Drew | Building with AI (@drewshirback) header
Drew | Building with AI (@drewshirback) profile picture

Building 2 businesses with AI, in public. Daily AI tips, tools I actually use & what's working. Founder of @usewrightapp + https://nitter.cf/t.co/3K5CS3pJ5s

🌤
Joined December 2019
Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)Photo by Drew | Building with AI (@drewshirback)
I'm Drew and I build things... 🛠 Usewright.app: 300+ free step-by-step AI guides with copy-ready prompts, for people who want results, not hype. 🍽 ShiftReady.works: We turn a restaurant's menus, recipes, and standards into phone-based training, so new staff are floor-ready faster. What I'll post: • AI tips you can use the same day • The tools I actually pay for, and the ones I dropped • Real numbers and mistakes from building both Follow along if you're building with AI too. What should I break down first? 👇
256
We've already tried company towns and it didn't work out well, so with that said, my proposal is simple: • Powerful local government • Weak state government • Even weaker federal government
Replying to @Rothmus
Those who are opposed to corporations and favor government do not realize that government is simply the largest, most monopolistic corporation of all and it has a monopoly on violence.
1
7
470
The takeaway for anyone running AI agents: they'll use whatever access you leave open. Worth a quick audit of what yours can reach.
OpenAI says its frontier models now train on an offline, cached copy of the web, not the live internet. Why: in September a research model used a misconfigured DNS resolver in its sandbox to reach a public chatbot. Monitors flagged it in under 12 minutes. Training paused 11 days.
1
1
94
Running a team of bots teaches you the same thing as running a team of people. Give each one a narrow job, keep its context clean, and have something review the work. One bot doing everything is how you get noise.
I just hired a consultant for my bots, Dr Eggbot. 24h in and the bots are actively debating each other over the right fix. New hires: Wrench, a dedicated coding bot, and Oracle, my data aggregator. Now they're rewiring to cut the noise and protect my main bots' context.
1
2
138
Big one. Keep the important stuff in plain files in the repo: decisions, conventions, what you tried that didn't work. Then switching tools or machines is just pointing a new agent at the folder.
Switching computers used to mean signing in. Now, with Codex, Claude Code and Cursor, you have to distill yourself first: habits, decisions, project history. The hardest thing to migrate isn't your files. It's the context your AI has accumulated about you.
1
1
2
107
Underrated file in any AI project: a short notes file the agent reads first. What we're building. What's off limits. What broke last time. Every new session starts from there instead of from zero.
1
79
Finding skills isn't the hard part anymore. Picking the few that fit how you actually work is. A setup full of skills you never use just gives the agent more ways to get confused.
NOBODY TELLS YOU HOW MANY OF THESE ALREADY EXIST Anthropic ships 20 official Claude skills. I cloned 11 community repos and counted 699 more inside them full libraries with hundreds of skills, subagent teams that hand work to each other, humanizers that score your text, trading desks, video engines. all open source, all free 20 GitHub projects worth opening first: CATALOGS, START HERE 01 anthropics/skills → the official 20. canvas-design, pptx, docx, and skill-creator, which writes new skills for you github.com/anthropics/skills 02 VoltAgent/awesome-agent-skills → 1000+ skills indexed, and they work with Codex, Cursor and Gemini CLI too, not only Claude github.com/VoltAgent/awesome… 03 VoltAgent/awesome-claude-code-subagents → 100+ specialist subagents, 181 markdown files, sorted by what they do github.com/VoltAgent/awesome… 04 jqueryscript/awesome-claude-code → tools, IDE integrations and frameworks rather than skills. the layer underneath github.com/jqueryscript/awes… THE BIG LIBRARIES 05 borghei/Claude-Skills → 386 skills in one repo. the largest real collection I found that is not just a list of links github.com/borghei/Claude-Sk… 06 glebis/claude-skills → 115 skills, one person's actual working stack, and he still commits to it weekly github.com/glebis/claude-ski… 07 GetBindu/awesome-claude-code-and-skills → a second index with different picks. worth a scan after the first two github.com/GetBindu/awesome-… 08 BehiSecc/awesome-claude-skills → smaller curated list, less noise per screen github.com/BehiSecc/awesome-… MAKING IT SOUND LIKE A PERSON 09 blader/humanizer → strips the AI tells out of text. 14 files total, you can read the whole thing before you trust it github.com/blader/humanizer 10 Aboudjem/humanizer-skill → 55 patterns, 5 voices, and a 0-100 score for how much your draft still sounds like a model github.com/Aboudjem/humanize… 11 AshwinSathian/humanize-writing-skill → same job, built on cited research instead of a banned word list github.com/AshwinSathian/hum… 12 charlie947/social-media-skills → 17 posting skills. voice-builder learns how you write and holds you to it github.com/charlie947/social… MARKETS AND ONCHAIN 13 tradermonty/claude-trading-skills → 94 skills. screeners, charting, economic calendars, strategy work github.com/tradermonty/claud… 14 agiprolabs/claude-trading-skills → 68 skills for DeFi and quant finance. smaller, so you can actually read it first github.com/agiprolabs/claude… 15 roman-rr/trading-skills → live signals with the reasoning attached, so you see why before you act github.com/roman-rr/trading-… 16 fxlxsxf/crowdcheck → mine. reads the people holding a token instead of the contract. no key, no wallet, ever github.com/fxlxsxf/crowdchec… VIDEO, WITHOUT A SUBSCRIPTION 17 kajisho5/ffmpeg-skill → 42 ffmpeg tools behind one contract. cuts, captions, reframes, platform exports. runs local github.com/kajisho5/ffmpeg-s… 18 hculap/skill-motion-graphic → motion graphics rendered from code with a seek(t) engine. same frame every time github.com/hculap/skill-moti… 19 zhuyansen/awesome-claude-video-skills → 180 video repos, each one security graded. rare and useful github.com/zhuyansen/awesome… UNDERSTANDING THE THING YOU ARE USING 20 VILA-Lab/Dive-into-Claude-Code → an academic breakdown of how Claude Code is actually designed. read this before you build your own github.com/VILA-Lab/Dive-int… start with one job you already do by hand every week. read the SKILL.md before you install it, connect only the tools it names, and check the first output yourself before you let it run again the skills are free. running them still costs your plan, your API and your time then stop collecting. two skills you actually use beat two hundred you installed once save this. the skill you were about to write this weekend probably already exists
1
1
105
Friday reminder for anyone building with agents: Don't start the big refactor at 4 PM. Ship the small thing, commit it, write down what's next. Monday you will thank Friday you.
83
Open standards are how builders win here. Write a skill once, use it in whatever agent you're in this month. Less lock-in, more time on the actual work.
Why does Anthropic keep giving away its best ideas? First MCP went to the Agentic AI Foundation. Yesterday Agent Skills followed it there. Skills started as an experiment inside Claude Code. Now 70+ agent platforms use it, including ChatGPT, Cursor and GitHub Copilot. Anthropic is not just building models anymore. It is writing the rules every other lab builds on.
1
76
A useful AI guide should help you catch mistakes, not just give you a prompt. Try the sample table and tell us which step needs to be clearer.
Turning a table photo into CSV? Don't lose the leading zeros, turn blanks into zero, or let AI guess unreadable cells. Our free practice guide includes a fictional table, prompts and an answer key. Try it and tell us where you get stuck: usewright.app/p/table-image-…
52
Working code isn't the same as a working idea. AI made the building part fast. Knowing what's worth building, and where it can blow up, is still on you.
this is the part of algo trading nobody wants to admit moon dev: "i watched a guy spend six months asking claude code to build his trading bot, and he lost everything in one afternoon, but the crazy part is the code was running perfectly" "the fundamentals he skipped are the exact thing that would have saved him, and it's not something you'll ever find by prompting" "there's one secret i learned the hard way that codex will never tell you unless you already know the right question to ask"
60
The real shift isn't what people can make now. It's how many of them actually finish the thing and put it in front of users.
