Last week I told you the configuration is the product. I gave you a 5-section SOUL.md framework. Some of you copied the executive assistant template, customized it, and immediately saw better results. Great.

But a bunch of you replied with the same problem: "I sat down to write my SOUL.md and stared at a blank file for 20 minutes."

Of course you did. Writing a good system prompt from scratch is hard for the same reason writing a good bio is hard — you're too close to yourself to describe yourself clearly. You know what you want, but you can't articulate it in the structured way the agent needs.

There's a fix. It's called reverse prompting, and it's the single biggest unlock I've found for agent configuration.

Instead of you writing the prompt, the agent interviews you and generates the prompt from your answers.

Traditional prompting: you → prompt → agent → output. Reverse prompting: you → goal → agent → questions → you → answers → agent → prompt.

That extra loop changes everything. The agent asks questions you wouldn't think to answer. It surfaces constraints you forgot you had. It structures the output in a way that's optimized for how LLMs actually process instructions — because it is an LLM.

OpenClaw's bootstrap files are loaded into every single prompt. They're not one-time instructions — they're persistent context that shapes every interaction. That means poorly written workspace files don't just produce one bad response; they produce systematically bad responses, forever, on every turn.

Reverse prompting fixes the input quality problem at the source. Instead of you guessing what the agent needs to know, the agent tells you what it needs to know.

Step 1: The brain dump. Open a conversation with your agent (or any LLM — Claude, ChatGPT, whatever you prefer for long-form work). Paste the reverse prompting meta-prompt from this week's Blueprint below. Then just... talk. Describe your job. Describe your frustrations. Describe what you wish your agent did. Don't structure it. Don't edit yourself. Ramble.

The meta-prompt tells the agent to interview you — it will ask follow-up questions, probe for specifics, and push you past the generic answers.

Step 2: The generation. After 5–10 back-and-forth exchanges, tell the agent: "Generate my SOUL.md based on everything I've told you. Use the 5-section framework: Identity, Philosophy, Communication Style, Rules, Proactive Behaviors."

What you get back will be surprisingly good. Not because the AI is creative — because it extracted structure from your unstructured thoughts. It found the patterns in what you said and formalized them.

Step 3: The refinement loop. Paste the generated SOUL.md into ~/.openclaw/agents/default/SOUL.md. Restart with openclaw daemon restart. Use it for 3–5 days. Then ask your running agent: "Based on our interactions this week, what's missing from your SOUL.md? What rules should be added? What behaviors should change?"

Commit the updates. Repeat weekly. This is the compound effect I talked about last week — but now the starting point is dramatically better, so the iterations compound faster.

The agent shouldn't ask you "What personality do you want?" That's as useless as an empty text field. A good reverse prompting agent asks things like:

These are the questions that surface the constraints you don't know you have. The answers become precise instructions that generic templates never capture.

Reverse prompting works because it solves a fundamental asymmetry: you know what good agent behavior feels like when you experience it, but you can't always describe it in advance. An interview extracts tacit knowledge — the stuff you know but haven't formalized.

This applies beyond SOUL.md. You can reverse-prompt your way to a better AGENTS.md, a better USER.md, better tool configurations. Any time you're staring at a blank config file, the answer is the same: don't write it. Get interviewed for it.

Copy this entire block. Paste it into a conversation with your agent (or Claude/ChatGPT directly). Answer the questions it asks. After 5–10 rounds, tell it to generate your SOUL.md.

Why this works better than writing from scratch: The meta-prompt constrains the interviewer agent to ask one question at a time (preventing the overwhelm of a 20-question wall), push past generic answers (the "be helpful" problem), and cover all five SOUL.md sections systematically. The output format is pre-defined, so the generated file drops directly into your workspace with no reformatting.

Power move: After generating SOUL.md, say "yes" when it offers to generate USER.md and AGENTS.md. You'll get a coherent set of workspace files built from the same interview, instead of three files written in isolation that might contradict each other.

What it does: Gives your OpenClaw agent the ability to search the web in real-time using the Brave Search API. Returns results the agent can read, summarize, and act on.

Setup difficulty: Easy. Create a free Brave Search API account, grab the API key, add it to your skill config. Five minutes.

Verdict: Immediately useful for morning briefings, link summarization, and research tasks. Without it, your agent is limited to what the LLM already knows — which has a knowledge cutoff and can't tell you today's weather or whether your flight is delayed. With it, your agent can answer "what happened overnight?" with actual current information.

The free tier gives you 2,000 queries per month, which is more than enough for personal use. If you're running cron jobs that search on a schedule (like automated morning news digests), you'll want to monitor usage — it adds up faster than you'd expect.

Watch out for: The agent will sometimes over-search — running 3–4 queries for a simple question when one would do. Add a rule to your AGENTS.md: "When searching the web, use one well-crafted query before trying additional searches. Prefer specific queries over broad ones." This cuts unnecessary API calls significantly.

Rating: High value, low effort. Install after gog.

Here's the irony of Issue #1: I gave you a template for writing SOUL.md, and some of you — I know because you told me — copied it word for word and changed nothing except the name.

That's the generic copy-paste problem I warned about, applied to my own template.

Templates are starting points, not destinations. The whole point of the 5-section framework is that your Philosophy section should be different from mine, because you have different work, different communication preferences, different things that annoy you about AI.

Reverse prompting fixes this. You can't copy-paste an interview. The output is inherently personal because the input was personal. Two people using the same meta-prompt will produce completely different SOUL.md files — which is exactly how it should be.

The best agent configuration is the one that could only belong to you.

See you next week.

— Michael