You've probably had this experience: you spend an evening teaching ChatGPT or Claude how your business works — your pricing, your tone, how you want customer emails handled — and it goes well. Then the next morning you open a fresh chat, and you're back to zero. Re-explaining everything. Re-pasting the instructions. The knowledge didn't survive the conversation.
Most small business owners treat this as an AI limitation to work around. Figma's product-management team treats it as a filing problem — and their fix is worth copying in a much smaller version.
The problem your chat window can't solve
Chat context is disposable by design. Close the tab and it's gone. Open a new thread and the AI starts blank. That's fine for one-off questions. It's terrible for operating knowledge — the stuff that should never get forgotten: how you quote jobs, what you promise in writing, which customers get special terms.
So people do the obvious thing: they keep pasting their instructions into every new chat. Some keep a "master prompt" document open on a second monitor. The knowledge exists, but it lives in the worst possible place — a place that resets every time you need to use it.
What Figma's PM team actually did
Ezra Mechaber, a growth product manager at Figma, described the setup on the How I AI podcast: an internal system the team calls a "PM OS" — not a product, just a repo. It holds about 31 documented skills covering the recurring work of being a PM: writing documents with the right reviewers looped in, briefing on open decisions, building small dashboards, even a first pass at coaching conversations.
His goal, in his words, was to be "lazy in the right places" — help with jobs he did poorly or forgot entirely, like looping support into a launch or finding who owns a piece of the product. Podcast host Claire Vo's framing is worth keeping: AI should "raise the floor," not just "take off the ceiling." The glamorous demos get the attention, but remembering to involve support removes a predictable failure from the operating system.
Two details matter more than the scale:
- The skills live in a repo, not in anyone's chat window. Shared parent context, private workspace per person. When one PM improves a skill, everyone gets the improvement.
- They prune it. Figma's PMs were invited to "roast" the skills at an offsite; the roasting decided which ones were worth keeping. Mechaber has been removing instructions as models improve. The repo is maintained operating context, not a permanent monument to prompt engineering.
Why "skill" is the right word
A skill is not a prompt. A prompt is what you say once. A skill is documented know-how that says when it applies and what it does. Four parts:
- When to activate. "When a new review appears on my Google Business Profile" — not "whenever."
- What to do. The exact steps: check the rating, classify the complaint, draft a response.
- How to sound. Friendly, plain, never defensive — your tone, written down once.
- What to reference. Your pricing sheet, your service area, your refund policy, so drafts cite real facts.
If you can't write those four lines for a task, you don't have a skill yet. You have a vibe — and vibes are exactly what dies in chat threads.
The five skills worth writing down first
A small business doesn't need 31. Five files cover most of the recurring operating knowledge in a service business:
- The morning briefing. Open estimates, unanswered reviews, overdue invoices, today's appointments. Mechaber's most-used skill is exactly this — he runs a "What's Up" briefing about six times a day. One detail worth stealing from their instructions: the briefing reads whole threads, not isolated messages, because a briefing is only useful if it can tell a new request from one that was resolved ten messages later.
- Customer replies. How you sound, what you promise, what you never put in writing — plus your FAQ: hours, service area, lead times.
- Quote and estimate follow-ups. Your rules for chasing: how many touches, over how long, when to let go.
- The weekly review. What shipped, what stalled, what the numbers say — run from your records, not from memory.
- "How we do things here." The employee-handbook skill: escalation rules, who can promise what, the standard a draft must meet before it goes out.
Five files. An evening of writing, once.
The repo part matters more than any single skill
Where the files live is what makes this work:
- One shared location, not five chat accounts. A single folder the whole team can see — a cloud drive, a repo, even a well-kept notes app. If your instructions are scattered across three chat threads and a text file on your laptop, they are already lost.
- One owner per skill. When someone spots a bad instruction, they fix it at the source — and every future use inherits the fix. Feedback through conversation, not a suggestion box that never gets opened.
- Pruning is the maintenance. A skill that keeps producing output you have to fix is worse than no skill. Figma's roasting session is the right instinct at any size: use, edit, delete what doesn't earn its place.
- The AI changes; the files transfer. Figma's team started in one coding tool and now mostly works in a different one. The durable investment was the context and skills, not loyalty to an interface. Whatever model you use next year, the skills come with you.
How to hand a skill to an AI agent
Written files only pay off when you point the AI at them. Two sentences is usually enough: "Before you draft customer replies, read replies.md and follow it exactly. Ask me about anything it doesn't cover."
Then let the skill do the delegation. "Handle new Google reviews per reviews.md — draft only, show me before anything posts" delegates the whole routine, not one message at a time. The same move works for a person: a new hire who reads "how we do things" on day one asks better questions by day three.
Two honest boundaries
Size isn't the point. Figma has roughly 60–70 PMs using this. A solo operator with five files gets the same benefit — the mechanism (documented, referenced, edited) doesn't require a team.
None of this is measured. Mechaber says he tried to build a benchmark and couldn't; he doesn't know how much the setup outperforms a vanilla agent. The honest case is operational: documented instructions beat instructions you keep re-pasting — because you stop losing them, and because your team inherits them. If that isn't worth an evening of writing, no benchmark will convince you.
The 30-minute version tonight
- Pick your highest-friction recurring task — the one where the AI's output usually needs fixing.
- Write the four parts: when it applies, what to do, how you sound, what to reference.
- Save it as a file in one shared place. Name it clearly.
- Point your AI at it next time the task comes up: "Read [the file] and follow it exactly."
- When the output disappoints, fix the file — not the chat. That edit is the whole system compounding.
Want a head start? The Boring Automation Pack from SMB Scale Up includes plug-and-play automation blueprints plus the standing-instructions worksheet — the boring half of this setup, pre-built.
Top comments (0)