (label) → Accept → Sink, with participation requirements configured in the Entry Gate.
Define a rubric and form fields for the label and explanation. Supply is the human-work node; the standalone Evaluate node is currently locked. Add further validation only where the workflow needs it, and define how disagreements are resolved.
Result: accepted labels and reasoning that can support evaluation or dataset improvement. Whether you can access and reuse the records still depends on delivery and usage terms.
Keep authoring and execution separate
Builder, the CLI, and the Campaign API are ways to author a Campaign. In Builder, Tiny-h / DSA can help analyse data and propose design changes for you to review and apply. A coding agent can also assist through supported CLI/MCP operations. An Agent Task is a configured step inside the Campaign's running workflow. Using Tiny-h while authoring does not automatically add an Agent Task to that workflow.
Whichever route you choose, define the required output, participation requirements, acceptance process, reward, budget, and delivery path. Validate the draft, test a small batch, and trace one real submission through to its retrieved record.
See Configure a workflow, existing operational use cases, and Start with a data need.