AI-DLC · the LEVELHARD method
The cycle we use to build software with AI, explained for people outside the field.
Six stations, an AI agent working in each one and a person approving every handoff. Nothing ships without proof. Here is the whole cycle, the apps that came out of it and where we took every idea from.
Writing code got cheap: AI already writes a good part of it. But writing code was always the smallest part of the job. What is still expensive is knowing what to build, checking that it works, putting it live and listening to the people who use it. A method that only speeds up the code leaves everything else the same size.
The code block shrank. The others are the same size, and almost nobody measures the last one.
Giving an industrial mixer to a bakery does not help if nobody knows which cake the customer wants, or tastes the batter before delivering it.
of new code at Google is already AI-generated and approved by engineers
Google 2026 (opens in a new tab)of a developer’s time goes to writing code
Atlassian 2025 (opens in a new tab)of technology professionals already use AI at work
DORA 2025 (opens in a new tab)slower, in a 2025 study, were developers who felt 20% faster
METR 2025 (opens in a new tab)AI amplifies what is already there, good and bad. And the feeling misleads: in 2026 the same study saw signs of a gain, still on weak evidence. Only what was measured counts. DORA 2025 (opens in a new tab)METR 2026 (opens in a new tab)
AI does.
An AI agent, a program that works on its own, acts at every step: it researches, writes, tests, prepares.
A person decides.
Between one step and the next, someone approves. The agent proposes; the one who presses the button is a human.
The machine checks.
Before saying "done", an automatic lock checks the proof. Missing proof, no pass.
Keep these signs in mind. They show up in every drawing from here on.
Six stations around a ring. The end of one lap is the start of the next. To keep it concrete, we follow a real app around the whole lap: Placarito, a scoreboard for truco, volleyball and other games.
There are two wheels. The big one is this one, the product cycle, turning at the pace of the people who use it: days and weeks. The small one turns inside the Build station, at the pace of AI: hours.
Listen before building.
sinais.md a list of hypotheses, with numbers and linksIt is likeThe restaurant that reads the competitor’s complaints before writing its menu.
The blueprint before the build.
intent.md · spec.md intent and spec, kept with historyIt is likeThe architect’s blueprint, approved before the first wall goes up.
Short stages, with an inspector at each one.
commit → spec tested code, with the proof attachedIt is likeA building going up floor by floor, with the inspector signing off each slab before the next.
The conveyor to the shop window, and a hand on the button.
store.config.json the app sent to store reviewIt is likeThe shop packs and checks the order. The owner is the one who opens the door.
The thermometer stays on after delivery.
cycle dashboard a dashboard with datesIt is likeThe doctor who follows the patient after discharge.
The end becomes the start.
the next intent.md the intent for the next lapIt is likeThe recipe book that changes with what customers asked for this week.
Not every agent can do everything. Autonomy goes up one step at a time, and moving up a step is a decision written in a file, not made in the heat of a conversation.
Reads only. Changes nothing.
Recommends. The person acts.
Proposes, the person approves, the agent acts.
Acts alone, and the person can switch it off at any moment.
Everything that goes to the store or goes live stays on the third step.
The agent never pressesSubmit
It is like a car: the one that only warns you left the lane, the one that corrects the steering, the one that drives while you watch. Nobody hands over the whole wheel on day one.
Gartner recommends exactly this, agent governance in four levels, and predicts that 40% of enterprises will demote or switch off agents by 2027 for lack of it. Gartner 2026 (opens in a new tab)SAE J3016 (opens in a new tab)
When the agent says "done", an automatic lock checks that technology’s list. Missing proof, the door stays shut.
The list for an iPhone app
done
It is the pilot’s checklist before take-off, and the cord in a Toyota factory that anyone can pull to stop the line. Here, the machine itself pulls it. Gawande 2009 (opens in a new tab)Toyota (opens in a new tab)
In one release, the package was missing three files that did not live alongside the code. The fix did not stop at that app: shipping now checks the package contents and refuses to submit without them.
A lock that always blocks ends up switched off in the first week. So this one gives up after three refusals in a row and turns into a warning. And each decision it makes is recorded in a file that outlives the conversation, with the version of the rules that authorised the delivery.
When AI writes more code, faster, there is more room for a flaw to slip through. So security is not a last step before shipping: it shows up at all six stations. And the testing is done by a tool of our own, open source: RoqueShield.
The question behind everything
Does zero findings mean it is clean, or that the test never ran?
A broken scanner and a clean system show the same screen: zero findings. RoqueShield was built to tell them apart. A scanner that could not verify goes into the report as skipped, and coverage shows as incomplete. It never becomes a silent zero.
It is the smoke detector with no battery: exactly as quiet as the one that works.
Sense
An opportunity that would force us to keep someone’s data on our server fails before it becomes a project.
Decide
The specification decides what will not exist: no account, no server, data only on the device. What does not exist cannot leak.
Build
Every change goes through the RoqueShield scan before "done". A third-party piece with a known flaw and a published fix does not pass, and every new package is checked: it exists, it has an owner, it is not an impostor name.
