The AI-Native SDLC, Translated for a One-Person Company

Code was the slow part. Not anymore.

The AI-native SDLC is Anthropic's playbook for how software should get built now that agentic AI writes the code (Anthropic, The AI-Native SDLC Playbook). Its one central claim: the old process was designed for a world where implementation took weeks, and that world is gone. The build now finishes in hours. Everything around it, deciding, reviewing, approving, watching, still runs at human speed. That is the new bottleneck.

I feel this in my own work, every week. For example, this month my system drafted a rental app for me faster than I could decide what I actually wanted it to do, using Claude Code as the engine. The building was not the wall. My own clarity was.

And the gap is real at every size. McKinsey found most organisations now use AI somewhere, but only about 6% see real bottom-line impact from it (McKinsey, State of AI 2025). The missing piece is never a smarter model. It is the process wrapped around it.

Code used to be the expensive part. Now the expensive part is you. 以前贵的是代码，现在贵的是你的判断。

— Weiss Ang

The six stages, in plain English

Anthropic's version has six stages, each ending with a committed file the next stage reads. Here is each one, translated out of engineering language.

Notice what runs through all six. Every stage ends with a file somebody signed. The chain of files is the audit trail: who asked for what, what got built, who approved it. That is the whole trick.

What this looks like when the whole company is one person

Treat this section as a real-world case study of the playbook at its smallest possible scale. An enterprise runs this playbook with a product owner, a tech lead and a release manager. I run it with none of those people, because there are no other people. My company is me, a laptop in Sabah, and a system of 23 AI staff members, each with a written charter saying what it owns and what it must bring to me.

So the org chart becomes rules. Before any build starts, the ask goes into a brief in my own words. Before any build ends, it has to pass a written finish line, a list of checks a script can run, where the script's answer is the verdict, not my mood at 11pm. And the mistakes I have corrected twice live in a handbook file the AI reads at the start of every session, which is exactly the playbook's rule for institutional knowledge.

I did not adopt this because it is elegant. I adopted it because a one-person company has no colleague to catch you. The tokens are cheap; I once ran 6.19 billion of them in 45 days, verified on a public leaderboard (the receipt is here). Without written gates, that much output is just noise moving faster.

An enterprise writes process so a thousand people agree. I write process so one person can be absent. 大公司写流程是为了让一千个人同步，一人公司写流程是为了让那一个人可以缺席。

— Weiss Ang

The three files that do the heavy lifting

In practice, strip the playbook to what a solo owner actually needs and three files remain. Everything else is elaboration.

• The ask. One page, your own words: what you want, why, what would make you stop. Write it before the AI touches anything. Anthropic's version is intent.md. Mine is a build brief with ten boxes, and the boxes only I can fill, the goal and the stop condition, stay mine.
• The handbook. One page the AI reads every single session: how your business works, your conventions, your known traps. The playbook's rule is blunt and correct: when the AI makes the same mistake twice, the correction goes in the handbook. Teach once, never re-explain.
• The finish line. "Done" written as checks a machine can run. Not "make it good". Not "I'll know it when I see it". A list where each row passes or fails. My builds do not end when I am tired. They end when the list is green and I sign it.

First the files, then the speed. Run it the other way around and you get fast output nobody asked for, which is the most expensive thing AI produces.

What stays human

The playbook is direct about this: humans remain accountable for every decision that requires judgment. The AI prepares the release; a person authorizes it. The AI drafts the finding; a person decides it matters. Attention does not disappear. It moves to the files worth reading.

In practice that means four standing doors in my system, and they are always manual: anything that costs money, anything published under my name, anything deleted, and the first go on any new build. My AI can reach every one of those doors. It cannot walk through them. That single design choice is why I can let it run for hours without watching it.

Start this week, no code required

You do not need to ship software for this shape to pay you. It works for any real work you hand to AI, a proposal, a report, a content system. The one-person version:

• First, write the ask. Next time you want AI to build or draft anything, spend ten minutes writing one page in your own words before you open the tool. What, why, what does done look like.
• Second, start your handbook. One file about your business: services, prices, tone, the ten questions clients keep asking. Paste it into every session, and add a line each time the AI gets you wrong.
• Third, make done checkable. Before the work starts, write the three to five checks that would prove it is finished. Judge the output against the list, not against your patience.
• Finally, put your gates in writing. Decide now what always waits for your yes. Money, publishing, deleting is a solid floor. Everything inside the gates, let it run.

The enterprises get the playbook. You get the advantage: no committee to convince. One evening is enough to set up all three files. 慢的从来不是 AI，是没人写下来的决定。

FAQ

Sources

• Anthropic — The AI-Native SDLC Playbook (the source this field note translates)
• McKinsey — The State of AI 2025 (only ~6% see real bottom-line impact)
• Viberank — AI coding usage leaderboard (the verified 6.19B token record)