Source: raw/Claude_Code_for_normal_people_-_skills_voice_mode_and_how_to_collaborate_with_AI.md — How I AI (Claire Vo), guest Grace Clark, youtube.com/watch?v=o_eg2TtXAO0.
Grace Clark’s thesis is that “prompt engineering is dead, but intent engineering is where we need to be focusing our time” — and the concrete practice behind it inverts a pattern this wiki has repeatedly documented as best practice. Rather than the user interrogating themselves to produce a well-specified prompt, Clark makes Claude do the proposing and reserves the human role for reaction. The single most reusable artifact is her voice guide skill, which she describes not as a style guide but as “a think-like-me guide… the best corpus of how I make decisions.”
Key Takeaways
- “Do not interview me ever.” Clark explicitly rejects the have-Claude-interview-you pattern: “The work should never be on you. Claude has immense amounts of context if you’ve connected it right. It should be studying you and then coming back with a really strong idea that you can react to.” The pressure goes on Claude to produce a proposal; the human’s job is to react to it.
- Voice-first capture, on mobile, while walking. The mobile app is “incredible for getting things out of your head in an unfiltered way.” Her worked example: 2–3 minutes of speech describing the problem and a rough want, ending with “Come back to me. What do you think this process could be? Is this even possible?” Claude returned a concrete proposal (a password-protected interactive artifact), and the conversation went from there.
- The ratio is the point. Roughly 10 minutes of talking, then an hour of Claude building — with the human pressure-testing and pasting screenshots of what they did and didn’t like. The prompt stopped being “a hyperengineered chunk of text” and became a three-to-four minute conversation.
- “HTML is the new markdown.” She no longer reads markdown unless teaching — she wants it visualized. Output is judged as a rendered artifact, not a document.
- The voice guide is the highest-leverage skill, and it autofires. It runs in “almost every single thing I do.” It contains her philosophy of communication, words she uses, and words she doesn’t — the anti-AI-slop list sitting at the bottom, not the top.
- You build it by asking Claude to study you. The originating instruction was simply “I think I need a guide that will make everything more like me,” and Claude proposed making it a communication-and-philosophy guide rather than a word list. This is intent engineering applied to building the intent-engineering tool.
- It is her teaching opener — “the number one thing I teach in my class because it is a quick win. It forces us to learn the power of ingesting context, and it teaches students how to push and collaborate with Claude and then commit something so that it is an invocable skill.”
- Automation targets the inbox, not the IDE. Her framing of email — “everybody has a key to your front door and can come leave things in your house” — motivates a Gmail-draft skill that pushes a draft and opens the browser, so replies get handled or dismissed without living in the inbox.
- Adoption is a muscle-memory problem, not a benefits problem. Vo’s line — that people need to feel the benefits to keep using it — is contradicted directly: “That’s actually not true. People mostly need to understand that we’re going to learn to collaborate. The real hump is defaulting to this and building the muscle memory of simply opening an app.”
Why the “don’t interview me” stance is worth taking seriously
It sits in genuine tension with guidance elsewhere in this wiki. R.I.T(E) and the broader prompting best practices push toward more front-loaded specification, and the interview pattern is a common way to get there for people who struggle to specify.
Clark’s position is not that specification is unnecessary — it is that the model should extract it rather than demand it, and that this is only true once context is genuinely connected. Her caveat is load-bearing and easy to miss: “Claude has immense amounts of context if you’ve connected it right.” Without that context the model has nothing to study, and proposing-before-asking would produce confident irrelevance.
The honest read is that these are two positions on a context-availability axis, not a contradiction: interview-first when the model knows nothing about you; propose-first once a voice guide and connected surfaces exist. That reframing makes the voice guide the prerequisite rather than a nice-to-have.
Try It
- Build the voice guide first. Start with the actual originating prompt — “I think I need a guide that will make everything more like me” — and let Claude propose the structure. Commit it as an invocable skill so it autofires rather than being pasted.
- Put the words-I-don’t-use list at the bottom, not the top. Clark’s guide leads with communication philosophy and decision-making, and treats banned vocabulary as a trailing detail. That ordering is the opposite of most anti-slop prompts and worth A/B-ing against the slop guide’s approach.
- Try the mobile-voice → propose → react loop on your next ill-defined task: speak the problem for two minutes, end with “what do you think this could be, is it even possible?”, and react to what comes back rather than specifying up front.
- Judge output as rendered artifacts. If you are reading raw markdown to evaluate work, you are doing the model’s presentation job — the “HTML is the new markdown” point, and the same reasoning behind the HTML-first output pattern used elsewhere in this project.
- For rollouts, target defaults not enthusiasm. If adoption is a muscle-memory problem, the intervention is changing which app people open first — not demonstrating ROI harder.
Open Questions
- How much of this depends on connected context? The approach is presented as general but conditioned on “if you’ve connected it right.” No account of what minimum connection is required.
- Does propose-first degrade for high-stakes or unfamiliar work? Reacting to a proposal is cheaper than specifying, but it anchors you to the model’s first framing. No discussion of when that anchoring is harmful.
- What is actually in the voice guide? Described by category (philosophy, decisions, words used/avoided) but not shown in full, so it cannot be replicated exactly from this source.
- Is the class curriculum published? She references teaching this; no link surfaced in the transcript.
Related
- Claude Prompting Best Practices — the specification-first tradition this argues against, and the axis on which they reconcile.
- Marketing Voice Prompt Patterns — the closest existing treatment of encoding a voice; Clark’s “think-like-me guide” is a broader version.
- Anti-AI Slop Guide — the words-I-don’t-use layer, positioned differently here.
- Agent Skills Overview — the mechanism the voice guide is committed as.
- Browser vs. Computer Use (How I AI) — same show, same format.
- Claude Onboarding Course — the non-engineer audience this episode targets, and where the muscle-memory framing applies.