Stop Optimizing Your Prompts. Start Collecting Them.

A prompt library is more valuable than a perfect prompt. The case for collecting, not optimizing, with a starter library of 5 prompts that pay for themselves in a week.

The optimization trapMost prompt advice is about optimization. "Add this phrase to your prompt." "Try this prefix." "Use this temperature." The advice is well-meaning, and some of it works, and almost all of it has the same problem: it treats the prompt as a one-off artifact, not as a piece of infrastructure.The optimization frame is wrong. A prompt you use once is an experiment. A prompt you use fifty times is a tool. The tool is more valuable than the experiment, because the tool compounds. The compound interest on a tool is the time you save the fifty-first, fifty-second, and fifty-third time you use it. A perfect one-off prompt gives you a great answer once. A good repeatable prompt gives you a great answer forever.The optimization trap is the belief that the next prompt tweak is the one that will make the difference. The library is the thing that makes the difference. The library is the system that does not have to be re-built every time.What a prompt library looks likeA prompt library is a small collection of prompts that you have used enough times to trust, organized by task. Mine is a folder in Notion with one sub-folder per task. Each sub-folder has the prompt, a one-paragraph description, an example of the output, and a one-line "last tested" date. The library is not fancy. The library is 14 folders, 14 prompts, about 4,000 words total.For each prompt, I track:The prompt itself. Copy-paste ready, including any system-prompt preambles, examples, or constraints.What the prompt does. One sentence, plain English, in my own words. ("Summarizes a long article into 5 bullet points and a 2-sentence summary.")The model it works on. Most of mine work on Claude and ChatGPT. A few are model-specific. I note the model because prompts do not always transfer cleanly across models, and the note saves me from re-discovering that.The output shape. A real example from a real use, with the date. The example is the part that tells me whether the prompt is still doing what I want it to do, or whether the model has shifted enough that the prompt needs to be re-tuned.That is the structure. Four fields per prompt. The structure is small enough to maintain (about 1 minute per prompt per year) and large enough to actually be useful.The 5 prompts that pay for themselves in a weekIf you are starting a library from scratch, these are the 5 prompts I would start with. Each one is a real prompt I use weekly, and each one will save you more time in a week than the 30 minutes it takes to set it up.The 5-bullet summarizer. "Summarize the following article into 5 bullet points (each one line) and a 2-sentence summary. Plain English. No filler. If the article does not make a clear point, say so." Use: every time you read a long article, a documentation page, or a transcript. The 5-bullet summary is the fastest way to know whether the article is worth a full read.The 3-field fix. (From [the 3-field post.) "GOAL: [what you want]. AUDIENCE: [who is going to read this]. FORMAT: [what shape the output should take]. [Optional: CONSTRAINTS, TONE, EXAMPLES.]" Use: every time you would normally write a prose prompt. The 3-field fix is the smallest structure that consistently produces useful output.The email triage. "Categorize the following email as ACT-NOW (action needed within 24 hours), ACT-LATER (action needed within 2 weeks), WAIT (waiting for someone else), or DONE-NEVER (no action needed, archive or trash). Output: just the label, nothing else. Email: [paste]." Use: every email you triage. The label is the decision; the decision is the bottleneck.The code reviewer. "Review the following code for bugs, edge cases, and security issues. Output: a numbered list of issues, each with (a) the line number, (b) the issue, (c) the suggested fix. If the code is fine, say so. Code: [paste]." Use: every PR you review, every script you write. The numbered list is the output; the list is the audit.The "is this worth doing" prompt. "I am considering [X]. Give me 3 reasons it is worth doing and 3 reasons it is not. Output: two bulleted lists, then a 1-paragraph honest summary. Do not soften either list. No 'it depends.' No 'let me know if you have questions.'" Use: any time you have a small decision to make. The 3-and-3 prompt is the cheapest way to get past the bias of your own initial enthusiasm.That is the starter library. 5 prompts, each one used weekly, each one saved in a folder with a one-paragraph description. Total time to set up: about 30 minutes. Time saved in the first week: about 90 minutes. Net win in week 1: 60 minutes. Net win in week 52: about 78 hours.What the library is notIt is not a Notion template you bought. It is not a course. It is not a 200-prompt mega-pack you will never use. The library is the 5-20 prompts that you have actually used enough to trust. The number is small on purpose. A library of 200 prompts is a library you do not use. A library of 14 prompts is a library you reach for daily.It is also not a one-time artifact. The library changes as the models change, as your work changes, as the prompts you used 6 months ago stop being the prompts you need today. I re-tune a prompt about once a quarter, on average. I add a new prompt about once a month. I delete a prompt about once every 6 months, when I realize I have not used it in a year. The library is alive, not static.How to maintain itOnce a quarter (I do it on the first Sunday of the quarter, but any day works), spend 30 minutes on the library. Walk through each prompt. Re-test it on a real use. If it still works, mark it as re-tested. If it does not, re-tune it. If you have not used a prompt in 6 months and cannot think of a use for it, delete it.The 30-minute quarterly maintenance is the part that keeps the library useful. Without the maintenance, the library decays. Prompts go stale as models change. Prompts that you have not used in 6 months are dead weight. The maintenance is the part that converts "I have a folder of prompts" into "I have a tool I use every day."What the prompt pack on this site isThe Blog Writing Factory (linked at the bottom of most posts) is a starter library for writing. 10 prompts, one ZIP, the ones I actually use to write the posts on this site. If you are starting a prompt library and writing is one of the things you do, the pack is a reasonable place to start. It is not a complete library. It is the writing folder. You will still need the email folder, the code folder, the research folder, the rest. But it is the writing folder, and that is the most leveraged folder for most of the people who read this site.The honest partI built my prompt library over the course of about 2 years. The first 6 months, I had a Notion page with 50 prompts I had copy-pasted from Twitter. None of them worked as advertised. The next 6 months, I deleted the 50 and built 8 of my own. Those 8 worked. The 6 months after that, I added 6 more, deleted 2, and re-tuned 4. The current 14 are the survivors of that 18-month process.The library is not impressive. It is small, simple, and obvious. It is also the single most useful piece of AI-related infrastructure I have built, more valuable than any model subscription, more valuable than any single prompt, more valuable than any tool I have tried. The reason is the same as the reason a code library is more useful than a clever function: the library is reusable, the clever function is one-off. The optimization is one-off. The collection is reusable.If you have been spending time optimizing individual prompts, stop. Spend the time on a library instead. The library will pay you back forever. The optimization pays you back once.Want a starter library for writing?The Blog Writing Factory is a 10-prompt starter library for the writing folder. It is not a 200-prompt mega-pack. It is the 10 prompts I actually use, in the order I use them. Yours for $14, no subscription, updates free when I ship new ones.Buy on Gumroad. $14

Send me the rough edges if you try it. I read every message.