Stop starting from scratch: saved prompts, Projects, and custom assistants

There is a moment in every AI-using team worth catching. Someone writes a genuinely good prompt, one that nails the company tone and gets the details right, uses it once, and lets it scroll away into chat history. Next week a colleague reinvents a worse version of the same prompt from zero.

That scroll-away moment is the difference between a team that uses AI and a team that is building a capability. The fix is unglamorous: treat your best prompts like the business assets they are. Name them, keep them, share them, improve them. Teams that make this shift are the fastest-growing group of serious AI users, and the mechanics take an afternoon.

Who this fits

Your team uses AI regularly and it is already saving real time, but everyone works from a blank chat window and the quality depends entirely on who is typing. If AI use is still one person's habit, start with the team article first.

First: the prompt library (one afternoon)

1. Ask everyone who uses AI to paste their three most-used prompts into one shared doc. No cleanup, no judging.
2. Name each one after the job it does: "Overdue invoice reminder, firm but warm." "Board update from bullet points." Names matter; nobody reuses "Untitled prompt 7."
3. As a group, rewrite the top five. Add what a stranger would need: your tone in a line, your policies, one example of a great output.
4. Put the library where work already happens: the same drive or Notion space as your procedures, not a new tool nobody opens.
5. Add a monthly fifteen-minute review. Drop what nobody used. Add what someone built. A library that never changes is a graveyard.

The immediate payoff is consistency: your best AI user just became everyone's baseline.

Second: promote your top jobs to assistants

A saved prompt still makes you paste the background every time. For your highest-volume jobs, take the next step: a Claude Project or a custom GPT, which is a workspace that permanently holds the instructions and reference files for one job.

Pick the job your team does most with AI. Then:

1. Gather what it always needs: your policies, two or three gold-standard examples, tone notes, the boilerplate.
2. Create a Project (Claude) or a custom GPT (ChatGPT) and load those in as instructions and files. Either works well; use whichever tool your team already lives in. Nobody pays us to say so.
3. Have two different people run their next real task through it and write down what it gets wrong.
4. Fix the instructions, then share it with everyone who does that job.

Now "write the proposal" is one sentence and an attachment, because the assistant already knows your business. Teams that work this way route a large share of their AI work through these standing assistants, and it is easy to see why: the setup cost is paid once instead of every morning.

Third: pair prompts with procedures

If you have written procedures, staple the two together: every SOP that involves writing or summarizing gets a link to the prompt or assistant that does it. New hires then inherit both the how and the shortcut on day one. If you do not have written procedures, this is a good excuse; AI is uncommonly good at turning a rambling voice memo of "how I do this" into a clean procedure doc.

What to expect

The honest gain at this level is not more hours saved on any single task. It is that quality stops depending on who typed the prompt, onboarding gets faster, and the time you already saved becomes durable instead of vanishing when one person leaves. Most teams feel the difference within two or three weeks, mostly in fewer "how did you get it to do that?" conversations.

The next level, when you are ready

Everything at this level still starts with a person opening a chat window and pasting things in. The next level removes both: workflows where AI runs as a built-in step on your actual business data, with a human checking a queue instead of kicking off every task.

The gate to that level is not a smarter tool. It is whether your information is organized enough to connect: records in known places, tools that talk to each other, a source of truth worth pointing AI at. If that sounds far off, that is normal, and it is exactly the gap the next article in this series covers.

This week: collect the three-prompts-per-person doc, and build your first assistant for the job your team repeats most.