From directing one tool to choosing between two

Two tools, two tabs, one question the intern courses left open

By the end of the intern-level AI courses, "reach for an agentic tool" meant something specific. You opened Cursor or Antigravity, pointed it at a request, and watched it read a file, write code, run a command and read the result. "AI tools: introduction (Cursor & Antigravity)" checked that loop against both tools' documentation. "AI prompting: zero-shot Ruby" covered phrasing a request precisely enough that the tool has nothing left to guess. Neither asked which model does the reasoning inside that loop.

A developer with Claude open in one tab and Gemini in another runs into that question on day one: which one gets today's hard task? The intern level already compared the editors. This course is about the models doing the thinking, Claude and Gemini. It focuses on the one place their differences are worth planning around: a large, tangled task rather than a five-line method.

The task that makes the difference visible

published?, the running example through most of the intern level, doesn't need this course. It's a few words of business logic with one branch. This app's Enrollment::UserGamificationQuery, at app/queries/enrollment/user_gamification_query.rb, does. In 48 lines it calls seven collaborators: Enrollment::ExperiencePoint, Enrollment::Level, Enrollment::LevelForXp, Enrollment::SkillRadarQuery, Enrollment::LeaderboardSnapshot, Enrollment::LeaderboardPercentile and Enrollment::StreakPolicy. It computes a leaderboard percentile from the learner's entry in a precomputed snapshot, with special cases for a learner who isn't ranked yet and for the top spot. It also computes a level-progress percentage with its own division-by-zero guard. Explaining it correctly, or extending it without breaking a collaborator, means holding all of that in mind at once.

That's the shape of task this course cares about: one where a wrong assumption made early quietly spoils a correct-looking answer by the end. Lesson 6 comes back to this file and tests lessons 2 through 5 against it.

What differs, checked against current documentation

None of the reasoning mechanisms here come from memory or an old blog post. Claude's and Gemini's current official documentation, read while writing this course, is the source for each specific claim in the next four lessons. Where a claim couldn't be pinned down, it's left out. Model lineups change fast, and both providers have shipped new models in the past few months, so a number from six months ago can already be wrong.

What this course covers

Lesson 2 covers what a "thinking" mode is, in both APIs' documentation. Lesson 3 compares the dials each one exposes for how much a model reasons, which is a different thing from writing a longer prompt. Lesson 4 compares context windows, checked against both providers' current model pages. Lesson 5 covers what reasoning depth costs in tokens and time. Lesson 6 applies all of it to UserGamificationQuery. Lesson 7 pulls the comparison together.

None of this replaces the review habit taught in "AI principle: human in the loop". It applies to a seven-collaborator query as much as to a one-line method. This course is about choosing how much reasoning a task needs before you get to reviewing the answer.