where do you need better context?
how this works
Use recent agent work as evidence. Feelings such as frustration, distrust, confusion, helplessness, or delight are signals about the quality of the context your agent received.
Write concrete answers to the four questions below. Do not try to solve every problem while you are answering. First identify the moments and patterns; then choose one context improvement to make permanent.
reflect (15 minutes)
- When do you feel bad while using your agent? Look for anger, frustration, distrust, overwhelm, confusion, boredom, or helplessness.
- When do you feel good? Look for surprise, delight, relaxation, safety, curiosity, engagement, pride, or focus.
- What common mistakes does your agent keep making? Use repeated behavior, not a general judgment that it is “bad.”
- What do you wish your agent always knew? Include your work, goals, preferences, constraints, files, or the context of what you are doing.
turn one finding into a fix
Choose one answer and classify the underlying problem:
- Missing context: the agent needs a fact, example, goal, preference, or source.
- Noisy context: the agent sees stale, irrelevant, or contradictory material.
- Unclear goal: the desired outcome or success condition is underspecified.
- Harness problem: the agent lacks a tool, permission, connection, or usable environment.
Make the smallest permanent improvement that addresses the cause. Add a durable fact to an appropriate AGENTS.md, create or improve a skill for a repeated procedure, start a fresh conversation, clarify the goal, or fix the relevant permission or connection. Then try the task again and check whether the behavior changed.
share the useful part
If you are doing this with a group, share one pattern and one change—not private details from your reflections. A useful share-back sounds like: “My agent kept doing this because it was missing that; I changed this file or instruction, and the next run did this.”