Case Study · 2026
Claude Navigator
A pre-execution intelligence layer that tells you what Claude will do before it does it.
Solo product design & build — Growth PM exercise, NextLeap
01 — Problem
People distrust AI assistants not because the assistant is wrong, but because they can't see what it's about to do until it's already done it.
02 — Context
Every Claude session hides the same decision three times: what did the user actually mean, which capability should handle it, and why. Power users feel this as unpredictability; new users feel it as a black box.
03 — Discovery
Working through real prompts against Claude's tool-use behaviour surfaced a repeatable pattern: the ambiguity almost always lives in intent, not in output quality. Fix the intent read, and trust in the output follows.
04 — Hypothesis
If a user can see the assistant's read on their intent and its planned route to an answer before execution, they'll trust — and correct — the assistant faster than if they only judge it after the fact.
05 — Solution
Built a three-part pipeline in front of the assistant: an Intent Classifier that names what it thinks you're asking, a Capability Router that shows which tool or path it will use, and a Decision Transparency Layer that renders both as an editable step before anything runs.
06 — Product Thinking
The hard call was where to put the friction. Showing the plan on every single message would kill flow, so the layer only surfaces when the classifier's confidence drops below a threshold — trust is asked for, not demanded.
07 — Execution
Built as a working interactive prototype wired to a live API, with a user journey map to pressure-test the friction point and a success-metrics dashboard to define what 'trusted more' would actually look like in data.
08 — Outcome
The interactive demo became the flagship artifact from the fellowship's growth exercise — good enough to anchor the first long-form case study under the Young, Dumb & Learning byline.
09 — Learnings
Transparency features are a UX cost, not a UX feature, unless you're precise about when to spend it. The next version needs real usage data, not just a demo flow, to know if the threshold is set right.