What Codex and Claude Code Recommend: Visibility for Developer Tools
Coding agents now suggest libraries and services while developers work. Here is how to see what they recommend about your tool and what to fix when it is wrong.
When a developer types "add authentication to this app," the assistant does not just write code. It chooses a library, a hosted service or a pattern. That choice is a recommendation, made inside an editor, often with no search page in sight. For developer-tool companies, that is a new place to win or lose a customer, and it does not look like anything in your classic SEO reports.
Why this surface is different
Three things set it apart from a chat answer.
- The recommendation arrives as action. The agent may install the package, not just name it.
- There is rarely a list. One suggestion, maybe two, with no carousel of alternatives.
- The context is the project. The agent reads the repo, the language and the existing dependencies, which shapes its choices.
That makes first place even more valuable than in a text answer, and the second place almost invisible.
What Promptwatch tracks here
According to the Promptwatch fact sheet, it tracks what Codex, Claude Code and OpenCode recommend. Codex and Claude Code are monitored among the real-UI coding agents, and OpenCode is tracked via API. The use case is developer-tool brands that want to know whether coding agents suggest them.
You treat these agents like any other engine in your prompt panel: write prompts that mirror developer tasks, run them daily and look at which tools are named.
Writing prompts for coding agents
Developers phrase tasks, not categories. Instead of "best auth provider," write the task: "add email and password login to a Next.js app." Tag prompts by language, framework and job to be done.
Vary the context. A request in a TypeScript project differs from one in Python. Persona stacking in Promptwatch lets you layer attributes such as language and experience level.
Include competitor-comparison prompts too. Developers often ask for alternatives or for migration paths.
What to read in the results
- Are you named at all? Many tools are not, even popular ones.
- In which tasks? You may win "send transactional email" and vanish for "set up background jobs." The gaps show where your documentation or examples are thin.
- What does the agent say about you? Sentiment and description reveal outdated or wrong claims, such as deprecated APIs the agent still suggests.
- Who beats you? Competitor names per task show who owns each job.
Why agents pick one tool over another
We have no sourced study on how these agents choose, so we will not offer a theory as fact. What you can observe and act on are the inputs agents can read: your documentation, examples, package metadata, READMEs and the web pages that discuss you. Each is something you control or can influence.
Practical steps:
- Make documentation crawlable. Check that AI bots can reach your docs. Agent Analytics shows IP-verified crawler hits and errors, and the real bot names to look for include GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot and others.
- Write task-shaped pages. A page titled for the job ("send email from a Next.js route") matches how the agent frames the problem.
- Keep examples current. Wrong code samples in old pages can spread.
- Publish a clear llms.txt and robots.txt. Promptwatch offers generators for both, though neither is a substitute for good content.
- Be present where developers talk. Reddit and YouTube citation reports show which community sources get cited.
Measuring change
Track the same prompt panel weekly. When you publish a new guide, add it to the Page Tracker and watch crawls and citations. Then check whether the agent's recommendation for the matching task changes. Expect a delay. We do not have data on typical lag, so measure yours.
If you use the Promptwatch MCP server, you can bring the data to where developers already work: the hosted MCP server has 110+ read and write tools, and the catalog lists Cursor and Claude among the agent integrations.
Be careful about attribution
A recommendation from a coding agent may not produce a referral you can see. The developer installs the package and never visits your site. Look for downstream signals you already have: package downloads, signups with a developer source, documentation traffic. Do not claim a link we cannot show.
What good looks like
A healthy position: your tool is named for the tasks it is built for, described accurately, and not displaced by competitors on those tasks. A common unhealthy one: named for an old API, or never named for the newest feature.
Recommendation
If your product is a library, an API or any service developers choose, add coding-agent prompts to your panel now, before competitors have defined the answers. Promptwatch tracks Codex and Claude Code in the real UI and OpenCode via API, alongside the crawler logs and citation data that explain why. Start with ten task-shaped prompts and read the results next week. The first surprise is usually a task you assumed you owned.