Your AI gets noticeably better when you stop typing the same instructions every time. Claude Skills and ChatGPT custom GPTs let you package the prompts, personas, and reference material you reuse weekly into a single tool you can call from any chat.

Join us Tuesday, May 26th at 10:00 AM PT / 12:00 PM CT / 1:00 PM ET for a 60-minute walkthrough showing how mid-market teams are building Skills for the workflows they actually repeat: client onboarding, weekly reporting, content drafting, vendor evaluation, and more.
Register here →

• OpenAI is preparing to confidentially file IPO paperwork in the next few days to weeks, with Goldman Sachs and Morgan Stanley leading. Sam Altman is reportedly targeting a September public debut at a valuation that could top $1 trillion, up from the $852B set in the private market last month. The filing push lands one day after Musk lost his breach-of-charitable-trust suit threatening OpenAI's restructure. The path-clearing didn't take long. (Bloomberg, CNBC, TechCrunch)

• An OpenAI reasoning model disproved an 80-year-old conjecture in discrete geometry that mathematicians treated as effectively settled. Paul Erdős's planar unit distance problem, posed in 1946, asked how many point-pairs can sit exactly one unit apart in the plane. The field believed square-grid arrangements were near-optimal. OpenAI's model produced a new infinite family of constructions that beats them, drawing on algebraic number theory. Fields Medalist Tim Gowers called it "a milestone in AI mathematics." It's the first time an AI has autonomously solved a prominent open problem in a major mathematical field. (OpenAI, Interesting Engineering)

• Coding agents moved from "chat assistant" to "autonomous worker" in a single month, and now they're running on a schedule. OpenAI Codex operates your computer on macOS, clicking and typing alongside you. Anthropic shipped /goal in Claude Code 2.1.139, which keeps the agent working across turns until a completion condition is met. This week OpenAI also added Codex Automations: scheduled tasks that run on a recurring interval, plus an auto-archiving inbox, mirroring Claude Cowork. The supervision question stopped being theoretical for any team running these tools. (OpenAI, VentureBeat, OpenAI Codex docs)

• Anthropic's enterprise push got concrete this week, with the customer math underneath it leaking out. Claude for Small Business launched with 15 prebuilt skills and connectors into QuickBooks, PayPal, Stripe, HubSpot, Slack, and Microsoft 365. PwC committed to training 30,000 professionals on Claude. Claude for Legal shipped with 12 practice-area plugins. Reporting separately cited 1,000+ customers spending $1M+ per year and $44B in self-reported annualized revenue.

Pricing is also shifting. Starting June 15, programmatic Claude usage will be metered out of a separate monthly credit pool, not your subscription. Claude Code, Agent SDK, and third-party agents authed via subscription burn against $20 (Pro), $100 (Max 5x), or $200 (Max 20x) in credit that doesn't roll over. Chat, desktop, and Cowork are unaffected. If your team automates against Claude, model the cost before mid-June. (Anthropic, PwC, Artificial Lawyer, InfoWorld, VentureBeat)

• Anthropic is in talks to rent Microsoft's Maia 200 AI chips, adding a third compute lever on top of its $100B Trainium deal and Google TPU agreement. Talks are still early-stage. The deal matters less than what's underneath it: even the frontier labs are now renting capacity from rivals to keep pace with demand. If you're planning around a single vendor's roadmap, the labs aren't. (CNBC)

• OpenAI stood up a separate $4B "Deployment Company" to embed engineers inside Fortune 500 customers. Backers include TPG, Bain, Brookfield, McKinsey, and Goldman. The new entity also acquired AI consultancy Tomoro for its 150 forward-deployed engineers. The labs are now competing with Accenture and Deloitte for the implementation budget, not just the API line item. (OpenAI, CNBC)

• SAP Sapphire 2026 put 200+ Joule agents into production for finance, supply chain, and HR. KPMG already has 270,000 employees using them. Salesforce added Multi-Agent Orchestration in the Summer '26 release. Microsoft brought Copilot Cowork to iOS and Android, so users can hand off tasks from a phone. The agent layer is now sitting inside the suites your team already opens every day. (CIO.com, Salesforce, Microsoft)

