Guide · Agent skill
How to use the Improve Convex Plugin Review agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
This guide explains the convex-improve-convex-plugin agent skill, a packaged set of instructions that lets a coding agent send your current session transcript to the Convex team for an AI post-mortem. It is aimed at developers who use Convex and have not used agent skills before. By the end you will know what the skill does, how to install it, how to trigger it, and what to check before any transcript leaves your machine.
Improve Convex Plugin Reviewopenclaw/clawhub · MIT Download the skill Quick start
- 1Download and unzip. Download the skill ZIP from this page and unzip it so you have a folder containing SKILL.md.
- 2Place the folder. Move the folder into ~/.claude/skills/ for all projects, or into .claude/skills/ in one project. Alternatively upload it under Settings → Capabilities → Skills in the Claude apps.
- 3Check your session. Confirm the session is one you are willing to share, with no confidential code, customer data or secrets.
- 4Ask in plain language. Tell the agent you want to send this session to the Convex team to improve the quickstart.
- 5Answer the consent prompt. When asked, choose Always, Just this once, or Never. The agent will not send anything until you answer.
- 6Review the summary. Read the highest-severity findings and wins the agent reports back, which focus on Convex's system rather than your app.
What does the convex-improve-convex-plugin skill make the agent do?
An agent skill is a folder containing a SKILL.md file of instructions that your coding agent reads when a matching request comes up. This skill tells the agent to run a helper script fetched from the Convex service. The script uploads the transcript of your current Claude Code or Codex session (a .jsonl file on disk) to an external review endpoint.
The review returns structured findings: where instructions were ambiguous, patterns where the agent got stuck, tooling failures, and wins. The findings are aimed at Convex's runbook, bootstrap script, skills and components, not at your app.
How do you install the skill?
Download the ZIP from the skill's page on this site and unzip it. The skill folder contains SKILL.md. Put the folder in one of these locations:
Use the user-wide location ~/.claude/skills/ so the skill is available in all your projects, or the project location .claude/skills/ inside a specific project so it is only available there.
Alternatively, in the Claude apps, open Settings, then Capabilities, then Skills, and upload the skill there.
- User-wide: ~/.claude/skills/
- Per project: .claude/skills/ in the project root
- Claude apps: Settings → Capabilities → Skills, then upload
How do you trigger the skill with a natural request?
You do not need a special command. Ask the agent in plain language, for example that you want to send this session to the Convex team so they can improve the quickstart. The agent matches your request against the skill's description: Send this coding session's transcript to the Convex team for an AI post-mortem that improves the quickstart system.
The agent will then run the helper and pass a one-line description of your app idea from the session.
- Send this session to the Convex team so they can improve the quickstart.
- Run the Convex post-mortem on this coding session.
- Share where the Convex setup instructions got confusing in this session.
How does consent work?
Sharing is opt-in. The first time, the helper asks you to choose Always, Just this once, or Never, and it remembers your choice. The skill's rules say the agent must not send a transcript until you have explicitly chosen.
If you have not answered yet, the helper exits with a CONSENT_REQUIRED message and exit code 4. The agent then asks you the question and re-runs the helper with your answer.
What do the output markers mean?
While the review runs, the agent watches for three markers. REVIEW_SOURCE means a transcript was found. REVIEW_SUBMITTED id=... means the review service accepted it. REVIEW_DONE status=done means the findings are ready.
If you see REVIEW_NO_TRANSCRIPT, no Claude or Codex .jsonl file was found on disk, so there is nothing to send.
- REVIEW_SOURCE: transcript found
- REVIEW_SUBMITTED id=...: accepted by the service
- REVIEW_DONE status=done: findings ready
- REVIEW_NO_TRANSCRIPT: no .jsonl transcript found
What do you get back and how is it summarised?
The agent summarises the highest-severity findings as title, then target, then suggestedFix, followed by the wins. The summary is meant to stay about the system rather than your data.
Before any upload, check what is in the transcript. The skill's script says it redacts keys and tokens, but the files shown do not verify that, so do not rely on it alone.
What do you need, and what about safety?
The skill requires curl and bash, access to the anteater review service run by Convex, and a Claude Code or Codex session transcript (.jsonl) on disk. It runs scripts and uses the network.
Do not use it on sessions that contain confidential code, customer data or secrets you do not want leaving your machine. The transcript is uploaded to a third party. The skill is a system-improvement loop only, so it is not a private code review tool or a place to send normal feature requests.
- Requires curl, bash, and network access to the anteater service
- Runs a script fetched from the Convex service
- Uploads the full transcript to an external endpoint after consent
curl -fsSL "<anteater>/send-transcript" | bash -s -- --idea "<one-line app idea from this session>"What people use it for
- Reporting where Convex setup instructions were unclear during a new project onboarding session
- Flagging a bootstrap script or tooling failure you hit while getting Convex running
- Giving a Convex onboarding team structured feedback drawn from real coding sessions
- Recording patterns where the agent got stuck on the quickstart so the docs can be improved
Frequently asked questions
Is the convex-improve-convex-plugin skill free?
The skill is published under the MIT licence. The source does not mention pricing, so check the page and repository for any terms about the Convex review service itself.
What do I need to run it?
You need curl and bash, network access to the anteater review service run by Convex, and a Claude Code or Codex session transcript in .jsonl format on disk. The skill runs scripts, so it is not a purely text-based instruction set.
Where do I install the skill?
Unzip the downloaded folder into ~/.claude/skills/ for all projects, or into .claude/skills/ in a specific project. In the Claude apps you can instead upload it under Settings → Capabilities → Skills.
Does it send my code to Convex?
It uploads the transcript of your current coding session to an external review endpoint after you consent. The transcript can include code and text from the session, so it is not suitable for confidential material. The script says it redacts keys and tokens, but this is not verified in the files shown.
Can I stop it from sending my session?
Yes. Nothing is sent until you choose Always, Just this once, or Never when asked. Choosing Never stops transcripts from being sent under that choice, and the helper remembers it.
Can I use it for private code review or feature requests?
No. The skill is a system-improvement loop for Convex's quickstart, docs and tooling. It is not an end-user feature feedback tool or a private code review tool, so use other channels for those.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
This skill suits Convex developers and onboarding teams who want to help improve the quickstart, and who are comfortable sharing a session transcript with a third party. Do not use it on sessions with confidential code, customer data or secrets, and do not pick it if you want private code review or normal feature feedback.
Download the skill