GPT HACKS: Editing prose at lightspeed with GPT

I've set myself the challenge of writing up my GPT learnings on LinkedIn. 40 articles in 20 days. I will probably fail but..

..as a famous plumber once said: "Here we go".

GPT is terrible at editing (well)

"Summarise this article"

"Improve my writing"

"re-write this in the style of..."

If you've tried these prompts, they probably did not spark joy. Maybe some piece of the output was impressive but then some other part of it wasn't, if you actually quite like writing or have your own style / aesthetic preferences for prose... good luck trying to get a GPT to imitate your style. Without fine tuning, it's really hard*.

Yes, we could wax lyrical about how impressive some of what GPTs can do editing-wise. But ultimately, when writing, brass tax is whether it saves time and keeps you in-flow. That precious burst of inspiration, once lost, is gone. You have to protect it.

?? Keys to the castle ??

I could do a big build up to how this trick works but that sounds tiring so let's just skip straight to it. If you have experienced the above pain, why this is so much better than the above approach will likely feel immediately obvious.

Note: my full prompt is at the end of this article

Finished draft below.
Please play the role of expert editor and suggest:

1. sentences of strings of words that aren't adding value and could be deleted.
2. Improvement suggestions
3. any general feedback on what could be improved.

Output format: Please output suggestions as a table with columns: Index | "verbatim words to change" | "words in context" (include some extra words around it for easier reading) | What to change to | Explain why

[this is where I pasted the full article in plain text]        

Here is the output of that prompt for the last article I wrote on "Writing Python at Lightspeed using GPT" (model: GPT-4-Turbo):

Get it?

Feel how much better this is?

If you didn't know the "output: Table with columns A|B|C|D" trick, you're welcome! It's a personal favourite. For me, it sparks joy.

Before:

Output: [complete re-write of your text]

Workflow:

1. Read para one, some bits are good, some bits are wrong in some way, you're not 100% sure why they feel wrong.
2. Go back to your para one, read your content.
3. Go back to GPT's edit,
4. Go back to your content
5. etc.

This pain get's even worse for para 2. Unless you have the working memory of an elephant-Einstein hybrid chimera, this does not an enjoyable experience make.

After:

Output: [Numbered table of suggested changes.]

Workflow:

1. Read through suggestions.
2. Type down the numbers you accept as you go.
3. GPT implements the changes for you.

So. Much. Better.

"NEXT HACK: GPT track changes"

THE MID-ARTICLE ADVERT!  

What am I selling? Me! I'm working out what to do next career-wise. More GPT consulting? One-man-GPT-python team? Run a team solving a critical GPT challenge? Sell AI products to LLM teams? 

Tell me if you know someone I should speak to, who might have ideas or leads.

I've worked with AI in various forms (classification, then language models, then AI safety via grant-making, training & prompt engineering since ChatGPT broke the internet a year back).        

>> Profile here <<

"NEXT HACK: GPT track changes"

If you read any of the GPT editing suggestions above, you might have noticed that most of them aren't actually good suggestions. That article [here] is written in my own personal style and GPT was trying to flatten it into samey professional mediocrity. Let's dive deeper into that in a minute.

For now: how to "track changes".

I want to agree with some of the suggestions and not others.

PROMPT:

Approved: 1, 2 (been able to improve much faster than I should have been able to), 3,4, 9,

Rejected: 5 (this is in quotes, not repeated), 6 (let's actually link directly to them as well, I want to make reccs when I back products), 7, 8, 10

Please return the relevant full sections (4-5 sentences each minimum) for the above edits with GPT-format-track-changes.
* Do not show the original then the corrected text, just show the full section, with changes formatted as below.
* Separate sections with line breaks.

GPT-Track-Changes = I want you to simulate how an editor might use tracked changes in google docs using markdown in your response.
Tips:
1. Show original text verbatim.
2. Use strikethrough (~~markdown) to show deleted or changed text.
3. Put new text in [] square brackets for easy comparison.        

