Scout Pattern
Before committing to a complex task, send the AI on a throwaway 'scouting' attempt to discover where complexity lies, which files are involved, and what questions arise.
The Scout Pattern is a reconnaissance technique for complex AI-assisted tasks. Before committing to a real implementation, you send the AI on a throwaway attempt—not to land the code, but to learn where the complexity lies.
The insight comes from Josh Bleecher Snyder: "Hand the AI agent a task just to find out where the sticky bits are, so you don't have to make those mistakes." The scout attempt is exploratory; you're gathering intelligence, not shipping code.
What scouts reveal: Which files get modified (shows AI's understanding of system architecture), where it gets stuck (reveals genuine complexity vs. routine work), what approach it takes (may inspire your own strategy), and what questions it asks (surfaces ambiguities in requirements).
How to run a scout:
- Create a throwaway branch:
git checkout -b scout/feature-exploration - Give the AI the task without heavy constraints
- Let it attempt the full implementation
- Observe what happens—files touched, patterns used, errors encountered
- Discard the branch and start the 'real' attempt with insights gained
When to use scouts: Complex refactoring where scope is unclear, new features in unfamiliar parts of the codebase, tasks where you suspect hidden dependencies, or migrations with unknown blast radius.
The key is that scouts are disposable. Run them in temp checkouts or throwaway branches. Let them fail safely. A failed scout that reveals hidden complexity is more valuable than a 'successful' attempt that misses critical edge cases. The Scout Pattern complements the Document & Clear pattern: after a scout reveals the landscape, document your findings, clear context, and start fresh.
Key Takeaways
- Use throwaway attempts to learn where the 'sticky bits' are before the real implementation
- Scout attempts reveal which files the AI modifies, indicating system understanding
- Failed attempts provide valuable context for the real attempt
- Low-stakes exploration surfaces ambiguities in requirements early
- Run scouts in temp branches or throwaway checkouts - you're not planning to land the code
- Pattern is tool-agnostic: works with any AI coding assistant
Visual Overview
In This Platform
When adding new content types to this platform, we often use a scout approach: first attempt to add a minimal example of the new content type without worrying about getting it right. This reveals which schemas need updating, which build script sections handle the content, and what validation errors arise. The insights from the scout inform the real implementation.
- build.js
Directoryfrontend/src/schemas/
- …