Shipping without code: a non-technical builder's playbook for AI coding agents
You can describe a feature in detail. You can sketch the flow and write the copy. For years, none of that shipped without an engineer, because the last step was code.
AI coding agents now do that step. Your job changes. You do not write code. You say clearly what you want, and you judge what comes back. If you have ever briefed a contractor or reviewed a design, you already do this.
This post is the how-to. For a short story of one real task, read Ship a product idea without an engineer.
Write the brief as a checklist
Agents build what you write. A vague brief gives a vague result. Three habits help.
Write acceptance criteria, not wishes. “Make the settings page better” is a wish. This is a checklist:
- [ ] The settings page has a dark-mode toggle
- [ ] The page remembers the choice after a reload
- [ ] The default follows the system setting
Each line has a yes or no answer. An agent cannot finish the task until each box is ticked.
Say what is out of scope. A short “not this time” list, such as “do not change billing”, stops an agent from guessing.
Give the context only you have. Say who the feature is for and what would make you reject it.
Short briefs also work for small tasks. One real task on our site was a single sentence: “Drop down arrow cut off on mobile. There is now also no way to get to your profile on mobile.” It merged for $20.87. See the example. Our guide on writing a brief has more patterns. The brief checker tests a draft, and starter briefs give you a place to begin.
Choose where a person approves
By default, no person approves a plan or a result. Agents plan, code, review, merge and verify on their own. Another agent does the review.
You can add a human check. An approval gate holds a task when it leaves a stage. An admin turns it on in the workflow, with Approval required to leave on a stage. This needs the Starter plan or higher, because it means editing a workflow.
- Gate the plan stage. You then read the plan before any code exists. This is the cheapest time to change your mind.
- Gate the review stage. You then approve the result before it merges.
When a plan waits for you, ask three questions. Does it solve what you asked? Did it keep the limits you set? Does it do anything you did not intend? Reject a plan with one sentence of feedback. That costs less than a rebuild.
Review what comes back
You do not have to read code. You would not audit an engineer’s changes line by line either. You trust the process around them. Here that process is explicit:
- Review: another agent reads the change against your checklist.
- Merge: the reviewed change joins the main code through a merge request.
- Verify: an agent confirms that the change works and that the checks on the main code pass.
Read the hand-off comments, the summary of what changed and the verify result. They are short and in plain language. Can you trust AI-generated code? explains why these gates matter.
Watch the cost
Every task shows what it cost, next to the task. Real examples on our site run from $3.05 for a missing menu link to $29.54 for a feature that touched 58 files. See all the examples.
You can also cap spend. Spend caps need the Pro plan. An owner or admin can set a default cap for every task in the account. When a task reaches that cap, its sessions stop and the task moves to blocked. No person gets a warning first. A budget on one task or on a workspace is different. At 80% of that budget, the owner gets a warning, and work continues. For more, read Cost-aware AI engineering.
Know when to ask a technical person
Agents ask you a question when they need a decision. Answer it in the Dashboard. If you want to watch, open any agent’s live session in your browser.
Some choices you cannot judge from a summary. Examples are how login works, how the system behaves under load, and whether a third-party service should hold customer data. Ask a technical person before you approve a plan that depends on one of these.
A developer must connect your repository and one device first. You do this once.
Start small
Pick one idea you can describe in a few lines. Write it as a checklist. File it and read the summary. Then do the next one. When you are ready to roll this out to a team, see the rollout guide.
Want to try it? Request access and file your first brief.