AI & craft
Directing Agents
Six sessions on steering an agent's attention: talk in small bites, check blast radius, summarize an owned plan, then act. Not more context. Better focus.
Built for: Developers who already use AI coding tools and want fewer surprises from wrong assumptions, dense plan dumps, and jump-straight-to-code habits
Welcome
You ask an AI coding tool for help. Sometimes with one line. Sometimes with a careful wall of context. Either way, the agent still surprises you: it starts coding too early, misses blast radius, or invents a path you never wanted.
The usual diagnosis is “I needed a better prompt” or “I needed more context.” This guide argues for a different miss: steering attention.
LLMs work by focusing on something. Hyper-focus on “implement this” makes confident mistakes. Diffuse focus makes the agent invent a destination. The skill is not writing longer English. It is steering what the agent (and you) pay attention to, then locking a short owned summary before implementation. The same directing pattern works with people; agents make the mechanism more visible because the emotional story falls away.
This guide is six sessions in plain language. No single product is required. The path you should leave with is simple:
Talk → summarize → act
You will not memorize prompt templates and move on. Each session starts with a question for you, walks through one ordinary task, and ends by asking you to compare what you believed with what the session showed. The thread through all six is the same job: add soft delete for user accounts, so abstract habits land on forks and blast radius you can picture.
By the end, you should notice new things in your own week: when a prompt made the agent assume “start coding”; when a formal plan dump made you lose the decision; when a short conversation would have been enough; when you were ready to summarize and move.
What this guide is (and is not)
This is not a list of clever prompts. It is not a claim that conversation replaces every plan artifact.
It is six sessions built around one ordinary task, add soft delete for user accounts, so each idea stays tied to work you already understand: assumptions, focus, blast radius, and ownership.
If you want the wider AI-native workflow (harness, repo rules, verification ladders, team norms), start with AI-Native Engineering. This guide sits beside it. That one is systems and habit infrastructure. This one is the attention habit before the loop runs.
Who might find this useful
| If you… | A gentle place to begin |
|---|---|
| Type “implement X” and regret the diff | Session 1, then Session 2 |
| Paste huge context and still get wrong focus | Session 1 and Session 2 |
| Use plan mode, but plans feel hard to own | Session 3 and Session 4 |
| Talk forever without a finish line | Session 5 |
What helps beforehand: basic comfort with a repository and with some AI coding tool (chat, agent, or plan mode). You do not need a paid product or a specific editor.
The six sessions
| Session | What you will explore |
|---|---|
| 1. The wrong assumption | Why the agent starts coding, and why more context is not the fix |
| 2. Steer the attention | Let the agent propose the path; why that brief works |
| 3. Plan mode’s gift and trap | Force planning without drowning in a formal dump |
| 4. Talk in small bites | Offload cognition into conversation-sized pieces |
| 5. Summarize, then act | Lock a short owned plan, then implement |
| 6. Check + carry forward | Quiz, full story, quick reference, and what to try next |
When you finish Session 6, a quick reference section there collects the prompt shape, talk loop, and summary template in one place for bookmarking.
Each session opens with a short question to answer in your own words, and closes with a reflection on that same question after you have read the session.
Open Session 1 from the sidebar when you are ready. Read at your own pace.