The prompt history search
Finding the turn from last Tuesday in three keystrokes.
A user who's used your product for three months has written hundreds of prompts. Their best ones are buried. Scrolling the sidebar is a non-starter. Rewriting from memory loses the nuance. The product needs a search that's fast, fuzzy, and ranked by what the user actually does: open-recent, use-often.
The pattern is simple from the outside: one keyboard shortcut, one search box, typed results with highlighted matches. Inside, it's a ranking problem, and the ranking is where products win or lose.
"Your prompt history is your workshop. Make it searchable, and the product becomes a tool. Leave it buried, and the product becomes a chat."
⌘K. Fuzzy match. Recency-weighted.
One shortcut opens the search. Typing narrows. Every result shows the title, the first line of the body, and the last-used date, with highlighted matches. Filters live above the input: pinned, archived, all. A 'copy to composer' button ends every row. The user can go from thought to new thread in three keystrokes.
Pull enterprise churn for last 30 days, group by region
Yesterday, 18:20 · used 6×
Type, filter, copy. The history stops being a receipt and starts being a workshop.
Users don't rewrite their best prompts.
The best prompt a user will write in your product was written weeks ago. If search is bad, they either rewrite (badly) or give up (quietly). Either way, your product stops compounding. A good search turns history into a productive archive instead of a scrolling graveyard.
Search that deserves muscle memory.
- Weight recency and frequency. The right ranking beats alphabetical by an order of magnitude.
- Fuzzy, not strict. Users misremember words. A search that doesn't forgive typos is a search that teaches itself to be unused.
- One shortcut to the input. ⌘K is standard for a reason. Don't reinvent.
History that looks fuzzy but isn't.
A search that only matches the title, or that ignores the body entirely, feels broken. Users will type a word they remember from the body, see no results, and conclude the history is useless. The index has to cover everything: title, body, pinned state, even which tools were called.
What this pattern gets wrong when it gets wrong.
- Stale memory
- A persistent fact about the user that's out of date and silently poisoning answers.
- Leaky context
- Content from another source, session, or user surfacing in a place it shouldn't.
Three shipping variants worth copying.
- A ⌘K search that fuzzy-matches prompt text and title
- Results ranked by recency then frequency of re-use
- Each result has a 'copy to composer' shortcut