Ship
Passwords and keys never ship inside the app. Store credentials come in through a vault, without showing on screen or in the conversation history.
Measure
For what is already live, RoqueShield has the outside tests, the way an attacker would see it: the running application, the network, and what the domain exposes on the internet.
Evolve
Every flaw becomes a fix bundle the agent reads: file, line and how to fix it. The fix only counts when the next scan no longer finds the flaw, with every scanner running.
Continuous security testing that refuses to stay silent.
What it does, in five steps
Scan
Runs the right tool for each kind of target: code, dependencies, live application, network.
Merge
The same problem found by two tools becomes one, with the higher severity.
Triage
An AI reorders severity, flags the likely false alarm and suggests the fix, reading evidence already stripped of personal data.
Confirm
On demand, an agent tries to prove one finding for real, inside a controlled attack environment.
Deliver
A fix bundle that another agent reads and turns into a correction. Then it scans again.
The kinds of test
The code at rest
Semgrep
A password left in the code
Gitleaks · Trufflehog
A third-party piece with a known flaw
Trivy
Cloud configuration
Checkov
The running application
OWASP ZAP · Nuclei
The network
Nmap · Greenbone
What shows on the internet
Shodan · subdomains · look-alike domains · GitHub leaks
The agent that confirms a finding runs real attack tools, so its limit is not a sentence in a prompt. It is code checked before every command: an allowlist for each phase, a validated target, no shell, a hard timeout, and lab mode on by default. Every decision, every refusal included, goes to the audit trail.
Personal and card data is scrubbed from the evidence before any AI reads it. With a local model and the cloud switched off, if the local model fails the test fails, instead of sending the finding out.
Open source, Apache 2.0 license: anyone can read it, run it and check how it works.
15 scanners behind a single contract · 302 automated tests · scanning in the acceptance list of 11 of the family’s 11 technologies · measured on Sep 23, 2026
The risk is not theoretical. In a study of 576 thousand AI-generated code samples, open models cited on average 21.7% packages that do not exist, and commercial ones at least 5.2%. Whoever registers one of those names first gets into the project of whoever trusted it. USENIX 2025 (opens in a new tab)
Cost does not show up on the bill by accident: it is decided on the blueprint. So the architecture is chosen at the Decide station, with the cost of running it on the table, and every station watches what it spends, the AI itself included.
The cheapest server is the one that does not exist.
The new apps are born with no account, no server, and processing on the device, AI included. There is no cloud bill that grows with the number of users, and personal data stays out of our reach by design: cost and security come out of the same decision.
It is what the FinOps Foundation asks for: everyone owns what they use, and business value drives technology decisions. And it is AWS’s definition of an optimized workload: the outcome at the lowest price that still meets the requirements. FinOps (opens in a new tab)AWS 2024 (opens in a new tab)
Sense
An opportunity that only works with a server, sync across devices or AI that runs only in the cloud fails before it becomes a project.
Decide
The architecture is chosen here: three designs compete, and the expensive advisor is called for the decision, checking on the map what else changes.
Build
Each task uses the model its size calls for, and whatever a plain program can check spends no AI at all.
Ship
The checklist runs on the machine, before the push, instead of on a paid cloud pipeline. And this site is static: there is no server to pay for or to go down.
Measure
Cost is measured like any other metric. How much context the kit loads into every session is measured before each of its versions ships.
Evolve
What costs more than it returns gets switched off with the number in hand. That is what happened to the paid pipeline and to the many-agent processes.
Three designs before one.
When there is more than one path, three independent designs compete, and the winner takes the best pieces of the others. The first idea anchors the ones that follow; the panel breaks the anchor.
The expensive model only at the expensive moment.
Three roles, three prices: a cheap model scans, a medium one builds all day, and the most expensive one is only called for architecture, irreversible actions, or a lock that failed twice.
Program before AI.
The 12 automatic locks and the checklists run as plain programs: zero AI cost and an answer in milliseconds.
Measure before cutting.
The paid integration pipeline was switched off after being measured: it always failed, for the wrong reason. The 7 many-agent processes stay paused, because one run costs more than a whole day of work.
None of this lives in anyone’s head. It lives in a plugin, a package that any AI session loads when it opens, and that carries a rule learned in one app to all the others.
What has already been decided and is not reopened, and the vocabulary of each product.
The step by step of what repeats: testing in the simulator, building the store page, measuring the impact of a change.
One per technology, looking only at what changed.
They fire on their own: a leaking secret, shipping outside the path, finishing without proof, text that sounds like AI.
Each technology’s acceptance criteria, written for the machine to run.
Whole processes with many agents. Off by default because of cost, and only run when someone allows it.
Store signals, dependency map, secrets vault, cycle dashboard.
34 pieces joined by 45 dependencies. Before touching one of them, the map says what else has to change, and in which order.
The same method runs with Claude, through the plugin in Claude Code and Cowork, and with Codex, through the instruction file it reads. It does not depend on a vendor.
counted from disk on Sep 23, 2026 · kit 2.18.1 · since Jul 18, 2026
The app factory started running the whole cycle on Sep 10, 2026. What follows comes straight from the App Store and from the code history.
