Most Prompt Engineering Is Tribal Knowledge

Most "prompt engineering" inside companies is tribal knowledge. Somebody tweaks a prompt, it works better, nobody writes down why, and six weeks later nobody can tell you which version is actually running in production.

Mark Shropshire and Mike Wilbur's DrupalCon Chicago talk was about Tessera, a tool they built that turns client brain-dumps into properly-structured Jira tickets. But the part I keep chewing on is quieter: they keep their context files and prompt rules as markdown, in a repo, committed like code.

In plain English: they can diff their prompts. They can see exactly how the thing got better (or worse) over time. They can roll one back.

That's the difference between "we're iterating on our AI tool" and "we're engineering one." Most teams I talk to are doing the first and calling it the second.

If your team is "iterating" on an internal AI tool with no version history, you're one personnel change away from starting over.

If you're trying to get an internal AI tool past the pilot stage, watch this one. Mark's post is in the comments.

(Great talk Mark. This is the kind of thing more teams should be adopting.)