kestra-flow（凯斯特拉流）

Kestra Flow Skill

Use this skill to generate production-ready Kestra Flow YAML grounded in the live schema.

When to use

Use this skill when the request includes:

• Generating a new Kestra flow from scratch
• Modifying, extending, or debugging an existing flow
• Translating a workflow description into valid Kestra YAML

Required inputs

• A description of the desired flow behavior
• Namespace (and tenant ID if applicable)
• Existing flow YAML if the request is a modification

Workflow

Step 1 — Fetch the flow schema

Fetch the full schema with curl and read it directly — do not pipe it through any interpreter:

curl -s https://api.kestra.io/v1/plugins/schemas/flow

Read the raw JSON output to validate every type, property name, and structure used in the output. Do not generate anything before the schema is available.

Step 2 — Collect context

Identify from the user message or conversation:

• id — flow identifier (preserve if provided)
• namespace — target namespace (preserve if provided)
• Existing flow YAML (for modification requests)
• Whether this is an addition / deletion / modification or a full rewrite

Step 3 — Generate the YAML

Apply all generation rules below, then output raw YAML only.

Generation rules

Schema compliance

• Use only task types and properties explicitly defined in the fetched schema. Never invent or guess types or property names.
• Property keys must be unique within each task or block.

Structural preservation

• Always preserve root-level id and namespace if provided.
• For modification requests, touch only the relevant part. Do not restructure or rewrite unrelated sections.
• Avoid duplicating existing intent (e.g., replace a log message rather than adding a second one).

Triggers

• Include at least one trigger if execution should start based on an event or schedule.
• Do NOT add a Schedule trigger unless a regular occurrence is explicitly requested.
• Trigger outputs are accessed via {{trigger.outputName}}; only use variables defined in the trigger's declared outputs.

Looping

• Use ForEach for repeated actions over a collection.
• Use LoopUntil for condition-based looping.

Flow outputs

• Only include flow-level outputs if the user explicitly requests returning a value from the execution.

State tracking between executions

• For state-change detection, use KV tasks (io.kestra.plugin.core.kv.Set / io.kestra.plugin.core.kv.Get) to store and compare state across executions.

JDBC plugin

• Always set fetchType: STORE when using JDBC tasks.

Date manipulation in Pebble

• Use dateAdd and date filters for date arithmetic.
• Apply | number before numeric comparisons.

Credentials and secrets

• Never embed secrets or hardcoded credentials.
• Use flow inputs of type SECRET or Pebble expressions (e.g., {{secret('MY_SECRET')}}).

APIs and connectivity

• Prefer public/unauthenticated APIs unless the user specifies otherwise.
• Never assume a local port; use remote URLs.

Quoting

• Prefer double quotes; use single quotes inside double-quoted strings when needed.

Error handling

• If the request cannot be fulfilled using only schema-defined types and properties, output exactly: I cannot generate a valid Kestra Flow YAML for this request based on the available schema.

Output format

• Raw YAML only — no prose, no markdown fences, no explanations outside the YAML.
• Use # comments at the top of the output for any caveats, assumptions, or warnings.
• The output must be ready to paste directly into the Kestra UI or deploy via kestractl.

Example prompts

• "Write a Kestra flow that fetches a public API every hour and stores the result in KV store."
• "Add a Slack notification task to this existing flow when any task fails."
• "Generate a flow in namespace prod.data that reads from a Postgres table and writes the result to S3."
• "Debug this flow YAML — it has a trigger variable reference that doesn't exist."