Here are the links and tools that caught my attention in June.
I saved about 70 links last month, and when I laid them all out, one cluster dwarfed the rest. A few months ago the interesting question was how to write a better prompt. In June, a dozen of my saves were about building the thing that writes the prompts for you.
Nobody prompts anymore
Peter Steinberger compressed the whole idea into one tweet: “Tell codex to maintain your repos, wake up every 5 minutes and direct work to threads.” He layers an orchestrator skill on top of triage, auto-review, and computer-use skills, so some of the work lands without him touching the keyboard.
Addy Osmani’s Loop Engineering is the piece to read if you want the full map. He collects the strong versions of the claim, including Boris Cherny, who runs Claude Code at Anthropic: “I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.” Addy’s anatomy of a loop has five parts: scheduled automations, worktrees, skills, plugins, and sub-agents, plus one memory that lives outside any single conversation. A markdown file counts. He also says plainly that this is early, that he’s skeptical, and that token costs can get away from you fast. The hedging is why I trust the rest of it.
The loops people are actually running got specific this month. Tom Osman fed Codex a /goal that inventories every feature in an app, writes a user story for each based on the code, tracks them all in one canonical spreadsheet, then switches itself over to testing the stories. Matt Pocock is building /loop-me, a skill that “interviews you about your work and finds opportunities for delegating your day-to-day work to AI.” And Katie Parrott built an idea farm for her writing: Codex watches her Slack, her workspace, and her agent conversations, scores potential pieces against her signature moves, and develops the good ones into pitches. Loops work on prose the same way they work on code, which I am professionally obligated to find interesting.
For sheer volume there was also MIKE’s guide to creating loops with Claude (1.8 million impressions) and Matt Van Horn’s Every Agentic Engineering Hack I Know, which roughly ten thousand people bookmarked, me included. Bookmarking an article about doing the work is much easier than doing the work.
If you want to try this without reading all of it: take Addy’s five pieces, cross off the ones your tools already ship with, and wire up the smallest loop that closes. One schedule, one repo, one job, one file where the agent writes down what it did.
Intelligence is borrowed
A frontier model got pulled, and shadcn’s reaction was the most clear-eyed take I saw all month: “A frontier model got pulled. If it happened once, it’s gonna happen again. Fable today. 4.9 tomorrow or maybe gpt 6 one day. So, treat intelligence as borrowed. Drain intelligence when it’s available.”
A surprising number of my June saves already treat models as interchangeable parts. fusion-fable is a Claude Code skill that fans a hard question out to a panel of models answering independently, then has Opus 4.8 judge the answers into a structured analysis and write the final one. The README’s case for the design: “even two cold runs of the same model diverge enough that synthesizing them beats running it once.” OpenRouter now ships the same shape as a server tool, so any model can convene a panel when it decides a question deserves one.
architect-loop splits the roles across vendors instead: Claude Fable 5 plans and reviews, GPT-5.5 Codex implements, and the repo is the shared memory. It runs on the subscriptions you already pay for (API keys optional). My favorite detail is that the builders are required to argue with the spec before they build; the README classifies silent compliance as a defect. JUMPERZ’s writeup of the same pattern puts it plainly: “fable thinks… codex builds, the repo remembers and you judge, that simple.”
Cross-vendor setups used to be a novelty you rigged up yourself. In June they came with install scripts.
Trust, but audit
If agents are producing more of the work, the tools I want are the ones that check it.
Alibaba open-sourced open-code-review, the code review agent they ran internally for two years across tens of thousands of developers. The design choice worth stealing: it deliberately trades recall for precision, and it burns about a ninth of the tokens of a general-purpose agent doing the same review. A reviewer that flags less and is right more often is a reviewer you keep listening to. That goes for humans too.
dupehound hunts a specific failure mode: the function your AI wrote twice. It fingerprints the structure of code instead of the text, so renaming every identifier doesn’t hide the copy, and it reports a repo-level “slop score,” the percentage of lines you could delete if every duplicate cluster kept one copy. The CI mode fails any change that duplicates existing code and names the original to reuse. It runs offline with no model involved, which felt almost contrarian this year.
coop handles the other direction of trust, running your agent full speed inside a container that mounts only the current repo, shadows your secrets, and keeps it away from your SSH keys. The README promises a box the agent can’t escape, “with a whip that will make them work all night.” One install drives Claude, Codex, and Gemini.
And Anthropic published agent identity, the access model where Claude stops borrowing your permissions and gets its own accounts: it posts in Slack as the Claude app, opens PRs as the Claude GitHub App, and hits your warehouse through an admin-scoped service account. The stat tucked into the post is that the length of task an agent can reliably finish on its own has been doubling roughly every four months. Ashwin Gopinath posted the counterweight the same day: “Claude Tag is a Trojan horse. Not because Anthropic is doing anything evil. Because the incentives are obvious.” Both things can be true, which is exactly why the admin-scoping part matters.
Getting recommended by machines
PostHog’s How to grow your AEO function is the best thing I read on AEO in June, mostly because it comes with receipts: their LLM-referred traffic grew 41x in 23 months, 7.4x year over year, and it converts “better than almost anything else we have.” The advice underneath is refreshingly boring. Most of traditional SEO still works, and the job is “answering the damn question instead of burying it under 400 words of ‘in today’s fast-paced digital landscape…’”. Their sharpest line: “if you’re not in the conversation, you’re not in the game.”
