Something significant has happened to the way software gets built, and it is worth understanding even if you never write a line of code, because it directly changes what you pay and how long you wait when you commission a custom build.
Claude Code is an AI coding agent. A developer describes what needs to exist, in ordinary language, and the agent writes the code, runs it, reads its own error messages, fixes what it broke and comes back with working software for the developer to inspect. It is not autocomplete. It is closer to a tireless junior colleague who drafts entire features and never minds being asked to redo them.
What actually changed
Software has always had two kinds of work in it. There is judgement: understanding the problem, choosing the approach, deciding what good looks like. And there is translation: turning those decisions into thousands of lines of correct, consistent code. Translation has always consumed most of the hours, and hours are what you pay for.
AI coding agents collapse the translation work. What took a developer three days of careful typing now takes an afternoon of directing, reviewing and correcting. The judgement work remains fully human, but it was never the expensive part by volume. The result is that a small, senior team now ships what used to require a department.
What stays human
It would be dishonest to describe this as software writing itself. The tool is powerful and completely unsupervised use of it is how you get plausible-looking systems that fail in production. The human parts are exactly the parts you would hope:
- Scoping: deciding what to build, and just as importantly, what not to build.
- Architecture: choosing structures that will still make sense in two years.
- Review: reading what the agent produced and rejecting what is not right.
- Testing: proving the system behaves under the awkward, real-world cases.
- Sign-off: nothing ships because an AI said it was done.
A worked example
Take a job every service business recognises: quotes go out, some are never answered, and nobody has time to chase them properly. The fix is a small system that watches sent quotes, waits a sensible number of days, drafts a polite follow-up in your tone, and logs the outcome in the CRM.
A few years ago that build meant weeks of developer time: wiring into the CRM, handling the edge cases, building a small interface, testing everything by hand. Quoted honestly, it rarely made sense for a small business, so the leak stayed open and the quotes stayed unanswered.
Built with an AI coding agent, the same system is a few days of work. The developer describes each piece, reviews what comes back, tightens the awkward cases and tests it against real scenarios. The judgement involved is identical. The typing is what disappeared, and the typing was most of the bill.
Why this matters when you commission software
The economics of custom software have moved, decisively, in favour of the buyer. Builds that were quoted in months now land in weeks. Ideas that were not worth the old cost, like a small internal tool that saves one person one hour a day, now clear the bar comfortably. The floor for what is worth automating has dropped.
Speed changes the shape of projects too, not just the price. When a working version exists in days, you react to something real instead of imagining from a specification document. Feedback arrives early, misunderstandings surface while they are cheap to fix, and the finished system reflects your business rather than a guess about it.
How we use it at Singularity1
We build with Claude Code daily, inside a process that keeps the owner in charge. The scope is agreed in writing before anything is built. The agent accelerates the build, and a person reviews and tests everything it produces. Nothing goes live without the client seeing it work and signing off, and every system ships with plain-English documentation and alerts that speak up when something needs attention.
The tools changed the speed and the price. They did not change the standard. Tested, documented, owner-approved: that bar stays where it has always been, whoever or whatever wrote the code.
Questions worth asking any builder
None of this removes the need to choose builders carefully. If anything, faster tools make discipline more important, because sloppy work now ships sloppier and sooner. A few questions surface the difference quickly:
- Who reviews the code the AI produces, and what happens when it is wrong?
- How is the system tested before it touches live work?
- What documentation arrives with the build, and can a non-technical person read it?
- Who owns the finished system at the end: the code, the documentation and the data?
If you looked into custom software a few years ago and walked away because of the quote or the timeline, the honest advice is that the maths has changed. The repetitive work in your week that never justified a big build now justifies a small one, and small builds are what modern teams do quickly and well.