A scoreboard for truco, volleyball and other games, with no ads mid-match.
"The ads won’t close. Impossible to use the app." (one-star review of a truco scorekeeper with 10 thousand reviews)
Stylish QR codes, Pix included, made on the iPhone itself.
"An absurd amount of ads or $$ for something so simple." (review of the QR generator leading the App Store search)
A packing checklist that runs entirely on the device, with a local assistant.
The founder’s pick to test the whole cycle on a simple app. Then nine laps driven by his own use.
A food diary that runs entirely on the iPhone, with the Brazilian food composition table.
Ads and subscriptions are the two biggest complaints across 50 thousand diet app reviews, and the best-known Brazilian app has been stalled since 2024.
A planner for small gatherings: the host’s tasks, shopping and schedule.
Invitation apps look after the guests. Nobody was looking after the host.
A guide and routine for leaving your pet with someone, readable by the sitter without installing anything.
Whoever travels needs to hand over the pet’s routine, medicine and contacts in a clear guide, without sign-ups.
the code history starts after the work did
The six above came out of the app factory. The rest of the house runs on the same kit, with the same locks.
The lab: an entire operating system in the browser.
Windows, a dock, Launchpad, a Docker container store and AI agents, all in one tab. The same URL runs on the web, iPhone, Android, Windows and TV, and the app is approved on all three stores. It is where the other products are born.
The musician's toolkit, running entirely on the device.
Tuner, metronome, real-time separation of a song’s instruments, a simple recording studio and a virtual guitar through the camera. It started as the founder’s own study tool and now has a community asking for features.
The native apps for Nelson Borges' devotional project.
myABBA is Nelson Borges’s non-profit devotional platform. The website and the content are his; the iPhone and Android apps came out of the same kit as our other products.
Stage light on your iPhone.
A dot-matrix LED sign at 120 Hz, a mode that pulses with the music and strobe on the flashlight. It works offline, in ten languages.
The app for opening Markdown on Windows.
A reader, an editor and both side by side, with synced scrolling. Free, no account and no ads, on the Microsoft Store. It is the app in the lap told just below.
Continuous AI security, open source.
It scans code, dependencies, the running application and the network, uses AI to triage the findings and hands over a fix package that agents such as Claude Code, Cursor and Copilot can read. It looks after security in the cycle, as level 06 shows.
MDViewer, our Markdown app for Windows, is the most complete example of a closed lap: a review on the Microsoft Store became a fix, the fix reached the store, and the reply went back to the person who wrote it, in the same place.
"Chief gripe is it does not remember screen position and always opens off-screen with only an edge visible… If this was fixed I’d pay for it."
Written and sent by the agent after the fix went live. It names the defect and states the real cause, which was not the large monitor.
Measure
The agent reads the reviews of the new versions in Partner Center, Microsoft’s area for publishers. There are six: four with five stars and two with three, each pointing to a concrete defect.
Evolve
The two three-star reviews become the next round. First, the cause: the user blamed his large monitor, but the app saved the window position and never checked whether it still fit on any monitor.
Build
The app now checks that at least 120 pixels of the title bar remain inside some monitor. The test deliberately saves a position outside all of them and confirms the window comes back to the centre. The same round brings what the other review asked for, a right-click menu in the editor and scrolling synced both ways, and the search panel now speaks the app’s three languages.
Ship
Version 2026.9.1300.0 passes Microsoft’s certification and reaches the store on September 13, three days after the review. It is checked against the Store’s own catalogue, not the dashboard.
Operate
The agent replies to the reviews directly in Partner Center. The rule was agreed before writing: no "working perfectly"; each reply names the defect the person hit and cites a number that can be checked, such as crashes, down from 14.52% in the old version to zero. The two three-star reviews waited for the fix to go live, by the founder’s decision.
From MDViewer’s evidence log, rounds 16 to 19, September 13, 2026. The version was checked against the Store’s own catalogue.
The cycle is the sum of public ideas, some of them decades old. What we did was assemble the version that runs, with locks, and publish what came out of it.
The cycle starts from what users say, not from someone’s request.
Done only with proof, and the proof is demanded by an automatic lock.
Each agent’s autonomy is written data, not a conversation.
A single toolbox for every product, with more than one AI agent.
The result is published, with dates.
Not here. It executes. The person decides what to build, approves every handoff and presses the final button.
The new apps took 2 to 8 days from first commit to the App Store, including Apple’s review.
The skeleton is public, and the sources are right below. What LEVELHARD shows here is the version that runs, with the locks and the numbers.
It does. This site was built that way, in rounds with tests, screenshots checked by eye and shipping only on the founder’s word.
Every outside number comes from a primary source, checked on Sep 23, 2026. When it is a vendor’s number, we say so.
We do not believe in slideware AI.We believe in the kind that runs.
The apps are on the stores and the door is open: AI in the development cycle, a product to build together, or plain curiosity about how this is made.