Let the Model Think
Updated
IntermediateWhat you'll learn
~14 min- Raise or lower reasoning effort to match tasks where logic determines the answer
- Ask for a plan or decision record (assumptions, options, chosen path) instead of a transcript of the model's thinking
- Recognize prompting habits current models no longer need: visible scratchpads, magic phrases, ALL CAPS
Before you change anything, give me a short decision record:- Assumptions: what you are taking as given about the data and the request- Options: two or three ways to do this, one line each- Chosen path: which one, and the reason it wins- What would change your mindThen do it.That prompt does not ask the model to think. It already does: current Claude, GPT and Gemini models reason internally before answering, at a depth you set but in text you do not read raw (Anthropic, OpenAI, Google). The prompt asks for the product of that thinking in a form you can check.
So you have two levers: how hard the model thinks is a setting, and what you get to inspect is a deliverable you ask for.
Lever 1: raise effort when logic determines the answer
Giving Good Instructions names the control in all four tools (/effort in Claude Code and Antigravity, effort inside /model in Codex, --reasoning-effort in Copilot CLI); the Vibe Coding Reference lists the levels. This lesson is about when to move it, and the three vendors agree:
- Raise it where an easily missed case decides the outcome: totals, a trade-off with competing constraints, a bug in code you did not write, a schema change (Claude Code:
/effort highorxhigh; Google: “complex coding, math, or multi-step planning”; OpenAI: “increase reasoning effort for difficult work”). - Lower it for brainstorming, a rename, recall or reformatting. There is no chain of logic to walk.
- Keep the top level rare. Claude Code’s guide says
max“tends to overthink” on ordinary tasks. For one hard question mid-session, Claude Code acceptsultrathinkin that prompt alone.
- Typed in the terminal:
claude --dangerously-skip-permissions - You type in the session:
/efforthigh - Mode:Effort: high, for this session
- You type in the session: The Q3 totals in reports/q3.csv don't match finance-summary.xlsx. Find out why. Before you change anything, give me assumptions, two or three options, your chosen path and what would change your mind. Show the mismatched rows as evidence.
- The tool did:Read reports/q3.csv
- The tool did:Read finance-summary.xlsx
- The tool replies: Assumptions - finance-summary.xlsx is the system of record Options 1. Re-sum q3.csv from raw transactions 2. Diff the files row by row and explain each gap 3. Overwrite q3.csv from the summary Chosen path: 2. Option 3 would erase whatever caused the gap. What would change my mind: the summary was hand-edited after export. Evidence - 3 rows dated Sep 30 are in q3.csv only (time-zone cutoff) - 1 refund is counted twice in q3.csv No files changed yet. Fix the refund and flag the cutoff rows?
Illustrative session; the exact wording varies.
The setting bought the depth; the prompt made the result checkable. You can reject an assumption in one line before any file changes.
Lever 2: ask for a plan or a decision record, not a transcript
- A plan before multi-step work. All four tools have a plan mode that proposes without editing: the plan, approve, build rhythm in Module 06’s Claude Code lesson;
/planin Codex and Antigravity;/planor Shift+Tab in Copilot CLI. - A decision record for a judgment call: assumptions, options, chosen path, and what would change the call. That last line tells you what to check.
- The evidence behind a result: the line items behind a total, the rows that differ. Anthropic lists “a short explanation, the evidence behind a result, or a summary of the actions it took” as what you can still ask for (refusals and fallback). For quotes from a source, Ground the Answer teaches quote-first; to pin the record’s shape, see Structure and Format.
This lesson used to teach a visible <scratchpad>. Retire it, and these habits with it:
- Visible scratchpads and thinking tags. On Claude Fable 5.2, Fable 5.1, Fable 5, Opus 5.5, Opus 5 and Sonnet 5.5, “a prompt that asks the model to write out its reasoning, for example in
<thinking>tags, may be declined” (Anthropic best practices). So may areasoningfield in JSON output (refusals and fallback); check saved prompts and skills too. - “Think step by step” as a magic phrase. OpenAI: prompting reasoning models to “think step by step” or “explain your reasoning” is unnecessary (reasoning best practices). Google: “it’s generally not necessary to have the model outline, plan, or detail reasoning steps” (prompting strategies), though it still says a short “Think very hard before answering” can help. Claude Code lists “
think hard,think harder,be thorough” under “What you can stop doing”: they nudge within the effort setting, not past it (best practices). - IMPORTANT and ALL CAPS. The same list says emphasis “mostly makes Claude over-apply the rule”; a rule that cannot bend belongs in a hook or deny rule.
On an older model with thinking off, asking it to reason first is still Anthropic’s documented fallback. OpenAI and Google do not say such a request is refused, only that it is unnecessary.
🧬In Your Field: Biotechclick to expand
Make the calculation a deliverable. For dilution, molarity or coverage math, raise effort and ask for a table: each step, the formula, the value, and the stock concentration assumed. A second scientist checks the table, and the table is the record.
📊In Your Field: MIS / Businessclick to expand
A decision record for build-vs-buy. Ask for the criteria, each option scored against them, the recommendation, and “what would change this call,” so you know when to revisit it.
You need a multi-step budget calculation to be right AND checkable by a colleague. On a current model, what is the best approach?
Key takeaways
- The model already thinks. Your job is to set how hard (the effort control) and to ask for something you can check.
- Raise effort where logic decides the answer; lower it for fast iteration; keep the top level for rare, hard, unattended problems.
- Ask for a plan, a decision record or the evidence behind a result, not a transcript of the thinking.
- Retire visible scratchpads, “think step by step” and ALL CAPS (as of 2026-11-07).
Next: Structure and Format