Build a prompt library, reuse your favourite prompts without copy-paste

You have favourite prompts. The one that condenses job postings. The one that drafts emails. The one that explains code diffs. Today they live in WhatsApp notes, in a browser tab, in your head. You type them every time.

A prompt library is not a tool, it is a habit. You put your prompts in one place, with three required fields, and you use them from there every time. After two weeks it is the first reflex before you send a new task to Claude.

This playbook shows 10 steps to build a library that works both in claude.ai (web) and in Claude Code (terminal), without premium tools. You need a note app that speaks Markdown (Obsidian, Notion, Apple Notes, Bear, even a folder with .md files is fine).

Step 1, one place for everything

First decision, where does the library live. For me it is a folder ~/notes/prompts/ with one Markdown file per prompt. Other people use a Notion database, others an Obsidian vault. Whichever tool, three rules, one place for all prompts, searchable in under 5 seconds, exportable as Markdown.

What does not work, spread across mail drafts, sticky notes and browser tabs. One source. Otherwise you search more than you use.

Step 2, three fields per entry

Every prompt entry needs three fields. Title ("Condense job posting"), use case ("when I find a posting and want to decide if it fits"), and the actual prompt text.

Four optional fields I added after three months, "Expected input" (what do I paste in), "Expected output" (what comes back), "Tonality" (casual, formal, technical), "Last updated" (date).

Do not invent further field bloat. If an entry has seven fields you never update it.

Step 3, categories, not tags

Three categories are enough at the start. For me, "writing", "analysing", "transforming". Every prompt lands in exactly one. If you constantly hesitate because a prompt fits two categories, the split is too fine. Reduce.

Tags are tempting but they go inconsistent. You do "Application" and next week "applications" and then you find nothing anymore. Three stable categories are better than 30 growing tags.

Step 4, the first batch is the hardest

Sit down for 30 minutes and think, which prompts do I write often. Look at your last 20 chats in claude.ai or ChatGPT. If you typed something similar there three times, it is a library candidate.

Realistic start target, 5 prompts. Not 30. You will test the first 5 next week and notice that three of them are not good yet. Only then you build further.

Step 5, sharpen a prompt

A library prompt is not the same as a throwaway prompt. It must work without context because tomorrow you paste it into an empty chat.

Four sharp edges make a prompt sharp. One, clear what the input is ("I give you a job posting as text"). Two, clear what the output should be ("List with 5 points, 1 sentence each, score 1-10 at the end"). Three, clear what is not (no cover letter, no greeting). Four, clear who you are ("I am a candidate for junior frontend roles").

Test, if you paste the prompt into an empty browser tab tomorrow morning, does Claude know what to do. If not, sharpen.

Step 6, variables with placeholders

Instead of typing your role, your project, the recipient new every time, use placeholders. I use curly braces, {ROLE}, {PROJECT}, {RECIPIENT}. When pasting, you replace them.

Example prompt for client emails:

You are \{ROLE\} and you write to \{RECIPIENT\} about \{PROJECT\}. Tone is \{TONE\}. Goal of the mail is \{GOAL\}. Write in Du form, short, no marketing speak.

You paste, you replace 5 braces, the prompt is ready. Faster than writing it new every time.

Step 7, work out the tonality block once

If you enforce a tonality across multiple prompts (Du form, short, no fluff), write a tonality block once as a separate note and paste it below each prompt that needs it.

For me tonality-default.md says, "Write in Du form. Keep sentences short. Avoid in today's fast-paced, at its core, let's dive in. Em-dashes off." When I use a writing prompt, I paste this block below. Change tonality, one file, all prompts benefit.

Step 8, slash commands for power users

If you use Claude Code, you can store your most important prompts as slash commands. Drop a Markdown file in ~/.claude/commands/ with filename condense-job-posting.md. Content is your prompt text. In the terminal chat you then type /condense-job-posting and the prompt is loaded.

Advantage, the prompt is always there and you can maintain it per project. Disadvantage, this only works in Claude Code, not in claude.ai. For claude.ai you stay on copy-paste from the library.

Whoever wants it even cleaner, packs a collection of slash commands into a plugin and shares it with colleagues, see the playbook Your first Claude Code plugin in 60 minutes.

Step 9, monthly cleanup

A library becomes trash when it is not maintained. Schedule a 15-minute routine every 4 weeks, walk through all entries, tag with "keep", "sharpen" or "delete".

Rules of thumb, if you have not used a prompt for 4 weeks and you see no concrete reason for next week, out with it. If you used it 4 times but the result was not good twice, sharpen it (step 5). If it runs solid, leave it alone.

A library with 25 sharp prompts is more useful than one with 80 dusty ones.

Step 10, keep versioning light

You will change prompts. Sometimes you find that the old version was better. Safety net, put the folder under Git. A git init in the notes folder, a git commit -am "prompt updates" once a week, and you can reconstruct every older version.

If you do not know Git, see L1 Git for AI Lesson 1. Alternatively suffix each prompt file with _v1, _v2 when your changes are drastic. But Git is cleaner.

What's next

When the library stands, the next lever is to combine it with project knowledge in claude.ai, see the playbook Claude Projects for non-coders. Whoever wants to maintain technical prompts in slash commands goes on to the lesson L2 System Prompts and to the playbook Your own slash commands for Claude Code. Whoever wants to build their own skill out of three recurring prompts, looks at First own skill in 30 minutes.

Official docs on custom slash commands in Claude Code, https://docs.claude.com/en/docs/claude-code/slash-commands.