LLM-Powered Operations

Loom can use LLM inference to generate or modify files as part of a module's operation sequence. This lets you produce content that is too dynamic or context-dependent for static templates — config summaries, documentation, policy files, or anything that benefits from natural language reasoning.

When to Use It

Use llm operations when:

• The output structure is too variable for a fixed template
• You want to generate prose (README sections, runbook entries, PR descriptions)
• You need to modify an existing file in a way that requires understanding its content

Use newFiles + templates instead when the output is predictable and structured — templates are faster, deterministic, and cheaper.

Basic Example

operations:
  - name: generate-deployment
    llm:
      provider: anthropic
      model: claude-sonnet-4-20250514
      systemPrompt: "Output only valid YAML. No explanation."
      prompt: |
        Generate a Kubernetes Deployment for {{ .serviceName }}
        in namespace {{ .namespace }} using image {{ .image }}.
      target: "deploy/{{ .serviceName }}.yaml"
      providerConfig:
        tokenEnv: ANTHROPIC_API_KEY

When loom run reaches this operation, it renders the prompt with the module's params, calls the model, and writes the response to deploy/payments.yaml in the target repository.

Two Modes

generate

Creates a new file. Fails if the target already exists — this is intentional to prevent silent overwrites.

llm:
  mode: generate   # default — can be omitted
  target: "docs/{{ .serviceName }}-runbook.md"
  prompt: "Write a runbook for {{ .serviceName }}."

modify

Reads the existing file at target, sends its content to the model alongside your prompt, and overwrites the file with the response. Use this to append, reformat, or update existing content.

llm:
  mode: modify
  target: README.md
  prompt: "Add a ## {{ .serviceName }} section at the end."

Providers

All major LLM providers are supported:

Provider	provider value	Auth
Anthropic	anthropic	ANTHROPIC_API_KEY
OpenAI	openai	OPENAI_API_KEY
Google Gemini	gemini	GEMINI_API_KEY
Google Vertex AI	vertex	Application Default Credentials
OpenRouter	openrouter	OPENROUTER_API_KEY
AWS Bedrock	bedrock	AWS credential chain

API keys are always read from environment variables — never put secrets in loom.yaml.

providerConfig:
  tokenEnv: MY_ANTHROPIC_KEY   # reads $MY_ANTHROPIC_KEY instead of $ANTHROPIC_API_KEY

Combining with Other Operations

llm is just another operation in the sequence. You can mix it freely with newFiles, patch, shell, and git operations:

operations:
  - name: scaffold-files
    newFiles:
      source: "."
      dest: ""

  - name: generate-readme
    llm:
      provider: anthropic
      model: claude-sonnet-4-20250514
      prompt: "Write a README for {{ .serviceName }} based on the Kubernetes manifests in this repo."
      target: README.md

  - name: commit
    commitPush:
      message: "feat: onboard {{ .serviceName }}"

Retry on Failure

LLM APIs can be flaky. Use retries and retryDelay to automatically retry on failure with exponential backoff:

llm:
  provider: openai
  model: gpt-4o
  prompt: "..."
  target: "out.yaml"
  retries: 3
  retryDelay: "2s"   # delays: 2s → 4s → 8s

Full Reference

See llm operation reference for all fields and provider-specific configuration.