Program is your prompt, written in natural language. It specifies initial inputs, "imports" external functions via tool descriptions, and implements business logic through control flow: sequential steps, loops, conditionals, and goto. Tool calls and user input are I/O.

1. Prepared information - Codebase docs, style guides, architecture overviews (baked into prompt or loaded from disk)
2. User input - Clarifications, corrections, new requirements during execution
3. Tool outputs - File contents, command results, API responses

Context is ephemeral: compaction will eventually wipe it. Context limits hit quickly with substantial state.

Solution: Serialize to disk using LLM-friendly formats:

• JSON - Structured data; LLM can surgically read/update specific fields via jq
• Markdown - Smaller unstructured data; load fully into context when needed

Payoff: Resume from any point with fresh context, sidestepping compaction entirely.

Outputs aren't limited to generated code. Like traditional programs producing console output, writing files, or displaying GUIs, LLM programs use tool calls to create:

• Generated code and diffs
• Editor actions (open files)
• Codebase statistics
• Change summaries
• Progress artifacts

These outputs serve multiple purposes:

• Help you review the work
• Provide input for next workflow steps
• Show program progress

Prompt (program)
  ├─ Inputs: docs + user + tools
  ├─ State: JSON (structured) / MD (unstructured) → disk
  └─ Outputs: code, diffs, summaries via tools

• Resumable - Fresh context can pick up from serialized state
• Compaction-proof - State survives context resets
• Explicit I/O - Clear boundaries between program and environment
• Composable - Workflows chain through shared state files