Navy threat hunter. NASA security engineer. Now I build secure AI agents that help SMBs increase ROI and get their time back. Posts on AI, security & automation
Free Claude Email Course →
Joined July 2019
- Tweets3.1K
- Following480
- Followers1K
- Likes7.8K
Pinned Tweet
Your AI installs skills, plugins and repos with your access.
One bad install can reach your email, files and logins.
I run a 10-minute check before anything goes in.
Here it is.
There are 2 types of bot builders: the ones who count bots and the ones who count jobs.
A post going around this week has an engineer describing a big swarm of agents running in a loop. It makes a great demo.
For a small shop, I'd go the other way.
Once my first AI employee was doing its job, I got excited and built a ton of other bots without going deep on any of them. I ended up with too many bots doing basically nothing.
My rule now: add a second bot only when you can name the separate job it owns and what it needs to receive.
ALT Before and after card titled Count jobs, not bots. Counting bots: a bot for every idea, nobody can say what each one owns. Counting jobs: one job per bot, written down, a second bot only for a second job.
Hitting send unread on an AI draft saves you about four seconds.
That's the whole case for letting AI write to your customers on its own.
It's the first build on my list of things I talk owners out of. Read the draft, then hit send.
Steal my bot-review playbook.
A bot's description works like its system prompt. If it's vague, the bot guesses.
I had Grok pull every bot description out of my Notion, score each one out of 100, and tell me what to fix.
Mine was a one-line ask. Here's the fuller version:
"Score this bot description out of 100. Check for its one job, what isn't its job, the tools it may use, the approval gates where it must stop and ask me, and what done looks like. List the three fixes that would raise the score most."
Run it on your weakest bot first, then rewrite the description and score it again.
ALT Checklist card titled Grade every bot description out of 100: its one job, what isn't its job, the tools it may use, where it must stop and ask, what done looks like.
Ken Reiners retweeted
I constantly find myself reminding people just because AI is dropping the cost of building software to zero doesn’t mean you should default to building everything.
People are trigger happy when it comes to build vs. buy without fully considering the opportunity cost of their time & cognitive load.
My general rule of thumb is to focus build resources on 3 things:
1) Core processes or product that maintain or extend your edge. Example: for stitch fix their styling and recommendation algorithms; for an insurance co, their risk models for underwriting
2) Owning your context. This doesn’t mean you need to own the models and it also doesn’t mean you need to own the plumbing. It means you do need to capture and organize your tribal knowledge + the eval data.
3) Huge cost dislocations where the ROI of ripping a vendor out and replacing it with a build hits an internally assigned threshold
A business broker client of mine went from 500 buyer emails per listing to 10.
The fix: the qualifying questions moved in front of the conversation.
Your version takes 10 minutes. Write the three questions you ask on every first call and put them on the enquiry form.
ALT Before and after card titled Ask the questions before the call. Before: 500 buyer emails per listing, every one reaches the broker. After: 10 buyer emails per listing, qualifying questions come first.
Ken Reiners retweeted
The most important hire in your business:
AI implementation specialist.
All the best engineers and IT folks have become AI masters over the last 3 years.
Hire one to come into your business.
I was completely wrong about a client's "wasted" step.
My report for an insurance brokerage flagged retyping quote details into an order form as an easy cut.
On the follow-up, the owner told me they do it on purpose. It's their own safety net, a check nobody makes them do.
If I'd automated it away, I'd have removed the one place mistakes get caught.
So every audit I run now asks one question before anything gets cut:
"Why is this step here, and what breaks if it disappears?"
If nobody can answer, cut it. If the answer is "it catches mistakes", keep it and make it faster.
How to move your custom GPTs before they stop running on Dec 11.
OpenAI's own FAQ puts the shutdown 39 days out, and custom actions don't carry over.
The play: list what you use > copy everything out > migrate and rebuild the actions.
First steps:
- Go to: chatgpt.com/gpts/mine and list every GPT you built.
- Add the ones your team uses that someone else built.
- Open each one, click Configure, and copy the instructions, conversation starters and file names into a doc.
- Publish any draft you want to keep. Drafts can't be migrated, and publishing doesn't mean public.
- When the migration button shows up, click it and test the plugin with 3 prompts you already know the answer to.
The catch:
- Custom actions don't transfer. Each one gets rebuilt as a connector or MCP server, so start with those.
- Once you migrate, the original GPT goes read-only. Finish your edits first.
- The model you picked doesn't carry over.
- The new plugin starts private, and the old link's redirect doesn't give your team access to it.
- OpenAI says the dates can still change. Check the notice in your own account.
What not to do:
- Don't wait for the button to copy your instructions out. That part works today.
