# How do I build an app with an AI coder?

Checked 2026-11-15 against GilbApp d5296208bbffa2ab70076e6194f5537c4be1ddf7.

Start Claude Code in Terminal or [a code folder](http://support.gilb.com/help/how-do-i-let-my-ai-build-in-a-folder-on-my-computer-from-gilbapps-chat), type the agent line with Agent.Coder and what to build, then read its handback and verdict.

## Before you start

Claude Code installed, [connected to your vault](http://support.gilb.com/help/how-do-i-connect-my-computer-to-my-vault) and [bound to its AI vault](http://support.gilb.com/help/how-do-i-make-an-ai-vault-and-connect-claude-code-to-it), each by its own answer, and your short name for it.

## Steps

1. Start Claude Code in Terminal, or connect it in GilbApp's chat and [choose a code folder](http://support.gilb.com/help/how-do-i-let-my-ai-build-in-a-folder-on-my-computer-from-gilbapps-chat).

   Open Terminal, type claude and press Return, as [the answer on installing Claude Code](http://support.gilb.com/help/how-do-i-install-claude-code) shows.

   Agent.Coder, the agent that builds apps, works on code kept on your computer. GilbApp's chat does not reach the files on your computer unless you [choose a code folder](http://support.gilb.com/help/how-do-i-let-my-ai-build-in-a-folder-on-my-computer-from-gilbapps-chat) while Claude Code is connected; otherwise code kept there needs Claude Code in Terminal. An agent is your AI following one written procedure, its standing instructions, for one kind of work. GilbApp has no steps for using Grok in Terminal.
2. Type the agent line with Agent.Coder and press Return.

   Type this line with your own short name, the one you gave your AI when you [connected your computer](http://support.gilb.com/help/how-do-i-connect-my-computer-to-my-vault), and press Return:
   /mcp\_\_gilb-&lt;short name\>\_\_agent Agent.Coder

   In a new Claude Code session it starts as Kai-Zen; this line loads Agent.Coder's standing instructions instead. Today four agents can be started this way or chosen in GilbApp's chat: Kai-Zen, Jackfruit, Agent.Concept and Agent.Coder.
3. Type what to build and press Return.

   Say what the app or the change should do, in your own words, and press Return. You do not need to code yourself.

   Agent.Coder builds apps on GilbPlatform in TypeScript, with React for the app, and reads the chapters of the Developer Guide its work touches. Before it writes a new piece of code, it looks for a ready one in the code library; [the answer on reusing app code units](http://support.gilb.com/help/how-do-i-reuse-app-code-units) shows that library. Before it writes code that stores a new kind of data, it hands the terms to Agent.Concept, the agent that finds and makes GilbVeda's terms; GilbVeda is the shared list of terms. It keeps its project files in the project's folder.
4. Read the handback and the tester's verdict.

   When it hands back, read the verdict, PASS, PARTIAL or FAIL, and the path of the tester's saved report. Agent.Coder has each piece tested by Agent.Tester, the agent that tests what others built, started as a separate agent that knows nothing of how it was built.

   The report is saved as QA-&lt;date\>-&lt;increment\>-&lt;project\>-&lt;verdict\>; a verdict that is only in a reply does not count. If Agent.Coder cannot start a separate tester, it tests the piece itself and marks the result self-QA, which may stand in for a PASS from Agent.Tester but is not equal to one. For a change to a screen it adds the path of a screenshot and the count of errors the browser reported. On a FAIL it fixes and tests again, and it stops after three different failures it cannot fix without more scope or your input. [The answer on reading what an agent hands back](http://support.gilb.com/help/how-do-i-read-what-an-agent-hands-back) explains the rest of the reply.

## Deeper

Can you build with your AI coder alone, without Kai-Zen? Yes: these steps start Agent.Coder, the agent that builds apps, directly. It still builds, has each piece tested by a separate agent \(if it cannot start one, it tests the piece itself and says so\), and asks you before it puts an app live. What you miss is what Kai-Zen, the planning agent, adds: a plan agreed with you of what the app must do and how well, steps picked to bring the most for the least effort, your check of each step once it is live, a measure of how far each goal has come, and a plan kept up to date with what was learned. Without Kai-Zen, you decide yourself what to ask for next.

Agent.Scaling, the agent that checks what was built, checks that each part has one job and deals with other parts only through their agreed interface, that nothing the app already has was built again, and the build against the project's Values for the whole project. If an agent called it, the work does not go on while a problem is open; if a person called it, the person decides. Agent.Coder applies its checks too. You cannot choose Agent.Tester or Agent.Scaling yourself: Agent.Coder starts Agent.Tester, and Kai-Zen has Agent.Scaling check each step it delivers.

For a project, Kai-Zen runs the whole round: it picks a step, hands it to Agent.Coder for an app on GilbPlatform, has what was built checked by Agent.Scaling and tested by Agent.Tester, and puts it live only with your go. [The answer on how Kai-Zen runs a whole delivery cycle](http://support.gilb.com/help/how-does-kai-zen-run-a-whole-delivery-cycle) tells more.