• The next 30 days reshape how Google and Apple do AI at the OS layer. At I/O this week Google unveiled Gemini Omni, a video generation and editing model built on a "world model" that reasons about physics like kinetic energy and gravity, with voice-driven character and background edits. Also new: Gemini 3.5 Flash, an Android-wide "Gemini Intelligence" agent that completes multi-step tasks across apps, and "Googlebook" replacing Chromebook as a Gemini-first laptop line. Apple confirmed WWDC for June 8 to 12 with Siri 2.0 and AI agents on the App Store. (CNBC, Google, eWeek)

• Cutting people isn't translating to AI ROI. A Gartner study of 350 executives at companies with $1B+ revenue found 80% of those who piloted AI reported workforce reductions, but reduction rates were nearly equal between high-ROI and low-ROI groups. The highest gains came from companies that used AI as "people amplification" rather than as replacement. (Fortune / Gartner, May 11, 2026)

• The frontline is hitting a "silicon ceiling" while leadership runs ahead. BCG's global AI at Work survey found 85% of leaders and 78% of managers use AI tools regularly, but only 51% of frontline workers do. Strong leadership support moves frontline positive sentiment from 15% to 55%, and five hours of training drives regular usage materially higher. (BCG, April 2026)

• AI is starting to reshape who gets hired into entry-level white-collar roles. Anthropic researchers found a 14% drop in the job-finding rate for workers aged 22-25 in highly AI-exposed occupations (computer programmers, customer service reps, data entry) versus 2022. Overall unemployment for highly exposed workers has not increased. (Anthropic, March 5, 2026)

Every 50-person company is running 5 to 10 unofficial workarounds: the spreadsheet bypassing the CRM, the Slack DM replacing the ticket, the manual rekey covering for broken automation. Most CEOs can name two. This prompt surfaces the rest, and names what each one is actually telling you about a broken official process.

Use it before quarterly planning, before any "improve operations" initiative, or after a new hire asks "wait, you still do it that way?"

The Prompt
Act as a senior operations diagnostician for a $5M-$100M company. I'll describe one function on my team (CRM, ticketing, expense, vendor onboarding, internal comms — pick the one I name) and the official tool or process we use for it.

Predict the 3 to 5 unofficial workarounds you'd expect to find given that setup — the spreadsheets bypassing the system, the DMs replacing tickets, the manual rekeys around broken automation.

For each workaround: name what broken official process it reveals, estimate what it costs in duplicate work or lost visibility per week, and give your verdict (Fix the system / Legitimize the workaround / Leave alone) with one-sentence reasoning.

Don't list things I already know about. Push back if my description of the setup sounds cleaner than reality usually is.

Return as a table: Workaround | What broken process it reveals | Weekly cost | Verdict (Fix / Legitimize / Leave alone) | One-line reasoning.

[PASTE FUNCTION + OFFICIAL TOOL/PROCESS]

Here's a small trick that changed more about my AI workflow than I expected: before I give Claude a task, I give it a role.

A few weeks ago I was reviewing some design assets and asked Claude to help me critique them. The first pass came back generic. "Good use of color." "Consider hierarchy." Surface-level stuff. So I added one line at the top of the chat: "Act as a seasoned B2B marketing graphic designer who's spent 15 years building decks and one-pagers for mid-market companies." Same designs. Same question. The next pass came back with the kind of pointed, specific notes I'd pay an actual designer for.

I think the reason is obvious in hindsight. When you ask AI a question with no context, it answers from the average of everything it's ever read. That average is mush. When you tell it who it is, you've narrowed the aperture before you've even asked the question. A "seasoned B2B designer who's tired of decks that try to do too much" answers very differently than "an AI assistant." Both are real options. Only one of them is useful.

If your team is getting mediocre output from AI, try this before you try a longer prompt or a better tool: three sentences at the top describing who the AI should be. Costs you nothing. Takes 60 seconds. Often gets you 80% of the way to "this is actually good."

One more thing while we're here.

This week's research was loud: cutting people doesn't get you the AI ROI you're imagining. The harder, better move is keeping the team and getting more out of every person. Most companies haven't asked "where could AI extend each person's reach?" because they're still stuck on the simpler question, "where could AI replace each person?"

If you want to figure out where AI actually pays off on your team, that's what we do for mid-market companies. The first conversation is short and free. Let's chat.