Directing AI coding agents

Last updated 2026-06-12 - 7 min read - By Divyam Chandel - Cohort instructor

A reference for non-engineers learning to direct AI coding agents: the standing brief, the verification ladder, and the loop that connects them.

The core skill of building apps with AI is not prompting. It is Agent direction Agent direction is the practice of giving AI coding agents context, boundaries, verification steps, and small next moves instead of asking for broad code generation. Open full term: giving the model a narrow job, enough context to avoid obvious damage, and a way to prove the work it claims to have done.

This page is the durable reference for that skill. The two long-form essays linked at the bottom carry the argument. This page carries the procedure.

The director’s loop

Every change a non-engineer makes to an AI-built app should go through the same five-step loop. The loop sits on top of two persistent artifacts (the standing brief and the verification ladder) and runs every time you ask the agent to do anything.

  1. Name the user job. Not the feature, the job. “The rep needs to be reminded on Thursday that Anna’s lead has not replied.”
  2. Cut the change down. One job per request. If the request spans two screens, split it.
  3. Tell the agent what not to touch. Name the files or flows that must not change.
  4. Ask for the smallest implementation. The version that passes the user job and nothing more.
  5. Run the verification ladder. Three minutes, five rungs, before you say done.

The loop is the same for every change. The work it produces compounds, because each successful pass leaves the brief stronger and the agent’s context cleaner.

The two artifacts

The loop only holds if two artifacts already exist in the project.

The standing brief. A markdown file at the root of your project (commonly named CLAUDE.md for Claude Code, AGENTS.md for Codex, .cursorrules for Cursor). The brief tells the model what your product is, what it must never do, what product-specific words mean inside your codebase, and the four-line self-check it must run before claiming a change is done. It is the durable context the model reads at the start of every session. Without it, every prompt is starting from zero. The companion essay I read 64 CLAUDE.md files walks through the four sections engineers leave out and the smallest useful version a non-engineer can write.

The verification ladder. A five-rung procedure the non-engineer runs after the model claims a change is done. The rungs are: read the agent’s claim word for word, list the files that changed, walk the flow you asked for, walk a flow you did not ask about, and ask the agent for evidence. Three minutes, no coding required. The companion essay The model said done. The app was broken. defines the ladder in full and walks through a worked example.

The brief is the before. The ladder is the after. The director’s loop above is the connecting tissue.

What goes wrong

Most failed AI coding sessions start with a request that is too broad. The agent tries to satisfy everything at once, then patches the patch, then rewrites an unrelated part of the app. Three days later the app is held together by hope.

The fix is not a clever phrase in the prompt. The fix is a smaller unit of work, a brief that anchors what cannot change, and a check the model has to run before it claims completion.

Anthropic’s own best-practices guide for Claude Code admits this directly: “Claude stops when the work looks done. Without a check it can run, ‘looks done’ is the only signal available.” The model is being honest about its own state. Closing that gap is the non-engineer’s job. The artifacts above are how you close it without writing any code.

What to read next

If you want a workshop instructor in the room while you write your first brief and run your first ladder, the Saturday workshop covers exactly this.