Skip to content

Agent Skills

Agent skills are portable Markdown guides for any AI coding assistant. Published repository:

https://github.com/dataflow-operator/skills

Unlike MCP, skills do not call tools — they provide operational knowledge from documentation and config/samples.

Skills vs MCP

Agent skills MCP
Purpose Deploy, config patterns, resilience checklists Generate and validate YAML via tools
Invocation IDE-specific or read AGENTS.md in skills repo MCP tools in chat
Best for Helm setup, prod manifests, idempotency review Draft YAML, Kafka Connect migration

Installation

Full instructions: skills/INSTALL.md

git clone https://github.com/dataflow-operator/skills.git /tmp/dataflow-skills
mkdir -p .cursor/skills          # or .claude/skills for Claude Code
cp -r /tmp/dataflow-skills/* .cursor/skills/

Cursor: @dataflow, @dataflow-config, …

Claude Code: copy to .claude/skills/ or link AGENTS.md in CLAUDE.md.

Copilot: reference https://github.com/dataflow-operator/skills in .github/copilot-instructions.md.

Any agent: open AGENTS.md in the skills repo.

Available skills

Skill Purpose
dataflow Entry point: DataFlow vs DataFlowCron
dataflow-deploy Helm install, CRD, verification
dataflow-config Write or review YAML
dataflow-fault-tolerance Idempotency, checkpoint, ack

Companion files in the skills repo: anti-patterns.md, samples-index.md.

Typical workflow (skills → validate → apply)

  1. dataflow-deploy — operator installed (Helm/CRD, webhook enabled in prod).
  2. dataflow / dataflow-config — choose kind, draft manifest from skills + samples; check anti-patterns.
  3. dataflow-fault-tolerance — pre-apply checklist (upsert, checkpoint, replicas, ack).

  4. Optional MCP — generate_dataflow_manifest / validate_dataflow_manifest for a quick draft.

    MCP validate is shallow

    MCP checks apiVersion/kind and that source/sink have type + config. It does not replace the operator validating webhook. See MCP.

  5. Server dry-run (admission) — before creating the resource:

    kubectl apply --dry-run=server -f manifest.yaml

  6. Apply and verify

    kubectl apply -f manifest.yaml
kubectl get dataflow,dataflowcron -n <namespace>
kubectl describe dataflow <name>

LLM crawlers

Machine-readable map of docs and agent entry points: llms.txt (served at the site root /llms.txt).

JSON Schemas (machine-readable contract)

Stable URLs under /schemas/:

Artifact Use
catalog.json Allow-list of source/sink/transform types; operatorVersion pin
dataflow.json / dataflowcron.json CRD envelope (top-level spec)
connectors/sources|sinks/<type>.json Typed connector config (CRD stores it as opaque RawExtension)
transforms/<type>.json Typed transformation configs

Workflow: draft from skills → optional MCP (shallow) → validate against schemas / kubectl apply --dry-run=server.

Maintainer

Skills are developed in the dataflow monorepo under skills/ and published to dataflow-operator/skills. Extensions for operator development: .cursor/skills/dataflow-* in the monorepo (stubs + maintainer-only files — see SYNC.md).

See also