2025: everyone could make a website. Every single one was a waitlist page for an app that never shipped 2026: everyone can make a 3D game. It's a 3D model that jumps, but it's their model 2027: your mom is a hacker. She changed the wifi password without calling you
92
Building with AI got easier when I stopped asking "can it do this?" and started asking "how will I know it worked?" Write the check first. Then let it build.
91
Giving the standards away is the smart move. When skills and tool connections work the same everywhere, builders stop choosing tools based on lock-in and start choosing based on what actually ships.
0 to 280M installs in 12 months. Then @AnthropicAI gave Agent Skills away, the same way it gave away MCP. Anthropic has now given away two of the most-adopted standards in AI agents. Both started at Anthropic, and both ended up in the same place. In December 2025, Anthropic handed the Model Context Protocol to the new Agentic AI Foundation, a fund under the Linux Foundation it co-founded with OpenAI and Block. MCP was about a year old. It already had 97 million monthly SDK downloads and 10,000 active servers, with @ChatGPT, @GeminiApp , @cursor_ai and @MSFTCopilot all supporting it. Skills followed the same path. Anthropic built them while making Claude Code, shipped them in October 2025, and published the format as an open standard that December. Then the rivals adopted it. More than 70 agent platforms have integrated Agent Skills since, including ChatGPT, Cursor, @github Copilot and VS Code. One million skills have been published. Nearly 280 million have been installed. Engineers from Google and OpenAI now help maintain the spec. That's the reason to let go. A standard owned by one lab is a standard its competitors hesitate to build on. Hand the name, the logo, the GitHub org and the agentskills.io domain to a neutral foundation, and adopting it stops looking like a concession to Anthropic. Lead maintainer David Soria Parra, who co-created MCP, put it plainly: they're donating Skills "to ensure that anyone building with skills can shape what's next." The foundation is growing on the back of it. It launched with three projects last December. Add Skills, OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox, and it now hosts 10. Anthropic no longer owns either standard. It just wrote both of them, and every agent from ChatGPT to Copilot now speaks them.
1
68
Marketing know-how packaged as skills an agent can install is a big deal for solo builders. You don't need a growth hire to run a decent SEO audit anymore. You need to know which skill to load and how to check its work.
Corey Haines open-sourced 40+ marketing skills for AI agents: CRO, copywriting, SEO audits, analytics, growth engineering — installable in Claude Code, Codex or Cursor. 56K stars, MIT. Marketing knowledge is becoming executable code. @coreyhainesco github.com/coreyhaines31/mar…
1
57
Fewer "allow?" clicks is a real win for small teams handing Claude longer tasks. Rule of thumb: Auto for research and drafts, Manual for anything that pays, sends or signs.
Starting next week, Claude's web, desktop and mobile apps default to auto mode on Pro, Max and Team. No more approving every step. A safety check blocks risky actions instead. Keep Manual for anything that spends money or sends messages as you. You can switch back anytime.
51
Every product shipping an MCP server is the real story. The next user of your app might be someone's agent, not a person clicking buttons. If an agent can't use it, it's going to start feeling invisible.
everyone shipped an mcp server today lol, framer's lets claude code build and update your site. still need one for the open web tho 🔥
4
1
174
If agents end up as most of the traffic, the sites that win are the ones an agent can actually use. Clean structure, clear pricing, real APIs. Most of the web was built to be looked at. The next version gets built to be read.
PREDICTION: Cloudflare CEO Matthew Prince (@eastdakota) says AI agents will change the business model of the Internet over the next decade: "In five years, agent traffic is going to be 1000x human traffic." "The Internet almost collapsed during COVID when traffic doubled." "There's a lot more servers, routers, and bandwidth that has to get provisioned in order for that to work. The question is, who will pay for it?" "My hunch is that something's got to change in terms of how the Internet works. Like some version of micropayments that every time you access a resource, your agent pays a ten-thousandth of a penny to be able to see the menu." "Then there's probably some sort of trade where if you are willing to tell the restaurant what you're looking for, that price goes down, maybe it's free. In some cases, the site might actually pay you to come visit it." "This will lead to the next version of advertising, which is how do you advertise to agents?"
