Shipping fast with AI

Shipping fast with AI means the first version of a small product is no longer gated on how much code one person can type. An agent drafts, the person steers, something goes online. In 2026 that is ordinary. It is not a new definition of indie hacking. The definition is still ownership and a customer. The editor changed. The base rate did not become kind.

This page is the state of that method. It is not a prompt pack. Anyone selling a prompt pack is on the course shelf.

What the names refer to

ChatGPT opened to the public on 30 November 2022. Within a couple of years, people who were not career engineers were launching small products, and people who were engineers were skipping boilerplate they used to write by hand. In February 2025 Andrej Karpathy named the loose end. Vibe coding: you describe the outcome, you accept a lot of generated code, you do not read every diff, you throw the project away if it is only a weekend toy. The name escaped the caveat. It now covers careful agent work and reckless agent work, which is inconvenient, because those have different failure modes.

Careful work still has a person who can tell that the charge flow double-bills, that the permission check is missing, that the agent “fixed” a bug by deleting the test. Reckless work finds out from a customer, or from a card network. Speed does not pick a side. The person does. Karpathy was explicit that the style, as he meant it then, was for disposable projects. A product that stores someone else’s data is not disposable. Treat the slang as slang.

What got cheaper

Setup. Auth, a database, a deploy, a landing page: the chores ShipFast sold as a repo in August 2023. Lou’s later note that AI hurt the boilerplate is the market saying the chore got cheap enough to stop buying. You can still buy a kit. You no longer have to if the only value was wiring.

Variation. A second product, a clone of your own idea with a different niche pasted in, a landing page test: the marginal experiment is an afternoon. Levels’ 2014 “twelve startups” year was a public stunt about volume. The volume is easier now. The hit rate was never the point of that year. Most of the twelve did not last. Faster volume does not fix the hit rate. It lets you reach the misses sooner, which is useful if you actually stop.

Vendor replacement, for someone who already runs a system. The September 2026 stack story is a year of agents on a Hetzner VPS, then a list of cancelled subscriptions, then a founder-reported savings near $25,000 a month. The scraper alone failed for weeks. He posted the give-up note and then the success note five days apart. That is the honest shape of “just vibe-code it.” It works until it does not, and the does-not can be the one vendor you cannot afford to lose. His hub keeps the cost story next to the revenue story so a savings claim is not a revenue claim.

What did not get cheaper

A customer. The start is still a buyer you can reach. Agents do not know your buyer. They will invent a persona if you ask, and the persona will not enter a card. Distribution is the scarce input the homepage has been saying out loud. Fast shipping increases the number of products asking for the same attention. It does not print attention.

Trust. A stranger has learned, sometimes the hard way, that a new tool might be a wrapper that breaks. Revenue proof, a name they have seen before, or a workflow they already live in matters more when the screen is easy to fake. TrustMRR is about fake numbers. The same mood applies to fake products. Shipping fast makes the mood worse for everyone who ships fast, including the careful ones.

A moat made of difficulty. Levels’ 30 July 2026 post is the citation: about $20 a month to build, his figure, and a customer who might build their own copy of a thin tool. The killing page stays with that claim. If your speed is the product, you are racing people with the same speed. Data, a network effect, a brand, a regulation, a workflow full of edge cases: those can still be slow in the right way. An agent does not absorb a decade of listings because you asked nicely.

Judgment about when to stop. The failure mode of cheap experiments is never stopping. Lou’s useful record is not “he shipped a lot.” It is that most of the lot did not become the company, and he kept going anyway, including through a period he describes as giving up. Volume without quitting the dead ones is just a busy repo. AI makes the busy repo cheaper. It does not install the quitting.

What fast shipping looks like when it is a business

Postiz did not win by being the fastest scheduler ever prompted. It already scheduled posts. The fast part was a surface agents could call: API, CLI, skill, MCP. The slow part was years of networks, auth, and failure modes. The product hub is that sequence. Speed on the margin, depth underneath. That is the pattern worth copying if you already have depth. It is not available to a weekend clone, which is why the weekend clone’s speed is less impressive than it feels.

A new product can still be fast and real. The test is boring. Does a named person have the problem this week? Did they pay? Does it work on the second Tuesday? Can you explain the bill if an agent wrote the billing code? If the answers are yes, the method did its job. If the answers are no, another afternoon of generation is procrastination with a modern interface.

Public builders who are good at this talk about what they refused to automate. Levels’ leftover list is domains, storage, machines, and models, in his posts, not a ban on every SaaS. A support desk he replaced is not a support desk you should delete if you have one human customer and no time to read the mail. Copy the criterion — would I rather own this failure? — not the cancellations.

The scene around the method

Courses that teach syntax are the collateral damage. CodeFast’s smaller 2026 month, in Lou’s recap, sits next to his sentence that AI hurt the course. The tool list on this site stays short because the vendors change and the method does not. Cursor, Claude Code, ChatGPT: names from the record, not a required kit. Next year’s names will differ. The $20 figure will date. The split will not. Building got cheap. Being believed, being found, and being worth more than a private replica did not.

Use the speed. Spend it on contact with a buyer and on the parts of the product that break in real use. The year is already full of people who spent it on a launch video instead. The launch video is also fast to generate. That is the tell.

FAQ