06-reference

technically sentry non engineers ship cms to git

2026-06-04·reference·source: Technically·by Justin Gage

"How my non-engineering team at Sentry learned to ship" — Justin Gage

Why this is in the vault

A first-person account of Sentry's marketing team moving ~2500 pages out of a CMS into Git + Markdown so coding agents could touch the whole site — the cleanest external proof yet that "the environment is the product" and that non-engineers ship once the environment is agent-shaped.

The core argument

The trigger was an asymmetry, not a cost line: a Claude Code session did a month of SEO link-fixing in an afternoon on the GitHub-resident parts of the site, then hit a wall at the CMS boundary (the blog and key pages the agent couldn't touch). That asymmetry — agent can update half the site instantly, half not at all — was the real pain, and it was compounding weekly as the team built more workflows around Claude Code. Velocity, not the Contentful bill, drove the migration (though the piece notes Cursor's CMS reportedly cost ~$57k/yr).

Execution: ~2500 pages migrated by 2.5 developers in a two-month window, most code written by Claude Code while engineers planned, reviewed, and directed. They migrated one directory at a time (docs/marketing first, then blog, then complex interactive pages), kept the CMS on a free tier for a soft cutover, and switched frameworks Gatsby→Astro mid-flight (build times 14min → under 6, then under 4 after a Marketo form cache). They consolidated ~200 bespoke pages into 3 templates — deliberately, because a simpler codebase is easier for agents to operate ("the agent doesn't have to learn the quirks of 200 one-off pages; it learns three templates").

The two reusable artifacts:

What was harder than expected (the honest section): design fidelity (pixel-matching old pages) took longer than the data migration — Claude is good at functional code, weaker at matching an existing design 1:1; a human-introduced form-validation rule tanked a conversion metric for days; build infrastructure was an under-appreciated failure source; and the team learning curve is real but smaller than expected — "I couldn't write code four months ago," and a handful of marketers now open PRs not because they became engineers but because the environment forced enough fluency to keep up. Next up: scheduled autonomous workflows (a Linear-ticket → Claude → PR loop already shipped for small well-scoped fixes) and AI code review to catch what human reviewers miss in 50-change batches.

Mapping against Ray Data Co

This is a direct, high-fidelity external validation of three RDCO positions at once.

Net: strong mapping. Reusable in positioning collateral and as phData-engagement framing. Worth lifting the "3 templates not 200" and "environment forced enough fluency to keep up" lines as positioning anchors.

Related