A pass on your own computer is not production, the live system people use. Agent.Coder does not do acts on production that need a person at the production machine. In GilbApp's chat you can also choose "Agent.Coder - Master of Coding", but there it cannot reach code on your computer unless you choose a code folder while Claude Code is connected. The answer [How do I let my AI build in a folder on my computer from GilbApp's chat?](http://support.gilb.com/help/how-do-i-let-my-ai-build-in-a-folder-on-my-computer-from-gilbapps-chat) shows how.

This answer is written for a computer screen.

Source: Standards read 11 November 2026: Agent.Coder \(A 06gf80g4e5xhzfm96gsrtbh8g4\) Role \("working with HumanAgent, who does not code"; Gilb stack\), GilbStack items 2 \(Dev.Guide.Overview, "read the chapters your increment touches"\), 4 \(terms to Agent.Concept before code\), 6 \(Agent.Scaling's checks still apply\), 7 \(look it up in the GilbPlatformCode Library first\), ProjectFolders, Solutions \("A pass on your own computer is not production"\), AfterIncrement \(separate Agent.Tester, RuleQatestIsolation; Fallback self-QA "may stand in ... but is not equal"; UI proof: screenshot path and console error count; "Stop after three distinct failures"; the page says "three different failures"\), HumanActs; Agent.Tester \(A 06gf80hsdxwmb8hjcqj901790m\) Role \("You did not build this code"\), Task \(QA-&lt;date\>-&lt;increment\>-&lt;project\>-&lt;verdict\>; "A verdict that is only in your reply does not count"; "Return the verdict and the report's path"; "with a verdict per RuleQatestPartialIsLawful"\); RuleQatestPartialIsLawful \(A 06gezpb2kxytqdxk62hbgsx904: "The verdict vocabulary is FOUR words ... PASS · PARTIAL · FAIL · UNWALKED \(a per-criterion mark\)"; "PARTIAL and a named FAIL are both lawful and WANTED"\); Agent.Scaling \(K agent-scaling-06gd2d2b4nz9152pb3tzcp4d5g\) Role, Parts, Reuse, Values \("the architect's Values and their Tolerable"\), Block. Agent list https://id.gilb.com/base/ai/agents \(read 11 November 2026\): picks Jackfruit, Kai-Zen, Agent.Concept, Agent.Coder \("Master of Coding"\); Agent.Tester, Agent.Scaling pick false; GilbApp ebd9415e \(live at https://gilb.com/app/version.json, read 11 November 2026\) src/components/agent/AgentSelector.tsx:15 \("Name - Description"\); gilb-connector 0.5.13 \(vendor/gilbplatformcode-gilb-connector, at ebd9415e\) README line 109 \(gilbplatform\_agent switches the agent\) and test line 1388 \(a name that is not a pick loads no standing\); GilbPlatform 498d29a0 \(live at https://id.gilb.com/version, read 11 November 2026\) src/ai/chat-agents.ts:113-149 \(only a pick loads\). Shared wording, word for word: "Which agents are there, and which do I use for what?" \("GilbApp's chat does not reach the files on your computer unless you choose a code folder while Claude Code is connected; otherwise code kept there needs Claude Code in Terminal.", the /mcp\_\_gilb-&lt;short name\>\_\_agent line\), "Can I use any AI, not only Claude Code or Grok?" \("GilbApp has no steps for Grok in Terminal."; xAI's own terminal coding tool Grok Build, https://docs.x.ai/build/overview \(read 13 November 2026\)\), "How does Kai-Zen run a whole delivery cycle?" \(Agent.Scaling "the agent that checks what was built"; the round\), "How do I read what an agent hands back?" \(Agent.Tester "the agent that tests what others built"; self-QA sentence\), "What counts as proof that something works?" \("A pass on the builder's own computer does not count as working in production, the live system people use"\), "Which agents are there, and which do I use for what?" and "How do I pick the next step?" \(agent gloss\). Rows: KaiZen.BuildWithCoder, TestIndependently, CheckBuildScales. Agent.Coder \(AfterIncrement: Agent.Tester as a separate agent; AskGoLive: never publish without the person's yes\), Kai-Zen.Cycle \(Step, Human, Measure, Update, Improve, Next\), Kai-Zen.ValueDeliveryStepRules \(Question: the most good for the least effort\). Live answers, sentences shared: How big should my first version be?; How does Kai-Zen run a whole delivery cycle? Code-folder exception: GilbApp 49b1ae53 \(https://gilb.com/app/version.json, read 13 November 2026\), src/components/agent/AgentChat.tsx:677-692 and :1084-1090 \("Choose a code folder", "Code in", Claude Code only\); live answer How do I let my AI build in a folder on my computer from GilbApp's chat?
