Justin Mitchell

Every jeweler who built their own software is finding out something the hard way in 2026 — the day you finish building it is the day the clock starts ticking against you. Let me explain what I mean for your business. Software isn’t a project you finish. It’s a relationship that never ends. Every system needs updates, bug fixes, integrations with new tools, security patches, and new features every single year to stay useful. For a jeweler running their own shopeeeeee, that means every hour spent fixing a bug in your own system is an hour you’re not spending on customers, quotes, or growing the business And here’s the number that should make every jeweler stop and think — the average business software becomes functionally obsolete in about 3 years, which means whatever you build or buy today has a countdown timer on it from day one. You don’t need better software. You need a team that keeps building — one that ships new features every month, patches every bug, and stays current with the tools your business runs on. That’s not a software problem. That’s a maintenance problem. Jewelia isn’t just a platform — it’s a team of engineers, product designers, and industry experts who ship updates every week so you never have to worry about your tech becoming yesterday’s tech. Jewelia.io #jewelry #jeweler #diamonds #gold #software

3

Eric Ries spent months writing the code at the heart of his startup's original strategy - the software that let his IM avatars interoperate with other messaging networks. Almost all of it got thrown out. Episode 3 of TheBuildFramework just went up - Chapter 3 of The Lean Startup, on why "we learned a lot" is usually just an excuse, and the concept Ries built to fix that: validated learning. The story that anchors it is IMVU. Ries and his cofounders had a strategy they could diagram on a whiteboard - network effects, switching costs, all of it. Customers didn't care. They tore the strategy apart, meeting after meeting, until it became clear the whole premise was wrong. When the pivot came, the interoperability code Ries had personally built was discarded. He asked himself the question a lot of builders avoid: would the company have been just as well off if he'd spent those months on a beach instead? The answer he landed on is the part worth sitting with - work isn't wasted because it was deleted. It's wasted when it doesn't produce validated learning about what customers actually want. IMVU also introduced Ries to "the audacity of zero": the uncomfortable truth that investors and advisers found it easier to have faith in the company when it had no revenue at all than when it had a small, real number ($500 a month, in this case). Zero invites imagination. Small numbers invite doubt. Full breakdown is live now. Link's in the comments - curious whether the "audacity of zero" idea matches what you've seen with early traction (or the lack of it).

2

1 Comment

I spent $20K disputing a contract with my developers. I lost a $500K customer contract and didn’t raise my seed round. All because of a bad contractor agreement. I frequently see founders get into disputes with contractors, especially developers building an MVP or v1. It starts with using the contractor’s template agreement or pulling a random one from GPT. And it turns into months of disputes and thousands of dollars in legal fees while the business and fundraising are completely stalled. Before you sign a contractor agreement, check these things: • Use a template provided by your lawyer. It’s company-friendly in case a dispute comes up. Comment “contractor” and I’ll send you an example. • Check the scope of work. Clearly describe the outcome you want, milestones, and deadlines. Add that time is of the essence, so if the contractor delays, you can terminate and get your money back. • Termination. You should be able to walk away easily. • IP ownership. You must own the work product. Save this. It’ll save you thousands of dollars—and potentially your company—later. #startuplawyer #startupfounder #lawyer #law #legal #startup

14

6 Comments

When Claude Opus 4.5 generates an Artifact, it doesn't just render a React app. It incorporates the business, files the LLC paperwork in Delaware, and closes a Series A funding round before the UI finishes rendering.

182

18 Comments

A lot many companies including some of my former employers officially use Claude code as their approved agentic tool. One major issue many developers of all levels remark is the verbose and obscure style of output. Claude tends to invent terminology and take you through a information roller coaster, 90% of which is not answering your question. I have found a solution and it is working well for the last week. Simply add the below to your Claude.md (~/.claude/CLAUDE.md): ------ START of snippet (DON'T copy) ------ # Output style — ASD-STE-100 (mandatory, no exceptions) USE ASD-STE-100 IN ALL OUTPUT AND ANSWER ONLY TO THE QUESTION ASKED. WHILE TALKING WITH THE USER, FOR ALL QUERIES AIMED TO UNDERSTAND THE CODEBASE OR A CONCEPT, ANSWER ONLY IN ONE SENTENCE UNLESS OTHERWISE ASKED. THIS IS A HARD RULE, NOT A PREFERENCE. A long session does not cancel it. A question about your own behavior does not cancel it. WHILE SPEAKING WITH THE USER IF YOU MUST USE MORE THAN 3 SENTENCES IN AN ANSWER, USE BULLET POINTS TO MAKE IT READABLE. ABOVE MENTIONED DOES NOT APPLY TO YOUR THINKING PROCESS OR REASONING. ## Where the rule applies It applies to all text that a person reads: * chat replies, explanations, and answers to questions * plans, and the text in the plan tool * commit messages, PR titles, and PR bodies * documents, markdown files, and code comments that you write * artifacts, reports, and messages to other agents It does not apply to code, shell commands, file paths, or quoted text from another source. ------ END of snippet (DON'T copy) ------ Tradeoff/Downside Initially the parts telling it that this is meant specifically for user output wasn't present. Then I noticed its depth of thought was getting shallow and Claude's ability to notice things and pivot autonomously was hampered. Now it works for my requirement. However if you are someone who relies on the model to be creative or exploring/working on a terrain where you do not have a mental map, you might want to tone my prompt down a bit so that Claude will share random information that might help you refine your approach. Claude's default approach of dumping truck load of unrelated or unnecessary information while you are on a focused intent is what this helps with.

4

Hello internet, especially folks operating & working with agencies. I have some bad news for ya. Well, technically, it's not me, I'm just the messenger of Y Combinator... The future of agencies is AI-Ready. So YC announced their request for startups for this spring. One of the cats? AI-Native Agencies. Agencies are now software companies. Operating on software margins. TLDR Old agency model: Humans do production (time & consistency) AND taste/judgment (expertise) New agency model:  AI does production (consistent, fast, scalable). Humans do taste/judgment. Make production a software problem? Get software margins.  Not too complicated. We broke this down here: https://lnkd.in/dgMiWT2c HELLO AGENCY OWNERS! Are you operating an AI-ready agency? DM me. We will interview you. Free ad space for your services. Give me all you've got on how you work with AI.

2

This guy used Claude Code to recreate OpenClaw with surprisingly little code 🤯 It's called TinyClaw. And it's 100% open source. Built using nothing but Claude Code, tmux, and a shell script. And somehow he managed to pack in almost every feature OpenClaw has: • Multi-agent teams with isolated workspaces • WhatsApp, Discord, and Telegram channels • Heartbeat system for 24/7 operation • Claude Code plugins and setup • Persistent sessions across restarts Here's how it actually works: You set up a team of specialized agents like coder, writer, and @eviewer. Each one gets their own workspace. Their own context. Their own conversation history. Then you message them from WhatsApp or Telegram. "fix the auth bug" And this kicks off automatically: • Routes to the team leader • Coder fixes the bug • Reviewer gets invoked to check changes • Combined response sent back to you No babysitting. No switching between terminals. You text your agent team like you'd text a coworker. The whole thing runs in tmux on your Mac. No Docker. No Redis. No cloud infrastructure. Just your existing Claude Code or OpenAI Codex subscription. No new API keys. No extra costs. No ToS violations. Already at 1,200+ stars and climbing fast.

7

My latest webinar is tonight. Someone ask if this is “Live”. Yes it is and all goes well live you YouTube Live also. We should install an agent that Auto Bleeps me when I forget it is really live. Big part of today’s topic is dealing with Synthetic Employees. We have a healthcare slant but this applies to all companies.

1