Make actions accomplishable by less capable agents over time.
What this means
Every time a capable agent (human or large model) performs a task, part of the work should be encoding that task so that a less capable agent could perform it next time. The trajectory is:
learn → document → curricula → skills → scripts → formalization
Each step reduces the inference required. A document records what was learned. A curriculum organizes it for teaching. A skill encodes it as repeatable instructions. A script automates the mechanical parts. A formal proof eliminates ambiguity entirely.
The goal is not to remove the need for capable agents — it is to move their effort from routine execution to genuine reasoning, while routine execution flows downward to cheaper, faster, more deterministic tools.
Operational implications
- When you finish a task, ask: what part of this could a dumber agent do next time?
- Encode that part as a skill, script, or frontmatter convention.
- Prefer deterministic scripts over inference-based skills where the task is well-understood.
- Design skills with enough detail that a less capable model could follow them.
- When frontmatter can encode what prose currently describes, migrate the information to frontmatter.