Balane TextDeck

How to save and reuse ChatGPT prompts

Four steps from chat history to a personal prompt library you can reach in three seconds, plus how to name, sort and maintain it once it holds thirty prompts.

Updated October 20, 20268 min read2 prompts to copy

If you use ChatGPT every day, your best prompts are scattered across chat history, a notes app and a few screenshots. Each one took ten minutes to get right, and next time you retype most of them from memory. This guide turns that into a small system: four steps that move a working prompt out of the chat and into a library you can reach from any app, then a section on organising that library and one on keeping it lean. Everything here works the same with Claude, Gemini or any other model, because the library lives outside the chat.

Why chat history makes a poor prompt library

Chat history is a conversation log. Search finds words in the replies as readily as in your prompts, there are no names or categories, and the prompt that finally worked is usually message four of a thread, after three corrections. Projects and custom instructions help with context, yet they stay inside one account on one service.

The test for any place you keep prompts is retrieval: can you get a prompt you saved a month ago into the text field in under five seconds? If the answer involves scrolling, opening a Notion page or remembering a thread title, the prompt is as good as lost. The four steps below are built around that five-second test.

Step 1: Extract the prompt the moment it works

Copy the prompt out of the chat as soon as an answer is good enough to use. Later is where prompts disappear: by Friday you will not know which of forty conversations held it.

Strip the model's replies. Then read your own follow-up messages, because they contain the real prompt. If you had to write "shorter", "no bullet points" and "keep the numbers" before the answer worked, fold those three corrections into the original prompt. The version you save should get the good answer on the first try.

To seed a library quickly, scan the last two weeks of history once and pull out every prompt whose output you actually used. Daily users usually find ten to fifteen. That is your starting set, and none of it has to be written from scratch.

Step 2: Turn the moving parts into variables

Read the extracted prompt and mark every word you would change next time: the topic, the audience, the length, the pasted text. Those become variables in double braces. Everything else is scaffolding you wrote once.

Three to five variables per prompt is the useful range. With two or fewer the template only fits the original case; with seven or more, filling it in takes longer than retyping. Reuse the same names across templates ({{topic}}, {{audience}}, {{tone}}, {{length}}, {{input}}) so you never have to think about what a field means.

Saved as a template

Write a {{word_count}}-word blog intro about {{topic}} for {{audience}}. Tone: {{tone}}, no jargon. End with a question the reader would ask themselves.

Saved as typed

Write a 150-word blog intro about switching from spreadsheets to a CRM for owners of small agencies. Friendly, no jargon, end with a question.

Four variables, one fixed rule. The closing question stayed in the scaffolding because it worked every time; only the parts that changed became fields.

Step 3: Name it and store it outside the chat

Give the template a name that describes the job, and add the variables when two templates are close: "Blog intro (topic, audience, length)" is findable in a month, "Content prompt 3" is not. Five seconds spent on the name pay back every time you search.

Store it somewhere that belongs to no single AI service. When you move a task from ChatGPT to Claude, or a new model changes how projects work, the library should not care. A plain text file, a notes app or a Notion database all clear that bar. What they lack is speed and a way to fill in the variables.

TextDeck is built for this step. You create a template, paste the prompt, mark the variables and pick a folder. It is free, needs no account and stores everything on your device, with optional sync through iCloud on Apple devices or a cloud folder of your choice on Android and Windows. It also ships 135 ready-made templates in 14 packs, so the library is not empty on day one.

Step 4: Insert it with a hotkey

A library you have to open and browse is one you stop using within a week. The bar is three seconds from "I need that prompt" to finished text in the field, from whatever app you are in.

In TextDeck that is ⌃⇧P on the Mac and Ctrl+Alt+P on Windows: pick the template, fill in the fields, insert the result. On iPhone and iPad the same recall runs through Shortcuts, Spotlight and the Action Button; on Android, TextDeck sits in the text-selection menu of every app.

