Three Teams Adopted Agent Skills Files. Only One Changed How They Promote People

Three engineering teams adopted AI agent skills files (AGENTS.md, .cursor/rules, Copilot instructions) in the same quarter, but only one changed its promotion criteria afterward. The other two teams saw cleaner diffs and reduced review time but kept rewarding story points and visible features, leaving human judgment unmeasured. The third team, a marketplace with ~80 engineers, paired its skills file rollout with explicit promo rubric edits: agents own boilerplate, humans own verification and cross-service judgment, and senior promotion requires documented examples of catching agent errors that passed CI.

Every team I walked into this spring had a skills file. Or wanted one. Or had a platform engineer “almost done” with one. AGENTS.md. .cursor/rules. Copilot instruction files. Claude project context. Different filenames, same promise: teach the agent how we work so output stops embarrassing us in review. I watched three adoptions up close — same quarter, similar stacks Python/FastAPI, TypeScript monorepos, one Java shop , similar AI tool budgets. All three rolled out skills files within six weeks of each other. Only one changed promotion criteria afterward. The other two got cleaner diffs and the same broken promo rubric. That’s the story nobody posts on LinkedIn, because “we added AGENTS.md” sounds like progress. It often isn’t. Cursor’s rules files, Copilot instructions, Claude project docs — the tooling names change monthly. The organizational question doesn’t: does your company know what humans are still accountable for? Skills files make that question visible. Promo rubrics answer it. Most teams only shipped half the update. A fintech client mandated AGENTS.md after a security review flagged agent-generated code bypassing internal auth patterns. Platform team wrote a thorough file: naming conventions, forbidden imports, required test patterns, "never log PII" repeated three times because agents need repetition. Adoption metric: 100% of repos had the file within three weeks. Success, according to engineering leadership. What changed in behavior: PRs started looking more uniform. Fewer obvious lint failures. Agents stopped suggesting eval — small win. What didn’t change: promo packets still rewarded story points and visible features . Staff promo still required “technical leadership across multiple projects” with no definition of what that meant in an agent-assisted world. I interviewed three engineers on that team for a coaching engagement. All three could point to their repo’s AGENTS.md. None could explain how their personal judgment showed up anywhere leadership measured. The skills file taught the agent the rules. Nobody updated the rules for humans . Six months later, one senior engineer told me: “We have the best-documented agents on the block and the same people get promoted for the same reasons as 2023.” A B2B SaaS shop with a vocal AI champions group went further. Multiple skills files per service. Nested rules for database access, API versioning, frontend component patterns. Weekly “skills file office hours.” Genuinely impressive craftsmanship. Output quality improved measurably. Review time dropped ~18% by their internal estimate — I believe the direction if not the precision. Promotion outcomes: unchanged. Worse, actually — a quiet resentment built among engineers who didn’t enjoy maintaining prompt libraries. They shipped features; champions shipped .md files. Come review season, champions had more visible artifacts. Not because judgment improved. Because documentation volume did. The skills file became a second job for people who liked meta-work. Everyone else kept getting evaluated on feature throughput while agents ate the middle of the ticket queue. This is the trend piece angle everyone shares: skills files spread because they work for code quality. They’re silent on who gets rewarded for what remains human work. The third team — marketplace, ~80 engineers — did something boring and radical. They adopted AGENTS.md too. Similar content to Team One. But their eng director paired rollout with a promo rubric edit , explicitly documented in the same PR that added the root skills file: → Agents own boilerplate matching the skills file. → Humans own verification, failure mode documentation, and cross-service judgment calls. → Promo to senior requires two written examples of catching agent output that passed CI but failed a production criterion. → Staff promo requires authorship of a skills file change driven by an incident or near-miss — not aesthetic preference. That’s it. No new tools. No headcount. Just aligned incentives. Results after two review cycles still early, I’ll be honest : The skills file wasn’t wallpaper here. It was the contract between what agents do and what humans prove they still do. I’ve seen this pattern across other clients since — not just these three. Skills files optimize the wrong side of the equation by default. They constrain agent output. They don’t constrain human evaluation , which still defaults to proxies from 2019: lines merged, projects led, on-call hours survived. Champions skew visibility. The engineers who enjoy writing AGENTS.md get credit for meta-work that doesn't map cleanly to old rubrics. Everyone else looks slower because they're doing verification work that doesn't close tickets. Leadership treats adoption as the goal. “100% repos compliant” is an easy executive slide. “We rewrote what senior means” is a harder conversation with HR. The teams that win aren’t the ones with the longest skills file. They’re the ones where the file and the promo rubric tell the same story about what’s still worth paying a human for. Not the syntax rules — everyone has those. The section that changed behavior was titled “Human gate — required before merge” : Agents don’t read that and self-correct. Reviewers do. It became the social contract. One engineer described it as “the skills file finally gave me language for what I was already doing in review, and promo season stopped punishing me for it.” That’s the REACH insight hiding inside a tooling trend: AGENTS.md is an org chart argument written in Markdown. Either your promotion process agrees with it or you built expensive wallpaper. I’ve since seen a fourth team copy Team Three’s rubric language verbatim without copying the enforcement. HR approved the words. Managers still graded on velocity because it’s easier to count. The skills file got longer. Promotions looked identical to last year. Copy-paste the rubric or don’t bother — half-adoption is the worst outcome because it lets leadership claim “we’re AI-ready” while engineers still get punished for doing verification slowly. If you’re an IC and can’t change the rubric, change what you put in your promo packet anyway. Two verification stories with incident numbers beat sixteen lines about “accelerated delivery with AI assistance.” Managers who can’t use it this cycle might remember next cycle. The ones who won’t aren’t going to save you with a prettier AGENTS.md either. Skills files will get longer. Vendors will sell “enterprise AGENTS.md governance.” Someone will certify you in prompt context management. The teams that come out ahead won’t have the prettiest file. They’ll have the shortest gap between what the file says humans must do and what promotion packets actually reward. Three teams adopted the same trend. One changed who gets to staff. The other two taught their agents the house rules and kept paying humans for speed. Which team are you on? Check the rubric, not the repo. If you’re interviewing for roles where verification is the real filter, the round-by-round mistake list beats another framework cram session. → Top 50 Backend Interview Mistakes Want to stress-test judgment calls in a format closer to production than prompt trivia? Froquiz’s Senior Dev Challenge runs scenario questions across backend, SQL, and ops — useful before you walk into a loop that asks what you’d catch, not what you’d generate. → Froquiz Three Teams Adopted Agent Skills Files. Only One Changed How They Promote People https://blog.stackademic.com/three-teams-adopted-agent-skills-files-only-one-changed-how-they-promote-people-515116932151 was originally published in Stackademic https://blog.stackademic.com on Medium, where people are continuing the conversation by highlighting and responding to this story.