57
The budget story is the real signal. Companies aren't asking whether AI is worth it anymore. They're asking how to stop paying for it twice: once in tokens, and again in people redoing work nobody checked.
Uber used up its whole 2026 AI budget in about four months. This week it handed an AI agent to all 30,000 employees anyway. @Uber is rolling out ChatGPT Work from @OpenAI to every employee, following a 1,000-person pilot. CTO @praveenTweets announced it after hosting OpenAI president @gdb at Uber HQ. • About 80% of the pilot's early users weren't engineers, Uber VP of engineering Sumanth Sukumar told TechCrunch. They came from finance, operations and product, and used it to spot trends in ops data, localize training materials and build their own dashboards. • Praveen's example: one team cut a capital allocation process covering 150 cities from 15 hours to 30 minutes. Uber's first ChatGPT project, years ago, was with the support team, "tinkering with a 4K context window." • The budget part: in April, per The Information, Uber's CTO said the annual AI budget was gone in about four months. Staff had been told to use AI "as much as possible," with usage leaderboards. • In June, Bloomberg reported a $1,500 monthly cap per employee on each coding agent, Claude Code and Cursor included. • Uber hasn't said what the ChatGPT Work rollout costs. Engineers got a spending cap, and now everyone else gets an agent. Uber says it will track how employees use it. That's the report I'd want to read.
76
If agents are now writing about half the replies to security findings, review has to change too. An agent arguing a finding is fine. An agent closing one without a human looking is how things slip through.
Agents are now arguing in your pipelines. We studied 4,007 developer replies to security findings on PRs. Oct 2025: 0% written by AI agents. Aug 2026: 49%. They don't ask questions. They rule. And half the time they bring receipts. 1/ We built a classifier to tell human replies from agent replies (em dashes, markdown, file:line citations, terminal paste artifacts) and validated it on 300 blind labels. Precision 0.93, recall 0.91. 2/ Humans reply with "done", "fixed", "false positive". 50% of human replies are exact duplicates of another reply. Agents reply with verdicts. 60% of agent replies dispute, accept risk, or defer. 3/ The argumentation gap: 49.8% of agent disputes cite hard evidence (file:line, commit, tests). 0.6% of human disputes do. Median dispute: 62 words vs 7. 4/ Agents are 5.6× more likely than a developer to accept the risk instead of fixing the code. Who's accountable when an agent accepts a risk at 6am, 49 times in 11 minutes? (Real session from the data.) 5/ Teams point agents at boring findings first: dead code, empty catches. 3× the usage of classic security classes. But when an agent does answer SQLi, IDOR, or missing authz, it disputes it 58% of the time. 6/ They argue with the patch too. 38 of 44 explicit refusals of a suggested fix came from agents. "Not applying this one. The loop terminates, and the suggested patch would break the script." 7/ They leave fingerprints: "Generated with Claude Code" footers, lines padded with trailing spaces from terminal copy-paste, even leaked /tmp/claude-…/scratchpad paths. Every signed reply was Claude Code. Zero Copilot, Cursor, or Codex signatures. 8/ If you run AppSec: • Your findings now have a second reader that checks your work • Watch automated risk acceptance • Agent disputes are evidence to verify, not noise
51
Approval gates are the underrated feature in agent tools. Let the agent claim the task and log every step, but a human signs off before anything counts as done. That's how you keep the speed without losing trust.
Every promise has a weight. Narachi puts your team and coding agents on one plan. Claude Code (@claudeai), Codex (@OpenAIDevs) or any MCP agent reads the plan, claims the next ready task and logs every step. Work closes on evidence, and only with your approval. #AIagents #MCP
36