Diff, the code review Grok Bot
Findings ranked by blast radius
Repro, the bug report Grok Bot
Reproduces it, writes the report, stops
Query, the SQL Grok Bot
A query that runs, with the assumption named
Probe, the QA Grok Bot
The paths a human would miss, as a checklist that can fail
Page, the incident comms Grok Bot
What broke, who was hit, what you are doing
Grok Bots for engineering: reviewing a diff ranked by what breaks in production, turning a vague complaint into a reproduction an engineer can pick up cold, writing the test plan for the cases people skip, writing SQL that runs with its assumptions named, and writing what customers read during an outage.
The thing that makes a coding Grok Bot useful is not that it can read code. It is the order it reads in. Diff ranks by blast radius, from data loss and security down to naming, so the output is something you act on top down rather than a flat list where a style note sits beside a race condition. Probe writes the happy path in one line and spends the rest of the plan on boundaries, interruption, concurrency and permissions.
The second thing is evidence. Every Diff finding carries the concrete input that fails and the fix as code. Every Probe check states what passing and failing look like, because a check that cannot fail is not a check. Repro reproduces a bug before writing a word about it, and narrows to the smallest case that still breaks, which is often half the debugging done before an engineer opens the file.
Diff for when the change touches money, auth, migrations or data
Repro for when somebody said it is broken and nobody can reproduce it
Probe for when you are about to release and want the cases people skip
Query for when the answer is in the database and being confidently wrong is the risk
Page for when something is down and customers are reading a status page
The ranking. Diff orders what it finds by blast radius, from data loss and security down to naming, and every finding carries the concrete input that fails and the fix as code. A general chat returns a flat list where a style note sits next to a race condition.
No, and it is told not to. It comments on an approach only when the approach causes data loss, a security hole, a correctness bug or a failure it cannot recover from.
Query asks for the schema, and where it has to guess it puts the guesses in a comment block at the top. It is read only by default and never runs a write.
The cases that actually break software: boundaries, interruption, concurrency, permissions, ugly text and the second device. It writes the happy path in one line and moves on.
These Bots review and test what you build, and the design templates for the tool you build it in sit alongside them: Lovable templates, Bolt templates, Replit templates, Claude Code templates, ChatGPT Codex templates, Cursor templates, Windsurf templates, Framer templates, v0 templates and Cline templates.