Ask HN: How do you manage your prompts in ChatGPT?

[nabi_nafio]

I use ChatGPT regularly for a lot of different tasks. For example, coding, health Q&A, and summarizing docs. The different prompts stack up in the sidebar which becomes very difficult to manage. For example, I frequently have to refer back to a prompt that I wrote previously. But I usually give up looking for it because of the tedious scroll and search process. I was wondering if there is an easier way. How do you manage your prompts in ChatGPT?

[andix]

This is actually one of my biggest issues with ChatGPT, that it's not really possible to create some kind of reusable workflows. The best option is Custom GPTs, to create a specific chat bot for one task.

There are many UI projects for LLMs, like openwebui.com for example. But even with the OpenAI API as backend they don't provide as many features as ChatGPT (Web search, Python processing of data, charting, image generation).

I think one of the most promising approach would be some kind of user scripts for extending the official ChatGPT UI. (user scripts in the browser with some tool like Violentmonkey, FireMonkey, or anything similar to the good old Greasemonkey). I don't use it though, and I don't know if there are any good extensions for ChatGPT.

[lulzury → andix]

I looked into creating a simple Firefox extension for ChatGPT which would let me branch a conversation off into a new chat so as not to "pollute" or interfere with the existing chat.

I had success extracting the existing conversation and adding it to a new window, but gave up after trying to modify the ChatGPT UI (i.e. to format what I'd just pasted in so it'd look like the conversation left off where I branched). The UI just kept re-rendering/re-painting itself non-stop, overriding what I changed. I didn't try to push past that. I'm sure I could use JS or something else to massage the UI further, but it didn't seem like a non-trivial task. Maybe something to look into some weekend.

https://github.com/lulzury/got-branch-convo

[fragmede → lulzury]

click the share chat button and open that url to clone a conversation

[H8crilA]

Why not just ask the bot to do something? I'm using it daily and don't have to spend more than 5 seconds thinking of a "prompt".

The only exception is function calling (or whatever they call structured output these days), but that is simply embedded in my or other people's programs that call the API.

[t-writescode → H8crilA]

This is what I've found effective.

I use conversational English for basically every prompt I work with ChatGPT on as a regular person.

For my application, I have prompts that I have stored in source code, but those need to have very consistent, exact inputs and outputs (mostly JSON), so creating a specific prompt is important for those.

For anything human where a human can parse it, regular ChatGPT works perfectly fine!

[adastra22 → t-writescode]

I think you guys are responding to the title, not the question. OP is wondering how to organize the history of chats.

[t-writescode → adastra22]

This is true, and you're right. I think the perspective we're trying to walk from is: "if you're creating custom queries just to find common things from ChatGPT - that is, you're not creating an assistant and you're not using it for a coding backend (which have their own solutions), then we don't recommend going that route and going a different route.

What I've said does smell of the Super Smart(tm) and completely unhelpful Stack Overflow response, I will give you that.

[chaosist → adastra22]

I made a private blog that if something seems really valuable, I blog the prompt and output. Private so I don't waste time on the delusion of someone else reading it and everything that comes with that time wise.

Beyond that, the previous chats could practically be deleted for me. Its not like my prompts are such works of art that I couldn't recreate the thought if needed. Not to mention, Sonnet has made me not use chatGPT for about a month now so the blog is nicely divorced from the LLM flavor.

[afpx → H8crilA]

I asked it for a solution and it told me to export the data and use a search indexer like typesense / InstantSearch.

[viraptor → H8crilA]

It depends how often you do a thing. It's ok to write things from scratch a few times. But if you often ask about the same database which involves dumping a long schema and including extra context for the information in it, you don't want to write it more than once. Or if you often ask for specific things like visualisation of data, but you don't want to manually explain each time that you want to use polars and pyplot and typed python and get timing around longer running functions and ...

If you're doing ad-hoc things, yeah, just ask. If not, prompt management saves lots of time.

[Terretta]

Things like:

https://lmstudio.ai/

https://untimelyunicorn.gumroad.com/l/machato

http://msty.app/

https://www.macgpt.com/

Or, of course:

https://llm.datasette.io/

[viraptor → Terretta]

Also https://chatboxai.app/ if you don't like Msty.

[spikey_sanju]

Simple. If you're using an Apple device, go to Settings → Keyboard → Text Replacement → Add Your Prompt.

For example,

Replace: !rw

With: "Rewrite this using simple words: {your_content}"

So, whenever I type "!rw", it replaces the text with "Rewrite this using simple words: {your_content}".

You don't need to switch between multiple tabs, use an extension, or refer to the documentation again and again.

[solarized → spikey_sanju]

Use autokey for linux and friends.

