Product case study · November 11, 2026
Use Claude Code plan mode to catch the decision you forgot
Practice Claude Code plan mode with an ambiguous shipping rule. Resolve the missing decision and define boundary tests before approving edits.

A useful plan exposes the decision that would otherwise become an invisible assumption in your code.
“Add free shipping over $50” sounds ready to build. Then someone places an order for exactly $50, applies a coupon, and discovers that the developer and the shop owner meant different things.
Claude Code can turn that sentence into code quickly. The useful reason to pause in plan mode is to catch the missing decision while it is still cheap to change. A long implementation checklist will not help if every step rests on the wrong rule.
This exercise uses a fictional shop, Bash on macOS, Linux or WSL, and an installed, signed-in Claude Code. It is a documentation-backed walkthrough, not a reported live Claude session. Official sources were checked November 11, 2026.
Give the ambiguity somewhere to live
Create a disposable folder and a tiny brief. These commands create practice files; they do not alter an existing project:
shipping_lab=$(mktemp -d /tmp/shipping-plan.XXXXXX)
cd "$shipping_lab"
cat > BRIEF.md <<'BRIEF'
Fictional shop exercise
Build a JavaScript function shippingCents(subtotalCents, discountCents).
Standard shipping costs 500 cents.
Orders over 5000 cents qualify for free shipping.
Inputs are integer cents, with 0 <= discountCents <= subtotalCents.
Decide whether the threshold uses the subtotal before or after discounts.
Ask the owner for that decision; do not silently choose it.
No UI, checkout integration, dependencies or deployment.
BRIEF
cp BRIEF.md BRIEF.before.txt
printf 'Practice directory: %s\n' "$shipping_lab"
The brief deliberately leaves the discount rule open. It already says “over,” so exactly 5000 cents is below the free-shipping threshold. That is a useful distinction: an agent should ask about the missing rule without treating every written requirement as undecided.
The prices here are invented practice inputs, not service pricing or a real shop’s policy.
Start with planning visible
Check your installed version, then launch the terminal session:
claude --version
claude --permission-mode plan
The CLI reference documents --permission-mode as a startup flag. Confirm the interface shows plan mode before sending the task.
Anthropic’s permission-mode guide says planning can read files and explore with shell commands while source edits remain blocked. There is an exception for interactive sessions where bypass permissions are available. Use a session without bypass enabled for this exercise; plan mode is not an operating-system sandbox. Shell-command handling also depends on configuration.
Paste this:
Read BRIEF.md and plan the smallest implementation.
Do not implement or change the brief.
First identify the product decision that is still missing and ask me about it.
Separate requirements stated in the brief from your proposed choices.
After I answer, list the files you would create and a small table of
inputs with expected shippingCents results. Stop at the plan for review.
If the flag is rejected or the interface shows another mode, stop and compare your installed CLI’s claude --help with the reference. Do not substitute a permission-bypass flag to get past an unfamiliar option.
Make one decision, then demand examples
The useful question is whether discounts change eligibility for free shipping. Answer with this fictional owner decision:
Use the subtotal after discounts. Free shipping requires a remaining
subtotal strictly greater than 5000 cents. Exactly 5000 still costs
500 cents to ship. Keep the valid-input assumptions from BRIEF.md;
input validation is outside this exercise. Revise the plan only.
Now compare the proposed table with these expected results, calculated from that rule:
| Subtotal cents | Discount cents | After discount | Expected shipping cents |
|---|---|---|---|
| 4999 | 0 | 4999 | 500 |
| 5000 | 0 | 5000 | 500 |
| 5001 | 0 | 5001 | 0 |
| 5500 | 500 | 5000 | 500 |
| 5500 | 499 | 5001 | 0 |
| 0 | 0 | 0 | 500 |
The last row may look odd: an empty order pays shipping. That follows from this deliberately small function’s rule. A real checkout might reject empty orders elsewhere. Keeping those responsibilities separate makes it easier to notice when a plan quietly adds product behavior.
My favorite row is 5500 minus 500. It catches both likely mistakes at once: checking the original subtotal and treating the threshold as inclusive. A happy-path example with a huge order would catch neither.
Anthropic’s best-practices guide recommends giving Claude a concrete check and making information it cannot infer from the repository explicit. This table supplies both. It does not need a framework, a database or an essay about architecture.
Review the plan as a small contract
For this exercise, a good proposal needs only a function file, a test file, the clarified rule and a command to run those tests. If it adds a checkout page or a dependency installation, send it back with the original scope.
If it skips the question and guesses the discount rule, correct it before approval:
That discount rule was an assumption. Replace it with my explicit decision
above, then recalculate all six cases. Keep this turn in planning.
When Claude offers plan approval, choose No, keep planning while any rule is wrong. The official guide also offers manual edit approval; approving a plan leaves planning and allows implementation under the selected mode. This exercise stops before that transition. Save the useful plan text in your own notes rather than approving simply to dismiss the prompt.
Exit the session with /exit (documented in the commands reference), then check the practice folder:
cmp BRIEF.before.txt BRIEF.md && printf 'Brief unchanged\n'
ls -la
cmp checks these two files only. The directory listing helps you notice unexpected local additions; neither command proves that nothing elsewhere changed. If the brief differs or implementation files appeared, inspect those changes and the session’s mode before continuing. Keep the folder as evidence instead of deleting it to make the check look clean.
You are ready to implement when the plan carries your explicit discount decision, all six expected results agree, and the proposed files fit the task. For your next real project, start with the Claude Code setup lesson, then use this pause wherever a plausible assumption could become expensive behavior. For a typo fix, I would skip the ceremony. For a rule that changes what customers pay, I would insist on the examples.
Sources and exercise scope
Primary documentation checked November 11, 2026:
- Claude Code permission modes: planning behavior, bypass exception and approval choices.
- Claude Code CLI reference: startup permission-mode flag and version command.
- Claude Code best practices: explicit requirements and verifiable checks.
The shipping policy, prompts and expected-results table are authored teaching examples. They do not report measured model performance or a successful agent run.
Put this to work
Judge a plan by its assumptions and expected behavior, not its length.
Try
Give Claude a fictional shipping brief with an intentionally missing threshold rule.
Prove it worked
Keep the clarified rule, expected results and a check that the brief stayed unchanged.
Keep in view
- Turn an ambiguous shipping rule into a concrete product question.
- Start a terminal session in plan mode and review its proposed changes.
- Define boundary cases before approving implementation.