156: Batten Down The Hatches

Ian explains that he spent a Saturday building a personal knowledge system after becoming frustrated with Obsidian’s lack of convenient AI integration. He used Amp “orbs” to turn a large specification into a working application in roughly three hours. The product was a text area that classified links, notes, questions, and lightweight to-dos, then stored and retrieved the material with AI. It resembled a deliberately limited “second brain,” not a complete life operating system. The prototype technically worked, but Ian quickly abandoned it because making it fast, polished, mobile-friendly, secure, and usable by others would require much more work.

Ian says he had possible names including Blackhole.app and jokingly considered “Oblivion” for the knowledge base. Aaron asks why he did not keep building a private collection of small tools rather than a public product, since a personal system would have a smaller attack surface. Ian agrees that security-by-obscurity would help and says he was not planning to store especially sensitive material, but maintenance still bothered him: he would own another application, another upgrade path, and another set of rough edges. The decisive problem was refinement, especially creating a mobile experience he would actually enjoy using every day.

Instead of finishing Blackhole.app, Ian connected Claude to Google Drive. Claude creates Google Docs for the stored material and maintains an index in a Google Sheet; a few skills tell it what to do with a pasted link, a request to log something, or a question. Ian likes that the data remains accessible through Google’s own applications, backed up, and available if he changes AI providers. He expects only about 100 to 300 items per year, so scale should not be an immediate issue. Aaron initially questions the choice but concedes that Google Drive is universal, secure, and already supported by multiple AI connectors.

Aaron suggests forcing the documents to use Markdown rather than rich text, because an agent should not have to construct or interpret elaborate Google Docs formatting. He also approves the spreadsheet as a structured index that could be queried or extended with Google Apps Script. Ian says the system uses Firecrawl through an MCP connection to fetch pages that ordinary tools cannot access, then stores the extracted material and keywords in Drive. Aaron recommends centralizing that crawling process in the Drive workflow so switching from Claude to another assistant will not require rebuilding the integration. Ian replies that moving the current few skills would still be easy, though he admits the system will probably grow.

The knowledge-system project also forced Ian to address his overloaded inbox. He gave Claude permission to read and relabel email, but not send messages; removing Gmail’s Inbox label effectively archives a message. An hourly Claude process created too many sessions and notifications, so Ian moved the workflow to Claude’s Duet offering, which can reuse a session and suppress its conversation updates, though he had to configure it not to write anything in order to eliminate push notifications. The agent now archives or labels newsletters and other messages automatically. Ian says his email is noticeably better, while Aaron notes that connecting an assistant to accumulated email and documents makes phone-based searches and spoken requests genuinely useful.

Aaron says he remains skeptical of agents that merely book flights or dinner reservations, but sees real value in building a personal corpus of tools, documents, and workflows that whichever assistant is convenient can access. Ian agrees that the ecosystem is extremely early: a future product could package the whole experience, or people may keep extending Claude or ChatGPT indefinitely. Their preferred strategy is to own the underlying data and useful processes while making the front door replaceable. Ian specifically does not want to become the developer trapped inside disposable AI-wrapper products; the durable asset should be the user’s knowledge, tools, and integrations.

Aaron reports that Laravel’s marketing team held its offsite in Fort Worth. Fourteen or so employees had individual hotel rooms, while a two-story penthouse served as their shared office, with a kitchen, living room, large table, couches, two bathrooms, and a television connected with Aaron’s HDMI cable. The arrangement avoided expensive conference-room equipment and created a comfortable base for concentrated work. They arrived Monday night, worked Tuesday and Wednesday, and left Thursday morning. Lunches were short walks to Corner Bakery, while dinners were at Joe T. Garcia’s for Tex-Mex and Terry Black’s for barbecue.

The offsite combined an intense launch-oriented work schedule with easy socializing. After the first dinner, some employees played Settlers of Catan until late; on the second night, the full group returned from ice cream and played five to seven rounds of Mafia, keeping even the people who normally go to bed early engaged. Aaron calls it the best company retreat he has attended. The team also ran AI sessions: Haley taught people to use Amp orbs, Aaron demonstrated Solo for people still using local development, and Steve showed his creative process using Claude CLI inside Ghostty without additional tools. Aaron managed from the screen while teammates orb’ed tasks and opened pull requests.

