He used the tools as a bounded learning assistant so he could rebuild the conditions for deep work himself.
The chain: block, bounded helper, human transfer.
1. Real learning block
The workload was real: exam preparation in Algorithms and Data Structures with mixed tasks, time pressure, and transfer friction.
The first step was to stabilise one clear block: "Was ist die nächste Aufgabe, die ich wirklich heute lösen kann?".
2. Agent builds the learning room
The agent helped generate structure, feedback loops, and interaction affordances for exercises. The role was bounded to:
scaffolding, consistency checks, and progress visibility.
No closed-loop tutoring, no hidden grading, no automatic final answers.
3. Samuel does the work
The learning step stayed human-owned. Samuel still had to derive invariants, choose approaches, validate edge cases, and transfer patterns.
This is the proof surface: how often he could complete tasks by thinking through evidence instead of copying.
4. Visible evidence
The result is a public trail of interactions and interactions that are still interactive and honest.
Below are the concrete ADS artifacts that are safe to show publicly.