The tooling is arriving fast. CrowdReply launched an MCP that “analyzes and ranks your website in AI search,” then implements the fixes itself. Distribb’s Agentic Mode pitches agent owners directly: “Your AI agent can write. It can’t rank. Yet.” You bring the model, they bring keyword data, a backlink exchange of 500-plus real businesses, and CMS publishing. On the paid side, Pipeboard’s meta-ads-mcp lets an assistant run Meta ads end to end, one node in a family of five ad-platform MCP servers with 230-plus tools between them.
I spend enough time on paid-media math at work that I went down the AI marketing platform rabbit hole too, seven saves in one day: NoimosAI, Omneky, and half a dozen more. The detail that stopped me was on Omneky’s site itself: they publish an llms.txt, an agent-skills file, and an MCP server card at their own domain. A marketing platform marketing itself to your agent. These companies expect the first visitor to be software, and increasingly they’re right.
If your growth model still assumes Google sends you humans, the PostHog post is the one to forward to your team.
The SOUL.md discourse
Hermes Agent, Nous Research’s agent, has reached the ecosystem stage where the documentation has a fan site. The Hermes Bible indexes all 169 pages of the official docs plus 37 community workflows, 15 videos, and 26 repos, searchable from one keystroke. Somebody built a bible for somebody else’s agent.
The month’s mini-drama was SOUL.md, the file that defines a Hermes agent’s identity and voice. A Reddit post shared the SOUL.md template behind one dev’s agent, Prajwal Tomar amplified it in fluent engagement-bait (“WAIT. This is actually insane. A senior dev dropped the SOUL.md template behind his Hermes Agent. Says he’s never shared this before.”), and nearly four thousand people bookmarked the amplification. Tony Simons had already written the counterpoint weeks earlier, an article titled “I’m Not Sharing My SOUL.md. I’m Sharing Something More Useful,” and it’s the correct instinct. A personality file tuned to someone else’s taste is a horoscope. The thing worth sharing is how you arrived at yours.
I also saved MIKE’s How to Become a Hermes Agent Operator, because apparently we are all operators now.
Stating the obvious
Jim Nielsen wrote Blogging Can Just Be Stating The Obvious, and the HN thread about it is worth the click too. His key ingredient for blogging: “have a willingness to state something that seems obvious to you but nobody else is saying it. Or if someone else is saying it, just link to them and say, ‘Yes!!! This!!!’” He builds it on John Gruber’s rant about popups, which contains a sentence I want on a plaque: “A webpage should show the webpage. An email should show the email. I should not have to explain this.”
I write a blog where I regularly state the obvious, so: yes. This.
An Ask HN thread on text-heavy blogs that are a joy to read sent me to Butterick’s Practical Typography, a full book on typography that lives free on the web and reads beautifully, which is the whole argument made as an artifact. In the same neighborhood, Jay Alto asked X for “the greatest blog post on the internet,” the kind that will “violently shift your entire perception of reality.” A question worth asking somewhere once a year.
One more that belongs here: lathe generates hands-on, multi-part technical tutorials on demand, then makes you work through them yourself, by hand, in a local UI. The README calls it “an experiment in using LLMs to teach you, rather than think for you.” It rhymes with Claire Vo’s Earn your content from last month’s list. Generating the tutorial is not the same as doing the tutorial, and lathe is honest about which half is yours.
Stuff I bookmarked
- claude-lamp: Claude Code hooks drive a Moonside lamp over Bluetooth. Purple means Claude needs your input. A build light for your agent, and the most charming thing I saved all month.
- TRMNL’s BYOD API and the Inkplate docs: in April I saved an e-paper board with no project in mind. In June I saved the actual documentation. TRMNL’s docs open by admitting a scratch build “is not an economically rational decision, but rather a labor of love.” Still no project. The research phase continues.
- Aside: a browser rebuilt for agents that works across your logged-in sites directly instead of through integrations, and claims the top spot on three browser-agent benchmarks.
- Jellypod: AI podcasts with one to four hosts, voice clones included, published straight to Spotify and Apple. Saved the same day as ElevenLabs, which probably means something.
- Make Interfaces Feel Better: Jakub Krehel’s agent skill distilled from his article on interface details. Optical alignment, concentric border radius, tabular numbers.
- Pure headless vs. hybrid headless CMS: we are still working through a website migration, so this comparison landed at the right time.
- vivek’s “how to be good at research”: an X article with 30,000 bookmarks, saved to find out what 30,000 people know that I don’t.
- Alex Lieberman’s five levels of work: a ladder for agency that starts at “There is a problem” and climbs toward showing up with the fix. He says he uses it with every employee.
- Marc Lou’s “32 Principles of a Viral Product”: I saved this on June 13 and again on June 29, which says something about either the principles or my memory.
- “How to Feed a Dictator”: Witold Szabłowski tracked down the personal chefs of Saddam Hussein, Idi Amin, Enver Hoxha, Fidel Castro, and Pol Pot. Nothing to do with AI, which by June was itself a selling point.