Working well: habits that compound

What this lesson shows

Scope the task, give the right context, keep the window clean, plan, and verify.

The question it opens with

You know how it all works, but your sessions still wander and burn tokens. Which habits actually fix that?

The walkthrough, in the lesson's own words

• Same task, two sessions. Tap the bar you think burned more tokens.
• Here's how to read a habit card, the sloppy way → the deliberate way.
• Flip the other four. Each one ties back to a mechanic you already know.
• Same tools, far less flailing, it's all in the habits.
• The turn and token counts are illustrative, invented to show the shape, not measured sessions.
• Same context window, same rot, same loop. So why two such different bills?
• Right, the harness re-sends the growing session every turn; caching discounts the stable prefix (6.4), but wandering turns are never free. The fix isn't a smarter model; it's five small habits.
• Other way around. The scoped goal runs a straight line; the vague one wanders, and the harness re-sends the growing session every turn it wanders. Caching discounts the stable prefix, but extra turns are never free.
• Card 1 is open: front was the sloppy way, the back shows the deliberate way, the why, and the payoff.
• You can now drive any agent harness deliberately, not hopefully, curate the context, keep it clean, plan, and verify.
• The tools will keep changing, new harnesses, bigger windows, sharper models, but these principles won't. Each habit attacks a fixed property of the loop: the model only sees the context window, that window rots as it fills, every turn re-sends the growing session (caching discounts the stable prefix, but wandering turns are never free), and a tool result is a claim, not truth. Bigger and smarter models don't repeal any of those, so curating, clearing, planning, and verifying keep paying off.
• Two people give a coding agent the same task. One types "fix my app" and burns a huge bill wandering for 20 turns; the other scopes it tightly and finishes in 4. Same model, same tools. What separated them?
• Habits, not horsepower. The harness re-sends the growing session every turn; caching discounts the stable prefix, but wandering turns are never free. Scope the task, give only the right context, plan before acting, and verify the output, those compound on any harness.
• The growing session is re-sent every turn, so wandering turns multiply the bill.
• Bar length = turns the session took. Tap the bar that burned more tokens.

Key takeaway

You can drive any agent harness deliberately, not hopefully.

What you can do after this lesson

You can apply daily habits, scoping, context hygiene, verifying, to get more out of AI coding tools.

Where it leads: You're driving the harness well, but how does it plug into your tools and data? One protocol connects them.

This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.