Ramble-Mode Onboarding
Onboard an AI agent by voice-rambling everything you need, not by wiring up APIs and structured fields.
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 3
- Confidence
- 90%
The highest-bandwidth way to configure an LLM agent isn't building integrations or filling structured onboarding fields — it's simply talking to it in plain, messy natural language about who you are and what you need. Doing this by voice on your phone the first time captures far more context, faster, than any form.
Origin
Credited by Claire Vo to Hillary Gridley, who framed it as the 'Yappers API' — the idea that chatting is the highest-bandwidth interface to an LLM.
Core principles
- 01Chatting is the highest-bandwidth API to an LLM
- 02Messy natural-language context beats sparse structured fields
- 03Voice removes the friction of typing and captures more nuance
- 04Describe the outcome you want, not the mechanism to achieve it
How to run it
- 1
Onboard by voice on your phone first
On first setup, use a voice note (e.g. in Telegram) instead of typing. Hold the record button and just talk — the agent transcribes it (via Whisper) and makes sense of it.
Pro tip It can talk back to you in voice too, making the whole onboarding conversational.
- 2
Ramble the full context, unstructured
Say who you are, what you run, your constraints, and the messy reality of what you want — 'I'm Claire, I run ChatPRD, I also have this podcast, my inbox is disorganized, here's what a good morning looks like.' Don't pre-structure it.
Pro tip Include the frustration and the desired end-state ('tell every recruiter with a bad pitch to kick rocks') — the agent infers the rules from your intent.
- 3
Let the agent interview you back
A good agent turns your ramble into follow-up questions — names, ages, activities, hard constraints, goals — and writes the answers into its own identity file. Answer conversationally and let it build the structure.
Pro tip State hard constraints explicitly ('nothing after 6:30, that's bedtime mode') so they get encoded from the start.
In the wild
Claire onboarded a new agent, Q, live by telling it in plain language it was to be 'an elementary school teacher and ex-professor/scientist' helping her kids with academics. Q interviewed her back — full names, ages, interests, homework load, off-limits family time — and wrote its own identity.md from the conversation.
→ A configured kids' homework-planning agent built entirely from conversation, with no structured forms or manual file editing.
Common mistakes
Reaching for APIs and structured fields first
Defaulting to 'is there an API for this?' or forcing users through rigid onboarding fields wastes the LLM's real strength. Rambling in natural language conveys far more context far faster.
Feeling you must author the config by hand
Users get overwhelmed thinking they must write all the identity/soul details themselves. In reality the onboarding conversation generates the file — you just talk and then iterate.
Is it for you?
Best for
Anyone setting up a new agent (or non-technical users intimidated by configuration) who wants the fastest path to a well-contextualized agent
Not ideal for
Highly precise deterministic configs where exact structured parameters matter more than rich context
From the transcript
“the highest bandwidth API for an LLM is just chatting to it”
“I often tell people when they're onboarding to an open claw, do it on your phone and do it in voice for the first time”
“the more information you give it, the more it's going to start to like try to discover about you and figure out tasks it can…”
From the episode
From skeptic to true believer: How OpenClaw changed my life
Claire Vo