How Agent Plugins Get Discovered and Recommended

Improve agent plugin discovery through useful outcomes, clear capability descriptions, activation, retention, and release quality across ChatGPT, Claude, and other runtimes.

To help an agent plugin get discovered, make its purpose clear and its results useful. Then measure whether people succeed and return. Telvine helps builders maintain that improvement process across ChatGPT, Claude, and other agent runtimes, with one plugin product record and platform-specific packages.

What OpenAI has said about recommendations

In his October 4, 2026 appearance on Lenny's Podcast, Tibo Sottiaux described retention, successful use, and quality as considerations for OpenAI's plugin recommendations. He also explained that poor plugins can stop being recommended.

That is a statement about OpenAI's approach, not a published ranking formula or a rule for Claude. Use it as motivation to improve the product's utility. Telvine does not control marketplace rankings or guarantee recommendations.

Make the capability easy to understand

Name the job the plugin performs, the inputs it needs, and the result a user can expect. Keep the listing, Skill descriptions, documentation, and examples consistent. A clear description helps people choose the right product and helps a runtime identify a relevant capability.

For example, an invoice-exception plugin could promise: review supplied invoice and payment exports, identify unmatched or partial payments, and produce a review table. That gives a builder concrete acceptance criteria and gives a user a reason to try it.

Use Skill triggering tests to check intended requests and nearby requests that should not activate the capability. Accurate triggering is a quality check; it is not proof of marketplace visibility.

Make the first run useful

Test the installed package on each supported surface. Check missing files, unavailable connectors, unclear date ranges, and conflicting records. Explain what the user needs to provide and preserve progress when a dependency fails.

For the invoice example, a useful first run identifies exceptions with reviewable references. It should avoid silently inventing matches or changing accounting records. Test those boundaries with synthetic fixtures before sharing the release.

Measure whether people return successfully

Track first observed use, first successful outcome, and repeat successful use. Define success for the workflow rather than treating every tool call as a completed job. Count distinct observed instances and choose a follow-up window that matches how often the work occurs.

A weekly exception review should offer value on the next reporting cycle. Repeated retries after a failure do not demonstrate that value. A monthly close workflow needs a longer window, and a one-time migration may be judged primarily on completion and quality.

Follow the worked cohort example. Separate platform and version results so differences between ChatGPT and Claude usage do not obscure release regressions.

Improve the release rather than chase call volume

Turn failure patterns and feedback into eval scenarios. Compare the candidate against the stable version on completed outcomes, repeat successful use, errors, and latency. Investigate tradeoffs: a faster result can still be less accurate, and more invocations can still mean more retries.

Keep configuration simple for users. Document the inputs, permissions, expected result, and recovery steps; let the implementation evolve as models and runtimes improve. Preserve deterministic checks where they protect correctness.

Publish and measure with Telvine

npm i -g @telvine/cli
telvine login
telvine publish ./my-plugin

Use skill.* events for Skill behavior and plugin.component.invoked and plugin.component.error for other observable components. Keep telemetry limited to permitted metadata. Never emit prompts, file contents, connector payloads, tool arguments, or model outputs.

Publishing to Telvine maintains your plugin release record. Distribution and review follow the requirements of each target marketplace. Pair release evidence with actual user feedback instead of assuming publication alone demonstrates product quality.

Frequently asked questions

Does this replace website SEO? No. Website search can help builders find your documentation. Marketplace discovery, runtime capability selection, and user retention are different stages. Measure each only where you have a reliable signal.

Does Claude rank plugins using the same signals? This guide makes no claim about Claude's recommendation algorithm. Successful outcomes and repeat use remain useful product measures regardless of the distribution mechanism.

Can Telvine guarantee recommendations or revenue? No. Telvine helps you evaluate and improve releases. Marketplace decisions and commercial terms belong to the platform operating them.