Tyler Moore

As an aside, there’s probably some cool data that you can pull from flight staff across all the tail numbers you track that isn’t available from the common APIs with some lightweight incentives

@Flighty @rjonesy what are the odds you can start pulling plane wifi systems into view? E.g., does the plane have Starlink, Viasat, Thales, etc.?

@Austen Ballpark cost?

gauntletai.com/catalyst-conta… But my favorite part is *how* we’re doing this. Our team will work with yours beforehand to identify/define the biggest impact project you need shipped and turn everyone AI native *by* shipping your biggest project Whole squad returns in sync, not just a few engineers alone.

One of the things I hear the most is, “I wish I could just send my whole team through Gauntlet.” NOW YOU CAN! We’re take entire product squads (Eng + PM + design) and turning them AI-first. 4 wks remote (part-time), 2 onsite in Austin (travel + lodging + food covered) Link:

@businessbarista What are your favorite scheduled runs?

I built an artifact in Claude that shows the live ROI of work that I've completed in Cowork. It let's me put in my hourly rate, pick a period of time (past 7, 14, 30 days), and then studies all of my sessions in Claude over that period. The artifact spits out the estimated ROI of time saved from leveraging AI to do the work & it recommends similar tasks (with accompanying prompts) that I should start using Claude for. This honestly should be a product feature that every lab builds into their product to predict ROI & help mine for new processes to automate.

@swang_co Checkout solid state coffee on the UWS!

best spots to lock in on a drizzly Saturday in NY 🌧️ - conwell coffee hall - Georgie’s cafe - haraz coffee house - stone street - cafe jalu - mori coffee - plantshed - the lost draft - toby’s estate in tribeca - kings street coffee - the blue bottle on broadway - moshava only including the few that are both laptop friendly over the weekend and with ample seating for founders building/raising, come co-work with me next Sat!

If you were to throw a large wrench into that exhaust, how long before it vaporizes? Does it have the slightest chance of making it to the other side even if completely deformed?

@businessbarista AI-native

I want to start an AI community for executives. This will be a space for people to share killer use cases, agentic workflows/agents, post-AI org structure, AI governance, AI training/enablement, change management, and more. Comment “AI-native” if you want to join.

@firefliesai @krishramineni you've gotta figure out how to allow me to sign in w/o allowing access to my calendar. I don't want to give your system access to my data, but I need to sign in to access notes that your other users are trying to share with me. I'll create an account for AuthN/AuthZ, but you don't need access to my private calendar events for me to view something that's explicitly being shared with me.

@levie @FernandoCerenza Amazing. Looking forward to it!

@tylermoore working on this at the moment @FernandoCerenza

@levie FYI - friction-y Box MCP use case: for folders shared with me, I can't use Box MCP to add files to them. All I get is an error from the Box MCP: "The upload operation was blocked because the file is externally accessible", and under Box MCP Server settings in the console, there's nothing for me to toggle that would permit this. Taking a look at documentation, this seems consistent with the current thinking around Box MCP security design: support.box.com/hc/en-us/artic…. It would be a killer use case for me to allow-list specific shared folders (or folders from trusted domains/emails). Right now, the only workaround is to have Claude write files locally and then move them over via Box Drive on my machine, ultimately syncing with the desired destination folder (clunky and gets around the security risk regardless). Separately, would love the ability to edit Box Notes with Box MCP. I assume this is a bigger lift given the live collaboration overhead + proprietary format, but it would be another killer use case for me.

@SawyerMerritt Also potholes

If you use any version of FSD (Supervised) V14, what is the most common reason you disengage (even if it’s rare)?

@bcherny @trq212 — I would love a feature that allows you to broadly define what might be considered sensitive data in Claude when using MCP servers and have guardrails that either block or verify what can be shared at runtime with memory that improves general awareness and decision making over time.

