Whiteboard: review what your coding agent built, not just its diffs
A YC-backed launch heads a wave of open tools for the new bottleneck—reviewing AI code: audit your prompts, drive the UI it shipped, break the diff.

Copy markdown
Whiteboard: diagrams, not 2,000-line diffs
Open-source (MIT) desktop app from /dev/fast (YC W26): coding agents draw architecture, sequence and database diagrams of their changes next to a Rust-based, AST-aware semantic diff, with a Decision Log tracing the reasoning and every diagram linked to code. Local-first, telemetry can be disabled, installable today at install.dev.fast. It's built for the case that's now normal—reviewing a large agent PR by intent and structure instead of scrolling a wall of diff. Early users include teams at Salesforce and Modal.
Anthropic ships a prompt-audit you can run in Claude Code
The anthropics/skills repo added a /claude-api prompt-audit command that scans your prompts and skills for anti-patterns that hobble frontier models—verification rituals like 'double-check your work', overbearing MUST/NEVER scaffolding, and stale few-shot cruft—then proposes diffs without auto-applying. Point it at your own SKILL.md and system prompts to claw back tokens and reliability before you ship them.
assay: does the page your agent built actually work?
An MIT-licensed CLI (pip install assay-ui) that opens a real browser, measures the visible controls, drives them, and reports contradictions—no LLM calls and no hand-written tests, roughly a 14-second median run. Aim it at a freshly generated web page to catch buttons that lie before your users do.
Canary: 'AI writes your code. Canary breaks it.'
Canary independently boots the changed app and actively tries to break the diff, pulling context from GitHub, Linear, Sentry, Datadog and Notion to surface the bugs that slip past every existing test. It's the adversarial counterpart to Whiteboard's comprehension view: one shows you what changed, the other tries to make it fail.