Variable history does the rest. TextDeck remembers the last value of each field, so a prompt you run daily with the same audience and tone only asks for the one thing that changed.

Worked example: meeting notes into decisions

The starting point is a prompt that took three rounds. The first answer summarised everything, the second dropped the open questions, and only the third separated decisions from discussion. Folded together, those corrections become one template.

It has two fields, {{meeting}} and {{notes}}. The rest never changes, which is why a template like this tends to end up among the five you use most.

Meeting notes to decisions

You are summarising notes for people who were not in the meeting. Meeting: {{meeting}}. Return exactly three parts: 1. Outcome: one sentence. 2. Decisions: a bullet list of what was actually decided. 3. Action items: a table with the columns Owner | Task | Due date. Anything discussed but not decided goes under a final heading called "Open". If an owner or a date is missing, write [MISSING] instead of guessing. Notes: {{notes}}

Try it: fill the variables

Worked example: a client follow-up email

This one started as a single message to one client after a call. Three variables cover every later use: who it goes to, what was agreed, and what happens next. A fourth, {{tone}}, stays pre-filled with "warm, brief" from variable history and rarely changes.

The two rules at the end are worth copying into other templates. They remove the opener and the invented deadline you would otherwise delete from every draft.

Client follow-up

Write a follow-up email to {{client_name}} after today's call. What we agreed: {{agreed_points}} Next step: {{next_step}} Tone: {{tone}} Keep it under 120 words and open with the most important agreed point. Do not start with "I hope this email finds you well". Do not promise any date that is not in the agreed points.

Try it: fill the variables

Name templates like a search query

Once the library passes twenty entries, the names do most of the finding. Lead with the job: "Summarise meeting notes", "Rewrite for a sceptical reader", "Review code for edge cases". Version numbers and dates have no place in the name.

For clusters of similar templates, a short prefix works better than a deeper folder tree: "email: cold opener", "email: follow-up", "email: decline politely". Type "email" into the search and the whole family comes back together.

Sort folders by purpose

Six folders cover most libraries: Communication, Analysis, Writing, Development, Research and Meta (prompts that improve other prompts). Every template goes into exactly one. If a seventh folder starts to feel necessary, merge two existing ones first; an extra folder costs more hesitation when saving than an imperfect fit costs when searching.

Skip folders named after models. "ChatGPT prompts" and "Claude prompts" turn one summary template into two copies the day you switch, and templates that only work in one model are rare. Where a prompt really is model-specific, for example one that leans on XML tags for Claude, put that in its name.

Then give your top five their own spot at the top, in a folder or collection of their own. In most libraries five templates carry the bulk of daily use: a summary, a rewrite, an extraction, a review and one specific to your job. Everything else can wait for search.

Maintenance: fix in place, prune monthly

Every template starts out mediocre. When an answer annoys you, fix the template right then, in the library, and the next use gets the better version. This is the real advantage over chat history: a prompt you run twenty times a month gets twenty chances to improve.

Once a month, spend ten minutes on two questions. Which templates have I not used in 60 days? Archive or delete them. Which of my top five produced an answer I had to edit? Tighten that one instruction. A library of 15 to 40 templates can be scanned at a glance; beyond that you are collecting, and the scrolling problem from chat history returns.

Turn anything that drifts into a variable. Dates, product names and word limits change more often than the task itself, and as fields you change them once while filling in instead of in ten templates.

Your first week

Day 1: pull ten prompts from the last two weeks of ChatGPT history, fold in the corrections, mark the variables and save each one under a name that describes the job.

Days 2 to 6: whenever you type a prompt from memory that you have typed before, save it. Whenever an answer needs editing, fix the template.

Day 7: delete the templates you never opened, rename the ones you could not find quickly, and move the five you used most to the top. From then on, the monthly check is enough.

Next guide

Free AI prompt library: 15 starter templates

Most "free" prompt libraries are lead magnets. This one is the real thing: 15 templates you can use today without signing up for anything.

Three keys. Every template.