Aaron says the team completed a large list of launch tasks by assigning pairs and groups from a shared screen while everyone worked in orbs. The setup gave nontraditional developers enough guardrails to make changes without installing and troubleshooting a local environment. It also let teammates see one another’s threads and preserve the reasoning behind decisions. Ian points out that Amp supports switching from AI interaction to coworker conversation inside the same thread, including at-mentions, so a teammate can ask why something is being done and then return to building. Both see this as a promising model for collaborative, nontechnical work around code and other projects.

Ian visited Philadelphia for a PHP meetup connected to Boss Man Chris. The event used a reverse Shark Tank format: contestants pitched deliberately funny ideas that were not expected to make money, and Ian judged alongside Alex Hillman and Andrew. Mary Perry won with Eff.fm, an app teaching users about English swear words; Ian suggested expanding it to other languages. His brother joined the trip and met the group, and they ate 1-900 Ice Cream. Ian also played poker with his brother, cousin, and Daniel Colborne, eventually moving from an unusually silent $1/$3 game to livelier $2/$5 and $5/$10 tables.

Ian describes trying to revive the dead $1/$3 poker table by playing every hand for an hour and announcing that he would play any hand containing a four. The table still would not talk or laugh, so the group left. At the later games, Ian won about $5,000, bringing him roughly back to even after a downswing. During an 11 a.m. to 4 p.m. session, they ordered Angelo’s cheesesteaks through Uber Eats rather than leave the table. Ian was surprised that the sandwiches traveled well: the crusty bread was wrapped in breathable paper and did not become soggy. He calls Philadelphia an increasingly strong East Coast poker destination.

Aaron cites Tyler King of Less Annoying CRM, whose podcast reported that the company had been flat and was beginning to see a downturn, with some customers saying they would build CRM functionality internally. Aaron argues that this expands the “AI is eating SaaS” story beyond indie hackers trying to save a few dollars. Ian says founder-led companies with five or ten employees may be especially willing to customize, but he distinguishes them from larger organizations. Replacing a $500-per-month product is not enough: meetings, implementation, maintenance, security, and staff time can quickly cost tens of thousands of dollars, so the custom system must create substantially more value.

Aaron pushes back that companies are groups of people, not perfectly rational calculators. A manager may value a tailored workflow, control, or convenience rather than merely subscription savings, and smaller businesses may simply build a tool without a formal IT process. Ian maintains that larger organizations face budgets, approval chains, boards, and opportunity costs, making self-built software harder to justify. They agree that AI makes it cheaper to combine the useful 10 percent of several products with company-specific features, but disagree about how widespread that behavior will be. Ian says he has never worked somewhere that casually replaces established systems; Aaron believes many smaller companies will do exactly that.

Aaron announces that he is moving personal projects from Forge, and some from Vapor, onto Laravel Cloud because he no longer wants to be responsible for servers, operating-system updates, or monitoring CVEs. He says even a capable engineer cannot reliably defend every application and would rather have Laravel’s team carry the operational burden. Ian agrees that the threat environment is worsening: automated scanners already probe Laravel and WordPress endpoints, and AI will eventually make it cheap to test every form and application for weaknesses continuously. He says Laravel Cloud needs to market and improve proactive security, including rapidly patching tens of thousands of pods without downtime.

The hosts mention recent vulnerabilities affecting Railway, Northflank, Trigger.dev, Ubuntu, Chrome, Playwright, and AI desktop tools, as well as a reported $6.3 billion Binance hack linked to North Korea. Ian says attackers currently appear especially interested in stealing environment keys and AI credentials, but ransomware could make the consequences much worse. Aaron argues that professional software engineering still matters because security, deployment, maintenance, and incident response are not solved merely by vibe-coding an application. Ian says Laravel Cloud recently stress-tested playbooks for patching running applications without forcing redeployments, but both expect the security situation to become more difficult rather than less.

The episode closes with Ian showing Aaron a substantially revised Outro.fm interface. Aaron prefers the faster, simpler workflow: a rich-text editor, a sidebar, keyboard movement between segments, automatic scrolling to the current position, and a mailbag page. The new canvas view displays segments as movable cards that can be grouped or stacked independently of the ordered list, allowing a producer to organize visually without changing the show rundown. Aaron asks Ian not to redesign it again, only to improve dark-mode contrast and ship it. Ian says he stayed up until 2 a.m. and used orbs to fix configuration and migration issues, with a public release possibly coming within the next week or two.