% apt install autokey-gtk

[danielrm26]

I humbly submit my project as a possible solution.

It's open-source on Github.

https://github.com/danielmiessler/fabric

The tagline is Human Augmentation Using AI, but really it's a crowd-sourced library of prompts.

Basically, I solve a problem once, to a satisfactory level, and then I upload it to Fabric so everyone else can do the same.

Over 22K stars just since January 2024.

[pulvinar]

I do a frequent data-export and then keep that text open in an adjacent tab to search.

The feature is kinda hidden: preferences (upper-right corner) > Settings > Data controls > Export data. You then get email and download it from that. Unzip and open chat.html.

[sixhobbits]

I use gpt through the api hooked up to a Telegram bot

I save common prompts with

    /create blahbot you always blah

And then I can ask any bot something by saying

    /blahbot tell me about blah

Kind of hacky and misses some of the qol features built into chatgpt but it's super convenient as I use telegram a lot on phone and desktop anyway and it's got pretty good search functionality (and cheaper to pay per token than the flat fee for me plus friends and family can use it too)

[FractalHQ]

I use Claude and Cursor because they’re so much better and don’t require any fancy prompting to not completely suck.

For snippets, I use the system built into Raycast. For non-programming questions, I just ask Perplexity as I’d ask a person and its orders of magnitude better than Google or any single LLM.

[mentos]

I have trello cards where I paste a link to the chatgpt4 session so I can reuse the context weeks or months later without having to spend time searching for it.

[srutisatish_312]

It's a tedious issue that I face everyday too. I solved for myself by creating a single prompt for different tasks.

What I do is create a few prompts that will be enough for my everyday tasks. Like if you do coding on a daily basis, create a single prompt that will make ChatGPT understand your coding requirements.

Open that single prompt for coding, anyday. No need to scroll up and down (which is very frustrating btw).

Hope this helps.

[fakedang]

Well somebody could integrate some form of Algolia/Typesense style search function into their own chat GUI, but I don't think that's a very defensible endeavour in itself - it would take OpenAI or Perplexity hours to integrate such functionality into their GUI.

[gourneau]

I have used https://github.com/lastmile-ai/aiconfig to encode some of my more common prompts into application like things. You can encode them into a yaml/json file.

[fruktmix]

Would be nice if you ask the same question twice, it would look in previous "stored questions" or not for your previous answer if you need to reuse an old answer. I found it tedious to look for an old answer I need in old conversations.

[pps]

I'm using Raycast, not ChatGPT, but I think you can simply create custom GPTs for each use case (with prompt as a system instruction) and then use @ in the default new chat to quickly switch to the desired bot.

[extr]

Apple Notes app

[Havoc]

Sounds like you’d do better using a front end like silly tavern or something.

Designed for role playing but you could easily create multiple characters. Ie one summarisation character etc

[muzani]

Claude has search built in. Paid version has Projects where you get to clump together context, files, even its own tone. Quality is very similar too.

[sandspar]

I use custom GPTs or I pin the prompt into the copy+paste section of my Google Keyboard on mobile. Neither are very effective.

[pryelluw]

I built my own prompt management system. It’s similar to a cms except there’s connections and taxonomic relationships between prompts and generated replies. Haven’t seen anything like it yet. Too busy to put up online for others to use.

[djbusby → pryelluw]

I have a similar webapp front for the APIs. And it stores the conversation thread like the Chat UI.

And I add some meta data to the prompt: tags, notes, links (just Markdown)

Wouldn't mind getting integrated into my NextCloud.

[dr_dshiv]

Why can’t one simply search ChatGPT (on the web UI)?

[throw_pm23]

I tried it when it was new and found it impressive and cute as a gimmick and intriguing from a technical/scientific point of view. Since then I haven't found any uses for it and stopped using it altogether.

[trav4225 → throw_pm23]

Hah! This resonated with me. Maybe I'm just "getting old", but I fail to understand the value of 99% of this stuff...

[outofpaper → trav4225]

LLMs are tools that allow us quick access to stochastic text. They are dream machines that snatch what might be and other maybes' from the ether. They are filtered monekys on typewriters and they are wonderful in a pinch.

Also to answer the question of what I use for managing prompts and other short blobs of text I need quick access to: Espanso.

[datavirtue → trav4225]

You don't see the value of getting answers without wading through a search results trash fire?

[TRiG_Ireland → throw_pm23]

I never even tried it. I cannot conceive of a single thing I'd want to talk to an LLM about, or a single question on which I care about an LLM's answer.

[lavren1974]

joplin

[phito]

I have history disabled.

[micromacrofoot]

in the native apps you can search past conversations