Check + Carry Forward

Before the quiz

You have read six sessions. Before the quiz, one question to see what stuck, not as a test, but as a mirror:

Quiz

Six questions from the guide. Tap an option to see whether you were right.

Quick check

Tap an option to see whether you were right. Wrong answers are fine; they show where a reread might help.

0 of 6 answered

1. 1.You write “Implement soft delete for user accounts.” What did the agent most likely assume?

   Propose options and wait for approvalThe job is to start implementingCheck blast radius before any planAsk clarifying questions until criteria are complete

2. 2.What is the explicit path rule in an attention-steering brief?

   Pre-write every fork so the agent cannot wanderYou own the destination; let the agent come up with the pathNever ask for blast radius until after the PRAlways implement in the same message as the goal

3. 3.What is plan mode good at, and where does it often fail humans?

   Good at forcing planning; formal dumps can still make you lose the decisionGood at coding; bad at planningAlways enough on its own if the document is longUseless; skip straight to implementation

4. 4.Why talk through a plan in small conversational bites?

   Because agents cannot handle long promptsTo offload human working memory into smaller decisionsTo avoid ever writing anything downTo make the session look more collaborative in screenshots

5. 5.When should you summarize the plan?

   Before any discussion, as a formal dumpAfter discussion, as a short owned artifact, then actNever; chat history is enoughOnly after the PR is merged

6. 6.What path should this guide leave you practicing?

   Implement → fix → clarifyTalk → summarize → actPaste repo → hope → mergeLonger prompt → longer prompt → longer prompt

The whole story: one task, six sessions

If you only remember one thread, remember this: the same soft-delete task walked through every session so attention habits had something concrete to attach to.

The one-liner said implement. Session 1: the agent assumed start coding. More context was not the first fix. Attention was.

The attention brief asked the agent to come up with the path, plus blast radius and next step. Session 2: that rule is explicit; the clauses steer tools, search, scope, and horizon. Length without that steering keeps the old job.

Plan mode produced a dense dump. Session 3: the redirect is the gift; approving fog is the trap.

Small-bite talk decided Door A without a novel. Session 4: offload cognition into turns you can hold.

The summary locked destination, road, blast radius, next step. Session 5: then act. Chat alignment alone drifts.

Session	The essence
1, Wrong assumption	The prompt chose the wrong job
2, Steer the attention	Let the agent come up with the path
3, Plan mode’s gift and trap	Redirect yes; formal fog no
4, Talk in small bites	Human-sized decisions
5, Summarize, then act	Durable ownership before code
6, Check + carry forward	Quiz, summary, practice

When you are ready: one loop to try

On one real task this week:

1. Do not start with “implement X.”
2. Ask for high-level steps, blast radius, and next step. No code yet.
3. Talk in small bites until you can retell the road.
4. Write a short owned summary (template below).
5. Implement against that summary only.
6. In the PR, note the road you accepted and the blast-radius rails.

If you want harnesses, repo rules, and verification ladders after code exists, continue with AI-Native Engineering.

Quick reference

Attention brief shape

I need to: [destination]
Let the agent come up with the path.
Tell me what we should do.
Give high-level steps.
Check blast radius for: [areas]
Suggest the next step.
Do not implement yet.

Sticky rule: You own the destination. Let the agent come up with the path.

Talk loop

1) Paths (2-3, plain language)
2) Choose (criteria: smallest safe change, …)
3) Blast radius (top risks only)
4) Next step (one move)
5) Stop when you can retell the road

Summary template

Destination:
Chosen road:
Blast radius / rails:
Next step:
Accepted: yes/no

Mode check

If the path is…	Do this
Still open	Talk → summarize → act
Settled by ADR / checklist	Prescribe steps in the summary → act → verify
Only “aligned” in chat	Write the summary before coding

When to return

If…	Reread…
Agents still jump straight to code	Session 1 and 2
You approve plans you cannot retell	Session 3
Formal dumps freeze you	Session 4
Good talks become bad diffs	Session 5

Come back anytime the implement reflex returns. That is not failure. It is the old assumption meeting a fast tool again.