OUTPUT:

This is much, much better than the regular GPT editing experience for speed, flow, dealing with the quirks of that GPTs ideas / misunderstandings etc.

Aside1: There's a general formatting output trick here which is just "Simulate X". Giving a mixture of both a simple qualitative instruction that conveys great specificity as well as examples or tips (use "instructions" if you think GPT won't 'get it' from the simulate instruction, use "tips" if you think it will but want to reinforce).

Aside2: The edits it made to the last paragraph are wrong. I'll try to come back and improve this prompt later, editing hasn't been a huge part of my recent workflow util a week ago.

Personalising your "GPT-Editor"

OK great. You've got an editor prompting setup.

You might have noticed that the numbers I referenced in the feedback I gave didn't align with the numbers from the first table.

There was an in-between step:

This step is fairly self-explanatory but the important part to note in this case is that, in it's initial set of recommended edits, >50% of them were trying to strip every centilitre of personality and flair out of my writing.

Personally, I don't love writing articles, despite loving writing in general. And similarly when I read, that little bit of personality added by the writer goes a long way for keeping me reading. Treading the line between classy flair and sounding amateurish is both extremely nuanced and deeply subjective.

Here are the two suggestion tables side-by side:

Use low-touch prompting for this

DO try to figure out what prompts you should include to your editor to personalise it to you.

DON'T give it too many, you'll throw the general instruction off and editing is already a task GPTs struggle with.

In general, the more GPTs struggle with a task, the less "instructions budget" you should allow yourself.

It's kinda like if you micro-manage a brilliant employee on an already-challenging-task. You can easily make them less brilliant. Although at least with GPT they won't complain to HR.

Pulling it all together

Your instructions might look something like this:

Finished draft below.
Please play the role of expert editor and suggest:

1. sentences of strings of words that aren't adding value and could be deleted.
2. Improvement suggestions
3. any general feedback on what could be improved.

Output format: Please output suggestions as a table with columns: Index | "verbatim words to change" | "words in context" (include some extra words around it for easier reading) | What to change to | Explain why

Notes on my personal style:

1. Preservation of Personal Style: You prefer a blend of formal and informal tones that reflect your unique voice and storytelling style. This includes the use of metaphors, colloquial expressions, and a conversational tone that makes the content more relatable and engaging.

2. Inclusion of Visual Aids and Humor: You value the use of visual aids and humor to illustrate points and maintain reader interest. This suggests that any editing should respect the inclusion of such elements as long as they serve a purpose in reinforcing the article's message.

3. Balancing Professionalism with Approachability: While you aim for a professional presentation of content, you also strive for approachability and reader connection. Future edits should find a balance that maintains professional credibility without stripping away the personal touch that makes your writing distinctive.

Tweak: please still suggest some deletes but don't include any content as "keep as is"        

Let me know how it goes!

Closing: Branching your "GPT-Editor"

Branching/Sharding deserves it's own post, which it will get. But in brief:

1. Once you get your "editor-GPT" working well and reliably, you'll either need to wrap all those steps into one opening instruction set, a few-shot instruction or similar or attempt to re-do the thread to get it back to being reliable and stable.
2. Branching your thread at the point that thread is performing well is generally going to be much faster and more replicable.

In ChatGPT you can achieve this using the little editing pencil, although it's not ideal. In an interface like big-agi this is built-in.

Branching is generally going to be much more reliable due to the greater amount of context that went into getting that GPT thread into 'being on good form'.

GPT Branching: clone your GPT when it's on 'good form'

E.g., To edit this article, I'm about to branch the editor thread I just built for the last article at the second step. If that doesn't work well, I'll try branching later and then earlier in the thread, and if that doesn't work I'll try compressing the various prompts into an opening instruction.

TBC.

* I do have a few techniques for this that almost work, I might write these up later if there's interest.