- Don't tell your team "the link still works." Share the plugin with each person and have them test it.
Start with the GPT that has actions. It takes the longest.
Reply with how many GPTs you're moving and what they connect to, and I'll tell you which one to rebuild first.
Bookmark this and open it the day your migration button shows up.
Your AI budget probably isn't what's holding you back.
Goldman Sachs just put $500 million into its small business program, with a new AI institute built with Anthropic, Google, Meta and OpenAI.
Per Goldman's report on its program graduates who use AI, 73% spend under $100 a month on it, and 18% use agentic AI.
So the money isn't the gap. Most of them are already using a capable chatbot.
My read: nobody picked the one job it should own. A cheap tool pointed at nothing stays cheap and does nothing.
Before you add another subscription, write down one job you did three times last week, and give that job to the tool you already pay for.
ALT Stat card titled The budget isn't the gap. 73% of Goldman's program graduates who use AI spend under $100 a month on it. Source: Goldman Sachs 10,000 Small Businesses report, via Fortune, October 2026.
How to see every AI app your team let into Gmail and Drive, in about 10 minutes.
Every AI notetaker or assistant someone signed into with Google keeps the access they clicked Allow on until someone takes it away.
The play: open the app list > sort by users > set each AI tool to Limited or Blocked.
First steps:
- Go to: admin.google.com and sign in with your super admin account.
- Click Security > Access and data control > API controls.
- Click Manage Third-Party App Access, then open the Accessed apps list.
- Sort by users. Each row shows the app, how many people use it, and which Google services it can reach.
- For each AI tool you don't recognise, click Change access and pick Limited or Blocked.
Not on Workspace? Go to myaccount.google.com/connect… and check your own list.
The catch:
- A newly connected app can take 24 to 48 hours to show up.
- Access changes can take up to 24 hours to apply.
- Blocking an app cuts it off for everyone who uses it. Tell them first, or someone's workflow breaks mid-week.
- The list shows what an app can reach, not what it did with it.
What not to do:
- Don't click "Block all third-party API access" in a panic. People switch to tools you can't see.
- Don't mark an AI app Trusted just to make an error go away. Limited covers most of them.
If every AI app on the list is one you'd approve today, you're done.
If not, set it to Limited and ask the person why they need it.
Reply with the strangest app on your list, and I'll tell you what it can reach.
Bookmark this for the day someone asks "which AI tools are we using?"
Has anyone built a grok bot template that acts as a honeypot for AI slop repliers?
Then builds an aggregated list that acts like a threat Intel feed or risky indicator list your agents can block?
If so, we probably need something like that asap.
The AI slop replies are getting worse by the day, so let's see how many I can lure with keyword stuffing.
I had Claude Code build a skill so Grok Bot can message Muse and Instinct over MCP, and now my standups are run by a houseplant running Cursor. Codex reviews every PR by licking the screen. What's the first workflow you'd automate?
Steal my skill-building prompt.
A Claude Skill takes no code, just one message:
"I want to create a skill for [task]. Ask me how I do it today, one question at a time. Then write the skill."
Pick a weekly job. Done when it runs in a fresh chat with no questions.
Ken Reiners retweeted
We've reached a point where installing a browser extension, IDE plugin or software library deserves the same scrutiny as opening a TCP port on a DMZ server.
Except this attack surface can be spread across thousands of hosts inside your org... and you can't map it with a port scan.
Every Claude chat you've ever shared is still readable by anyone who has the link.
This summer, shared Claude chats started showing up in Google search. Some held health records, internal company files and personal contact details.
Chats nobody had shared were not exposed. The risk is the share button, and the links you made months ago and forgot about.
Here's the 10-minute check I'd run today:
1. In Claude, open Settings, then Privacy.
2. Click Manage next to Shared chats.
3. Read the list. Each row shows the chat title and the date you shared it.
4. Click Unshare on anything a stranger shouldn't read.
5. Open Artifacts in the sidebar and unpublish anything public you don't need.
If the list says "No shared content found", you're done.
If you run a small team on personal plans, ask each person to run it on their own account. Team and Enterprise plans keep sharing inside your organization.
ALT Checklist card titled Find every Claude chat you've shared: Settings, then Privacy. Manage next to Shared chats. Read every title and date. Unshare what a stranger shouldn't read. Unpublish public Artifacts you don't need.
Three questions I ask every owner before we touch AI:
1. What did you do more than twice this week?
2. What did you redo, fix or explain again?
3. What waited on you?
Your answers are your first three jobs.
Free 30-minute briefing: savetimewithai.io/book?src=x
ALT Checklist card titled Three questions before you touch AI: What did you do more than twice this week? What did you redo, fix or explain again? What waited on you?