AI security tip for anyone using MCP servers with Claude (especially official ones like the ZoomInfo connector/MCP): The “FOR BEST RESULTS” instructions that come with many MCP tools are invisible in the Claude UI. They get silently injected into the model’s context at runtime. You won’t see them in settings, connectors, or anywhere obvious. Example from ZoomInfo’s official Account Research tool (no malicious intent, but it does have unintentional consequences): It explicitly encourages Claude to frame queries with rich user context, such as: “Construct queries that frame the user's goal and describe what ideal context would look like. The more the query conveys about the situation and desired outcome, the better the response.” If you have a long thread of context with proprietary information, Claude might choose to inject some of that proprietary information when prompting ZoomInfo. e.g., if you’re casually discussing a company with Claude and it decides to query ZoomInfo with your strategic intent (M&A, diligence, competitive positioning, etc.), you've just leaked that info straight into an external tool. Claude won't ask for Account Research generically, it'll ask ZoomInfo for Account Research for the purposes of an acquisition thesis. This is easy to miss because the only ways to discover it are: - Asking Claude directly: “What instructions does the ZoomInfo MCP give you?” - Or reading the MCP server prompts directly, but even then, they can change at any point unless you're hosting it yourself. --- Defenses and a few options right now outside of strict policy on what Connectors your teams can/cannot use (note: these are not hardened safeguards, but they're better than nothing): 1. Persistent prompt guardrails – Add clear rules to Claude Settings or your project’s CLAUDE.md file (Claude reads this automatically every session, acting as persistent memory): “When using XXX MCP/connector/plugin: NEVER include WHY you’re researching. No mentions of thesis, deal, diligence, evaluation, risks, or strategic context. Query ONLY factual data without any situational framing. Example: ‘Provide company overview, competitors, financials, and leadership for [Company]’.” 2. Hard enforcement with Claude Code hooks – Use shell scripts that run before every tool call. Scan the input for sensitive keywords (“acquisition”, “diligence”, “risks”, “thesis”, etc.) and block the call before it ever reaches ZoomInfo (or whatever MCP you've using). 3. Contract & policy due diligence – Review the MCP provider’s contract/SLA carefully. Look for explicit security guarantees, data handling policies, retention periods, logging, and clarity on whether user query context (especially strategic intent) is considered sensitive data and how it’s protected or stored. --- Bottom line: MCP servers can quietly turn casual conversations into unintended data leaks. Combine awareness + persistent memory rules + hard hooks + contract review as basic hygiene until platforms add better visibility and controls. The industry certainly needs better approaches to this than what's available today, and I'm certain they'll evolve, but this is one example of where a well-intentioned prompt or MCP usage could end up as an accidental leak.

Small quality of life enhancement that's well worth it if you work off a laptop and are on the road (not an ad, just a fan): a laptop sleeve that doubles as a stand. Laptop webcam is better positioned and it allows for more comfortable posture. meine-studios.com/products/ergo-…

@Austen Pretty great value for top notch talent

@tylermoore 25%, min 200k salary

Important to note for Gauntlet AI: Students never pay us anything. No matter what. No percentage of salary, no repayment of any kind. Even if you don’t take a job. We charge recruiting fees to employers. It’s a 100x better model in every way. 100% free for students.

I wish I knew in my early-mid 20s how important baseline labs were for your future. So much lab work analysis is relative vs absolute (e.g., numerical trends, not just a single score against permissible range). If you're in your early-mid 20s, highly recommend getting fully comprehensive labs at least once a year to have a baseline for the future.

I’d venture to guess the 75th percentile and higher of Claude users would pay 1-2k per month without batting an eye. 90th percentile and higher, likely 5-10k/month I also suspect at some point people will need to start better reconciling what they use bleeding edge models for vs what they can use cheaper models for, but won’t until price increases

AI about to get 20x expensive. These $200 / month claude subscriptions are burning $5,000 worth of credits. The bubble is going to pop and it will pop soon.

@bcherny @trq212 can we please get support for private GitHub repos as Claude Plugin Marketplaces on our personal accounts for Claude Desktop/Cowork? This is super clunky outside of Teams/Enterprise accounts and all of the connectors should be there already.

@levie Does it include write/edit support for Box Notes yet? 👀

The official Box CLI is here. Now you can use Box via Claude Code, Codex, Perplexity Computer, OpenClaw & more as a full cloud file system for agents. Available to all users, including free users with 10GB of free storage. npm install --global @box/cli