When to ask for deeper reasoning

The three-gear model from Foundations — fast, reasoning, deep research — applies at every desk, but the careful workhorse has its own relationship with the middle gear. As of September 2026, Claude offers extended reasoning that can be dialed up for harder problems, and knowing when the dial earns its wait is a per-desk skill. This lesson is that skill, for this desk.

What deeper reasoning actually buys here. Not smarter prose — better load-bearing structure. The difference shows on tasks with internal constraints that must all hold at once: a schedule where nothing overlaps, an argument where section five can't contradict section two, a plan whose steps genuinely depend on each other. Fast mode drafts each part well; the reasoning dial is what checks the parts against each other before you see them.

The trigger list. Reach for deeper reasoning when the task has: many constraints that interact; numbers with consequences (paired, always, with the shown-steps protocol); a long document where the answer depends on distant sections agreeing; or a judgment call where you'll want the trade-off structure, not just the verdict. Skip it for drafts, rewrites, and anything your own read will verify in a minute — the everyday majority.

Ask for the reasoning you can audit. Whatever the mode, the habit that matters is yours: "before the answer, lay out the constraints as you understand them and how they interact." A constraint list you can correct — "the kiln can fire twice a week, not once" — fixes errors at the cheapest possible moment: before the reasoning runs on wrong premises. Deeper thinking on wrong premises is just wrongness with better posture.

Long thinking and long documents together. The strongest combination at this desk: heavy material plus deeper reasoning plus a structured output shape. "Using the attached contract and the decision log, reason through whether clause 8 conflicts with what we agreed in March; show the chain." This is where careful earns its name — cross-references your own read would take an hour to trace, walked in a minute with the chain visible.

Budget honesty, one more time. Deeper reasoning costs more time and, on paid plans, draws limits faster — the same aiming rule as everywhere: volume on the fast path, the dial for the handful of tasks a week whose structure genuinely interlocks. If you can't name the interacting constraints, you probably don't need the dial; naming them is the test.

Mode names drift; the trigger list doesn't. What the dial is called, where it sits, how many notches it has — all of that is as of September 2026 and will be rearranged by some future redesign. The durable asset is your trigger list: interacting constraints, consequential numbers, distant-sections-must-agree, trade-off structure wanted. When the interface changes, re-find the dial in a minute and carry on; when a new depth option appears, test it against the same list before adopting it. Furniture moves; the sorting rule is yours.

One mistake to skip: using deep reasoning as a truth serum. The dial improves structure-keeping, not fact-knowing — a hallucinated fact reasoned about carefully comes back polished, not corrected. Verification stays a separate discipline with its own three steps, and the next lesson applies it here. Mode chooses depth; only checking chooses truth.