# I built an AI code review bot in 30 minutes - here’s how

How I AI (Claire) mini-episode: build a PR risk-scoring / auto-approval bot with Vercel Eve plus Codex, inspired by Intercom's AI-approved PRs and Rewind's Diff Vader. Explicit contrast: Eve is easier than Hermes Agent / OpenClaw for Slack plus GitHub internal agents.

Tags: youtube-inbox, source, code-review, vercel-eve, github, slack, agents

## Summary

Engineering orgs drowning in AI-authored PRs should not human-review every diff. Pattern: wait for CI green, score risk (blast radius, reversibility, data/security, ops, verification gap), auto-approve low-risk, escalate medium-high to Slack. ChatPRD's bot (Merge Mommy) was built in about 30 minutes on Vercel Eve; repo rules still require a human click for SOC 2, so auto-approval is a signal plus a two-click Slack merge.

## Key points

- Thesis (2026): you can safely not review every PR if AI scoring is auditable and in policy. Intercom: 2-3x PR throughput; AI-approved PRs 5x faster; lower revert rate than human-only; labeled/traced for SOC 2/HIPAA. Counter to "AI code is less safe."
- Inspiration: Intercom how they made AI-approved PRs safe; Rewind Bot / Diff Vader (blast radius, correctness, checks).
- Why Eve (Vercel): simplest way for the speaker to deploy AI agents in Slack and GitHub without the torture of Hermes Agent or OpenClaw. Agent equals a directory of instructions, skills, code; sandbox; Vercel connectors wizard for Slack/GitHub (tokens/refresh). Built on Chat SDK (open-source multi-channel chatbot). Speaker is porting internal Slack agents to Eve when not already on OpenClaw.
- Build: one Codex prompt (internal GitHub bot that reviews PRs after checks are green, grades low/medium/high, auto-approves low risk) plus steer to Eve. Chrome browser-use hack for Slack app plus GitHub App permission screens (2FA/save clicks by human).
- Flow: GitHub App event after checks, Vercel GitHub channel, sandbox checkout, skills/tools score diff, comment plus gray check / request-changes, Slack ping. Instructions are about 4-5 paragraphs of markdown; scoring skill about 1 page. Thresholds (agent-chosen): under 24 low, 25-64 medium, 65+ high. Diff size does not set risk. Docs low, feature logic medium, auth/billing high.
- ChatPRD policy: bot cannot satisfy required human reviewers (gray check). Low-risk auto-approved PRs still need a human smash-approve in Slack. Merge conflicts block auto-approve even if docs-only (example score 6/10 blocked). Deprecation/API-behavior PR scored 45/100 medium, not auto-approved.
- Evals (Intercom bonus, not demoed): log every review; engineers grade whether the scorer was right, same eval discipline as customer-facing AI.
- Josh relevance: Eve as a third shape next to Grok Bot (managed computer) and Hermes (self-hosted OS) for managed Slack/GitHub agents if already on Vercel.

## Quotes / memorable lines

- "We figured out how to code with AI. Now, what do we do with all these PRs?"
- "Eve has become the simplest way for me to deploy AI agents in the enterprise in Slack and GitHub without the torture and pain of something like Hermes agent or my beloved OpenClaw."
- "We can put AI to work for us or we can have AI put us to work, and this is an example where you do both."

## Sources

- YouTube: https://www.youtube.com/watch?v=cmATJGbA8bI
- Raw ingest: `raw/youtube-inbox/cmATJGbA8bI/`
- Show: https://howiaipod.com
- Channel: How I AI https://www.youtube.com/@howiaipodcast
- WorkOS mid-roll sponsor (not a source for the bot architecture)

## Related pages

- [[hermes-agent]] (Claire's harder-than-Eve foil for Slack/GitHub bots)
- [[grok-bot]]
- [[grok-bot-vs-hermes-agent]]
- [[grok-bot-privacy-and-security]] (approvals / audit / SOC 2 framing)