I kept giving AI the same product checks

I was building more products with coding agents, and I kept running into the same problem. The first version arrived fast. Then I had to ask the agent to go back and check the things that make it a solid product.

The prompts were often the same. Check the message. Check the main action on mobile. Check keyboard navigation and accessibility. Check performance, translations, metadata, and the details that keep the quality stable from one release to the next.

AI could build the first version. It still needed the product standards I had learned from design, product, and years of shipping software. Repeating those prompts is what led me to build FixFlags.

Put the product standards in one review

I called it Product QA because I was checking more than the interface. I wanted one review that looked at whether the product explained itself, whether people could use it, and whether the right information reached search engines and social platforms.

Those three areas became Message, Experience, and Reach. Message checks the positioning, hierarchy, trust, and next step. Experience checks interaction, mobile behavior, accessibility, performance, and the paths people need to complete. Reach checks SEO, metadata, structured data, and social previews.

Every check has to point to evidence from the live product. If FixFlags cannot show what it found, the check does not belong in the report. This keeps the review tied to the product instead of turning it into a score based on taste.

The Product QA review

The report groups each check under Message, Experience, or Reach, then shows the Flags and the evidence behind them.

• FixFlags logo, orange F and wordmark on white: FixFlags identity
• FixFlags report overview with ranked unresolved Flags: Report overview
• FixFlags rubric evidence organized across Message, Experience, and Reach: Rubric evidence

Show what to fix first

A score can tell you that something is wrong without telling you what to do next. I built the report around ranked Flags instead.

Each Flag names the problem, shows where it appears, includes the evidence, explains why it matters, and provides a fix prompt. Severity and confidence help put the work in order. The report can say: fix this message first, repair this interaction next, then add the missing SEO or metadata.

The fix prompt includes the context, the intended change, a limited scope, and a way to verify the result. That matters when a coding agent can edit the product. A broad prompt can change too much. A focused prompt gives the agent one problem to fix and tells it what to preserve.

The evidence and the fix

Each Flag shows the problem, the evidence, and a focused prompt for fixing it without changing the rest of the product.

• FixFlags Flag detail with severity, evidence, and explanation: Flag detail
• FixFlags bounded fix prompt ready to copy into an AI builder: Fix prompt

Send the fix to the coding agent

The report only solves half the problem. The change still has to happen in Cursor, Claude Code, a terminal agent, or another builder. I did not want people to copy the finding and explain it all over again.

The first option was a fix prompt that could be copied into any builder. Then I added MCP, so a compatible agent can read the report, list the unresolved Flags, and pull the full context for one fix. The CLI brings the same review and fix workflow into the terminal.

The coding tool stays where the changes happen. FixFlags gives it the product check, the evidence, and the instructions for the fix. I used the same direct approach for the brand: one orange F, a compact wordmark, and a report that looks like work to do.

Run the checks again

Once the changes are live, an update review runs the same checks on the same URL. The report shows which Flags were resolved, which ones remain, and what changed. The work stays in one place instead of starting over with a new audit.

Deep reviews handle important journeys that need more than a page check. They walk through the steps, capture what happened, and return evidence from the experience. Product reviews show what is unfinished on the surface. Deep reviews show where the path breaks.

Reports are private by default. Sharing is a separate choice and can be turned off again. A product can stay private while it is being fixed, then the report can be shared when another person needs to review the work.

FixFlags is in open beta at fixflags.com. It is the product I wanted when I was repeating the same prompts: check the live product, show me what is wrong, send the fix to the coding agent, and check it again.

Fix it and check again

Prompts, MCP, and the CLI send the work to the coding agent. An update review checks the live product again and shows what changed.

• FixFlags update review showing resolved and remaining Flags: Update review
• FixFlags MCP and CLI workflow for bringing report context into an AI coding agent: MCP and CLI workflow