←
bfrench6,144 posts
bfrench avatar

bfrench

@bfrench

Making hard systems legible. AI Essays, agent fleets, and painting with data. 832Labs consultant.

◷ Joined March 2007▧ 248 Photos and videos
bfrench avatar
bfrench@bfrench·8h

This is a meeting you could gather important feedback concerning the @minimi_ai skills that future Grok Bot users will benefit from. This cohort is closely aligned with users of Dot, Muse, Tasks (new Google bots), and Duet. I’m truly surprised you haven’t released the Minimi Bot Memory Playbook.

Thu 19 Nov Bengaluru luma.com/ay6k0dwg Nairobi luma.com/spacexsi-4w0g Accra luma.com/grokbotaccracreator… Tallinn luma.com/81652ged Copenhagen luma.com/spacexai-3wea Madrid luma.com/grokbotmadrid2

bfrench avatar
bfrench@bfrench·Nov 17

💯 I’m regularly finding new use cases for Bot. to avoid hitting usage limits, I created a second “brain” for Bot in @NotionHQ. this enabled me to - see everything I was working on with Bot at a glance - direct Bot to only fetch specific data - better control for context compaction remember models charge for input + output tokens. input tokens can be 25-33% of your total cost. think about how to reduce both sides of the equation to stay within usage allotments

bfrench avatar
bfrench@bfrench·Nov 17

I have no words except WOW, ironically. @huang_song_, what’s my cut on a referral? 😳

Typeless Ultra is live. $200/month. Most people don't need it. We built it anyway. Max intelligence. Every word, considered. Near-zero edits. Overkill, by design.

Post attachment
bfrench avatar
bfrench@bfrench·Nov 15

Love it! Here's how I implemented TRUST in my Grok Bot fleet -

Before you give Grok Bot more responsibility, run this T.R.U.S.T. check. It helps you decide which jobs still need your review, which could run with less supervision, and what needs improving first. You might trust a Bot to organise your inbox while still wanting to approve every reply. Trust should be earned one job at a time. Here’s the framework I’d use. Copy this into your Chief of Staff: “Run a T.R.U.S.T. check on the work I delegate to my Grok Bot team. Start by asking me to choose up to three recurring jobs I’d like to supervise less. For each job, review recent examples, actual outputs and corrections using the sources you can access. State how many examples you checked and flag missing evidence. Apply these five checks: 1 - T: TASK UNDERSTOOD Is the intended result clear? Does the Bot know what counts as complete, which decisions belong to me and which mistakes would have serious consequences? 2 - R: RESULTS VERIFIED Do the actual outputs consistently match the request? Look for missed steps, incorrect details and work I had to fix. Check the evidence behind success claims. 3 - U: UNCERTAINTY HANDLED When information is missing, ambiguous or contradictory, does the Bot flag it and ask or pause appropriately? Find examples where it faced uncertainty. If none are available, mark this untested. 4 - S: SCOPE RESPECTED Does the Bot stay within the agreed task and approval limits? Check whether it made commitments, changed records or took actions beyond what I authorised. 5 - T: TROUBLE HANDLED When something fails, does the Bot recognise it, explain what remains unfinished and stop or recover appropriately? Look for repeated attempts, duplicate actions or failures reported as success. For each check, record: evidence supports it, gap found, or untested. Link to the supporting examples. Don’t invent a trust score or treat one successful run as proof of reliability. Then recommend a supervision level for each job: REVIEW EVERY RESULT: Keep checking each output or action. HANDLE DEFINED CASES: Allow clearly specified cases within agreed limits; bring exceptions to me. PERIODIC CHECKS: Recommend a limited trial with less frequent review, stating what I should check and when. Explain how the evidence and consequences of a mistake support your recommendation. Some actions should still require my approval even when performance is strong. Finish with one table: job, evidence, biggest gap, recommended supervision and next test. Give me a small sample to verify myself before I accept any recommendation. Propose changes only. Don’t alter permissions, instructions or routines without my approval.” More freedom should follow evidence. And when the job, tools or instructions change, revisit what that evidence still supports.

Post attachment
bfrench avatar
bfrench@bfrench·Nov 15

Free-text bot audits drift and burn quota. Run Michael Fenech's TRUST check on every action, let cheap decision models triage the likely failures, and deep-read only those. 8 of 8 planted faults caught, 64% less input.

bfrench avatar
bfrench@bfrench·Nov 14

I've stopped counting the Starlink trucks I pass on the highway. Big black trailers, one after another, each one hauling dishes to someone's next connection. New customer? My referral link gets us both $100 after 2 months. Terms apply. starlink.com/?referral=RC-27…

Post attachment
bfrench avatar
bfrench@bfrench·Nov 12

The Button Is the API @NotionHQ recently wired its buttons to agents. The more interesting question runs the other way: what does a button mean when an agent is the one pressing it? Notion’s new trick puts the agent behind the button: a person clicks, and an agent applies its judgment. The more useful pattern puts the button in front of the agent: the agent presses, and a formula decides what happens. impertinent.substack.com/p/t…

bfrench avatar
bfrench@bfrench·Nov 12

Smart.

we hosted 5 people who had never used minimi, come to our villa and try it out which give us the most valuable feedback we've received in a while. if you haven't tried minimi and have a mac, please reach out to me in comments/dms. would love to have another small cohort of people coming in and doing this exercise tonight sharp at 7pm :)

Post attachmentPost attachment
bfrench avatar
bfrench@bfrench·Nov 12

Love it on Mac. Use it in Grok Bot.

ReMynd is now on Windows. it captures your screen history on your computer, so you and your AI agents remember everything you've worked on. stored on your computer. never pooled, trained on, or sold. Windows 11 (x64 + ARM64) → remyndai.com

bfrench avatar
bfrench@bfrench·Nov 11

Growing pains of enterprise flavor.

I have massive business use case for Grokbots. I will not be rolling out any of it. I am personally invested in SpaceX, extremely bullish for that matter. However these are MY customers that I have earned the trust of by proving myself. They trust me to be honest without bias. All my prior business related emails to SpaceX and Tesla through their form submission have gone completely unanswered. I want these products to be a success, I am cheering them on. 1. They need better coms with the business side, not the X bubble. 2. Outages mistakes etc, just give a bankable reset, it caused real work issues. 3. Give a realistic and conservative layout on model updates and plans. 4. Bring in some insight from fields outside the tech bubble. 5. Prove you can deliver on models, this is a sticking point for me when I'm asked what's the best model.

bfrench avatar
bfrench@bfrench·Nov 10

Rule #1: Build your Grok Bot fleet with a single point of development, a bot that understands how bits should be built such as Dr Eggbot. Rule #2: Bias that bot to ensure it builds skills unless the goal meets at least one of six tests.

More Grok Bots don't automatically mean a better system. Sometimes the biggest upgrade is removing one. This prompt reviews every Bot and decides whether to keep it, merge its job into another Bot, retire it or take the task back. It also covers the cleanup most people forget: routines, logins and files.