We Fixed an Embarrassing Problem We Kept Seeing in AI-Heavy Teams

Last quarter we lost a frontend engineer on our team. Talented guy. The code handoff went fine — PRs merged, docs updated, the usual checklist. Two weeks later, our tech lead pinged the team Slack:

"He had this prompt + Skill combo that ran Claude Code reviews in like 30 seconds. Does anyone have the config?"

Crickets.

We spent a full week reverse-engineering it from scattered chat logs. Every tweak, every iteration — gone, because it lived in his personal Claude account. That week cost us more than the AI subscription ever did.

And here's the thing: we're not some disorganized startup. We have Git workflows, code review checklists, documented architecture decisions. But AI-era production knowledge — prompts, Skills, trained Memory — sat completely outside our asset management. We were paying the AI bill. We weren't inheriting the output.

The real tab isn't the subscription fee

We started tracking this across the team. A marketing colleague had a prompt that cut competitive analysis from three hours to twenty minutes. It lived in her chat history. A backend engineer had a Skill workflow that auto-reviewed PRs with structured feedback. The config was buried in a local JSON file, zero version control. Our product manager had spent months feeding the AI team naming conventions, API styles, and user persona context — hundreds of conversations distilled into Memory. If she left tomorrow, the replacement would start from zero.

These aren't minor conveniences. They're production assets. And nobody was cataloging them.

The math is straightforward. A 30-person engineering team easily burns $1,000+ a month on AI subscriptions and API calls. Over a year, that's five figures. How much of the reusable output stays inside the company? Our honest answer was: almost none.

What we built

We didn't want to build another AI tool. There are enough of those. We wanted a governance layer — something that sits between your team and whichever AI providers they use, and does two things:

Cost attribution. Real-time tracking of who's calling which model for what. If someone's burning a top-tier model on basic summarization, you see it before the month-end invoice hits. Broken down by project, by department, by model. FinOps for AI, basically.

Asset cataloging. Prompts, Skills, and Memory get automatically recognized when committed, tied to projects, and versioned. A marketer's tested prompt shouldn't require her to manually share it. The system knows it's a reusable asset and drops it into the company directory with the right permissions. People leave — the knowledge doesn't.

Dogfooding numbers

We've been running this internally for about two months. Here's what happened:

• Our internal prompt library went from zero to 47 reusable assets
• The marketing team's competitive analysis prompt alone saves roughly 2.5 hours per report — and is now searchable by anyone on the team
• We caught two cases of GPU-tier models being used for basic text tasks and rerouted them to cheaper alternatives, saving about $200/month with zero impact on output quality
• Three prompts that had been iterated independently by different team members turned out to be 80% the same. Merged into one, maintained in one place

The hardest part wasn't the tech. It was getting people to stop treating their prompts like personal notebooks and start treating them like code — committable, reviewable, reusable.

What we're still figuring out

Honest admission: the asset cataloging piece works well for structured things like Skills and prompts committed through the platform. The long-tail stuff — the impromptu one-off prompts, the Memory that accumulates organically over casual conversations — that's harder to catch. We're iterating on discovery, trying to surface valuable assets without creating noise.

The other open question is governance. Once you have a shared prompt library, someone needs to maintain it. Who reviews submissions? Who decides when a prompt is deprecated? This is less a product problem and more an organizational one, and we're feeling our way through it.

If any of this resonates — if you've lost good prompts to departing teammates or discovered your AI spend was 3x what you thought — we'd love to have more teams kicking the tires. You can check out what we're building at https://aikeylabs.com/zh/i/ih12.

Enterprise inquiries: [email protected]