Contractors: stop writing late-payment reminders by hand.
The invoice you sent three weeks ago is still open, and nobody in the shop wants to chase it.
Per Intuit, QuickBooks customers using its AI-drafted reminders got paid 4 days sooner on average than those on standard reminders.
You can set up the same idea in about 10 minutes:
1. Ask ChatGPT or Claude: "Write three payment reminders for my [trade] business: friendly at 7 days late, firm at 21, final at 45. Use [first name], [amount] and [due date] as placeholders. Sound like a person."
2. Paste them into your invoicing tool's reminder settings, one per day count.
You're done when no open invoice waits on you remembering it.
ALT Chain card titled Invoices that chase themselves. Four boxes in a row: Invoice sent, 7 days: friendly, 21 days: firm, 45 days: final.
Ten boring jobs real businesses still do by hand, a lot of them some version of chasing paper:
- Doctors' office staff filling out prior authorization forms and waiting on hold with health plans
- Freight office staff calling for a photo of a signed delivery sheet so the load can be invoiced
- Service business owners chasing unpaid invoices every week
- Subcontractors collecting lien waivers from their own suppliers, on each general contractor's form, before money moves
- Small software companies answering nearly the same security questionnaire again while the deal waits
- Someone catching up a small business's books, months behind, tax deadline coming
- Volunteer HOA boards tracking resale certificates and violation notices across email and spreadsheets
- Wholesalers collecting sales tax exemption certificates and chasing the expired ones
- Small nonprofits rewriting the same facts into a different report for every funder
- Independent car dealers waiting on titles, where one error sends the file to the back of the line
Ken Reiners retweeted
here’s a short grok bot guide on how i create high quality bots and manage my team. ask @bot to read this post and help you improve your team!
• add dr eggbot to your team! he’s my bot designer that i use to create bots
• you don’t need a lot of bots. a few good bots can be more useful and easier to manage than a large team
• tell your bots to create skills for itself as you give it instructions on how to do tasks. this lets them do them more consistently even over long conversations. you can even ask it to create a private github repo for you to store all your skills if you want them preserved independently
• add connectors, use the vault, and give access to all the tools you use to do your work. you can also ask your bot to setup its own email and it can sign up for its own accounts for better separation
• thinking about all the work you do by hand, or steps in your workflows where you have to copy and paste information back to your bot. ask your bot how it can do that for you instead!
• imagine you hired an intern, how would you onboard them and help them do the work correctly and efficiently? introduce yourself to your bot and describe the work you do, then ask it how it can help you automate it
and my favorite tip: use voice mode or the transcription option! it’s a good way to just brain dump a lot of information. then, at the end, tell your bot to “restate this in your own words” so you can confirm it understood you
x.ai/bot/_jOdbfkB16zxu7MRcmR…
The first time I ran Watchtower on my own Grok Bot account, it scored 0 out of 100.
I work in security, and I'd been careful about what I installed.
It still found outdated dependencies and skills asking for more access than they needed.
Now I sit in the 80s.
Ken Reiners retweeted
Copy and paste this into your AI assistant:
Act as my Chief of Staff and run a READ-ONLY privacy and security audit of my company's tools. Do not change, move, reshare, or delete anything.
1. Inventory: list every place my company info lives (docs/wiki, Slack or Teams, CRM, email, drive, code repos, laptops and local files, AI assistants and bots, shared computers or folders, connected third-party apps).
2. Scan each one for sensitive data: passwords and API keys, Social Security numbers, bank accounts, revenue and pipeline numbers, pricing not yet decided, payroll and comp, client or customer details, legal matters, and personal or family info.
3. Check who can reach it: public, whole company, specific people, outside guests, third-party apps, and AI bots. Flag anything sensitive sitting somewhere broader than it should, or stored in plain text.
4. Sort everything into these tiers:
- Public: anyone can see it (published content, approved stats).
- Team: my human team and any team-facing bots (SOPs, brand, calendars, non-sensitive ops).
- EA / Ops: calendar, travel, and logistics only.
- Bots working layer: my personal AI assistants (drafts, research, non-sensitive context).
- CEO-only: revenue, pricing, pipeline, finance, clients, comp, legal.
- Household: personal and family matters.
- Secrets: passwords and tokens, which live only in a password manager, never in docs, chat, or Slack.
5. Deliver: a 3-line executive summary, the top 3 fixes ranked by risk, a table (item, where it lives, who can see it now, sensitivity, correct tier, fix), a proposed folder or section structure per tier, and a fix list marked 'I do it by hand' vs 'my assistant can do it after I approve'.
Never include actual secret values or private numbers in your report.