Session 1
The Wrong Assumption
Why “Implement soft delete” fails: not missing context, but the prompt steered attention at the wrong job.
- See the wrong-job assumption behind jump-to-code prompts
- Separate missing context from mis-steered attention
- Recognize hyper-focus and under-focus as twin failure modes
- See why the same directing skill is older than agents
A situation you may recognize
A ticket lands: add soft delete for user accounts.
You type:
Implement soft delete for user accountsThe agent starts coding. Fast. Confident. Then the surprises arrive: an archive table you never asked for, cascades into billing, auth changes as a side quest, or a huge diff that is hard to review. You clarify. It pivots. You paste more context. It still feels like you are cleaning up a road you never chose.
The common diagnosis is: I should have given more context.
Sometimes that is true. Often it is not the whole story.
The prompt already steered attention
That one-liner did more than under-specify soft delete. It steered the agent’s attention at a job: implement.
LLMs work by focusing on something. When the strongest signal is “implement,” the model hyper-focuses on producing code. Blast radius, alternatives, and “should we do this another way?” fall out of focus. Adding more files or more paragraphs of background does not automatically change the job. It can even feed the same assumption with richer fuel.
| Prompt signal | Where attention goes |
|---|---|
| ”Implement X” | Start coding X |
| Huge context + “implement X” | Start coding X, with more material to use |
| ”Make it clean” | Invent a destination that looks clean |
| ”Propose a path, check blast radius, suggest next step” | Plan and advise before driving |
The third row is under-focus: no clear job, so the model invents one. The first two are hyper-focus on implementation. Both miss for different reasons. The shared miss is steering, not word count.
This guide’s name for the skill is simple: steer the agent’s attention. Session 2 shows the concrete mechanisms. Session 1 only needs the diagnosis.
That move adds context while keeping the same steering: start coding.
If the miss was attention, you need different focus targets first: what should we do, what is the blast radius, what is the next step. Session 2 turns that into a practical brief and explains why those clauses change the run.
This is familiar if you have worked with people
The pattern is not new.
Hand a teammate “implement soft delete” with no path, no blast radius, and no next step, and you often get the same shape of surprise: they optimize for shipping a delete-looking change, miss sideways damage, and you feel the cleanup after. Design review and good pairing already steer attention: clear outcome, visible plan, hard questions, then execution.
Agents make the same gap louder and cleaner. There is less emotional fog. You are less likely to narrate the failure as personality (they ignored me, I already told them) and more likely to see the mechanism: the brief steered focus at the wrong job. Detaching the emotional story is useful. It does not invent a brand-new skill. It reveals an old directing skill that was always there.
Don’t Prescribe the Road approaches the same family from the other side: a turn-by-turn recipe steers attention at your pre-written forks. The one-liner steers attention too narrowly at coding. Different shapes, same miss: the brief chose the wrong thing to focus on.
What Session 1 asks you to notice
Before the next sessions give you a replacement habit, practice one notice this week:
When an agent surprises you, ask: where did my prompt steer its attention?
If the answer is “start coding” and you still needed a path, blast radius, or a next step, you found a steering miss. More context may still be useful later. It is not the first lever.
Before Session 2
Session 2 shows how to steer attention on purpose, starting with an explicit rule: let the agent come up with the path. It also explains why those clauses change tool use, search, scope, and stopping rules, not only the wording on the screen.