"Interview me in depth using free text, or askuserquestion tool when multiple choice works, about relevant parts of my life you don't know about yet and save it all to memory."
That's the whole prompt. On September 11, 2026, Thariq Shihipar (@trq212), who works on Claude Code at Anthropic, posted it as a direct answer to a quote from Blake Robbins (@blakeir): "i have long thought ChatGPT, Claude, etc. would be 10x more useful if they just spent 20 minutes interviewing me... really dig into my life, how I think, what I care about, what I'm trying to do." The reply crossed 232K views.
The idea is a small inversion of how people normally build up an AI's memory. Instead of memory accumulating passively — a "remember this" aside here, a correction there, over weeks of sessions — the interview prompt front-loads the whole process into one conversation, with Claude asking the questions instead of waiting for you to volunteer context.
TL;DR
| Question | Answer |
|---|---|
| What does the prompt do? | Has Claude actively interview you and save the answers to memory in one session, instead of memory building up passively over time |
| What tool does it lean on? | AskUserQuestion for multiple-choice branches, free text for open-ended ones |
| Who posted it? | Thariq Shihipar (@trq212), Anthropic — quote-replying Blake Robbins (@blakeir) |
| Does it fix stale memory? | No — Thariq's follow-up prompt (below) is for that, run periodically, not automatically |
| Is this the same as CLAUDE.md? | No — CLAUDE.md is a hand-written, project-scoped file; Claude's memory is account-scoped and updates live from conversation |
| Biggest criticism in replies | Staleness (Ishu Agrawal) and lack of curated, toggleable memory "blocks" (light*/@perceptronHuntr) |
Why this is a genuinely different move than "remember this"
Most people who've used Claude's memory feature build it the slow way: mention something, Claude notes it, move on. That works, but it means memory reflects whatever came up in passing — not necessarily what actually matters for Claude to know about you.
The interview prompt reverses who's driving. Claude proposes what it doesn't yet know, and picks the right input mode for each question — free text for anything open-ended ("what are you trying to build this year"), the AskUserQuestion tool for anything with a natural small set of options (timezone, industry, preferred communication style). That mode-switching is the detail that makes the prompt work in one shot instead of turning into an open-ended essay Claude has to parse for structure afterward.
It's the same shift Robbins described wanting: not a chatbot waiting to be told things, but one that treats "getting to know you" as a task worth spending a session on.
The follow-up prompt: keeping memory from going stale
Thariq didn't stop at the interview prompt. In a reply to Ishu Agrawal (@ishuagra02), who raised the obvious problem — "my biggest problem with memory is that it can get stale very quickly and that it isn't useful enough to risk that" — Thariq offered a second prompt:
"Interview me about anything in memory that might have gone out of date and update it."
Run periodically, this is meant to be the maintenance pass: instead of Claude silently carrying forward information that stopped being true (a job you left, a project you shipped, a city you moved out of), it re-checks with you directly. That's still a manual trigger, not automatic drift-detection — you have to remember to run the second prompt, which is a real limitation Ishu's original point still stands on. The interview prompt solves cold start; it doesn't solve staleness on its own.
What people pushed back on
The replies are where the actual design debate happened, not the original post:
- Ishu Agrawal (@ishuagra02) framed the core tension plainly: memory's usefulness is capped by how quickly it goes stale relative to how much risk you're willing to accept trusting it. A rich profile built in one interview is only as good as the next six months of your life not changing it.
- light* (@perceptronHuntr) raised a structural complaint: sometimes you want Claude to know about you, sometimes you don't, and right now memory is one lump rather than curated blocks you can sift between at will — closer to how people already organize separate notes per project than how account-level memory currently works. That's a request for scoping, not just accuracy.
- Zach (@ZryMiller) asked the practical question builders ask first: "Why not just add to CLAUDE.md?" — worth answering directly, since the two systems get conflated constantly.
Interview-me memory vs CLAUDE.md — they're not the same tool
Zach's question deserves a real answer, not a dismissal. CLAUDE.md and Claude's chat memory solve adjacent but different problems:
| Claude memory (interview prompt) | CLAUDE.md | |
|---|---|---|
| Scope | Account-level, follows you across chats | Project-level, lives in one repo |
| How it's written | Claude writes it, from conversation | You write it, by hand |
| What it holds | Personal context — how you think, what you're working toward, preferences | Stack conventions, commands, architecture notes |
| Update mechanism | Live during chat, or via a re-interview prompt | Manual edits, or /init to bootstrap |
| Best for | "Who is this person, across every conversation" | "How does this codebase work, in this one project" |
The pattern Anthropic has documented elsewhere — see explainx.ai's coverage of MEMORY.md as a cross-agent persistence pattern — is the more general version of the same split: durable facts belong in a structured, reviewable file; live context belongs in whatever the agent's native memory layer is. The interview prompt is a fast way to seed the second kind, not a replacement for the first.
Try it — and check what actually got saved
To run the technique yourself:
Interview me in depth using free text, or askuserquestion tool when
multiple choice works, about relevant parts of my life you don't know
about yet and save it all to memory.
Then periodically:
Interview me about anything in memory that might have gone out of date
and update it.
Two things worth doing before either prompt turns into a habit:
- Review what Claude actually saved. Claude's unified memory rollout keeps sensitive topics off by default and never saves certain PII categories outright — but "an agent decided this is worth remembering about me" is still worth a manual glance, especially the first time you run a deep interview.
- Scope it if you're on Team or Enterprise. Memory defaults and sensitive-topic handling are admin-gated on those plans — check your org's Capabilities settings before assuming the interview prompt behaves the same way it does on a personal account.
The security stakes here aren't hypothetical. explainx.ai has covered a real memory-exfiltration technique via web fetch — a reminder that account memory is a real asset worth protecting, not just a convenience feature.
Why this matters beyond one viral prompt
The interview prompt is small, but it's a preview of where personalization is heading: instead of users doing the work of deciding what an AI needs to know, the AI asks. That only works if the underlying memory system is trustworthy enough to act on what it collects — which is exactly what Ishu's staleness point and light*'s curated-blocks request are both really about. explainx.ai's field guide to Thariq's own work on Claude Code covers more of this same "make the agent do the structuring work" philosophy, including the thin-prompts, thick-artifacts framing that this interview trick is a small instance of.
Related reading
- Claude Unified Memory: Chat and Cowork Share One Context Layer
- What Is CLAUDE.md? Persistent Memory That Transforms Claude Code Sessions
- What Is MEMORY.md? AI Agent Persistence Explained
- Claude Memory Heist: Web Fetch Exfiltration
- Field Guide to Thariq Shihipar, Anthropic AI Engineer
- Thin Prompts, Thick Artifacts, Thin Skills
- What Are Agent Skills? Complete Guide
Prompt text and engagement figures are accurate as of September 11, 2026. Memory defaults and admin controls referenced here reflect Anthropic's August 2026 unified memory rollout and may change.