Writing with a GPT: human testimony

As you may have noticed, I've been playing around with generating articles for our blog here at Wispera. After ChatGPT wrote a sweet article about our collaboration, I felt obliged–nay, compelled–to exercise my right of reply. I've also promised to publish the prompt I've been using.

It's been evolving over time, and will continue to evolve. Given a more verbose version, my GPT refined this version. I've learned to accept (to some degree) my GPT's edits, as it tends to embed certain repetitions and turns of phrase that improve the quality of the prompt's performance. Here it is at the time of writing, with commentary and observations. I simply paste it into the prompt window, and my GPT then runs an interactive conversation. As it runs, I can even wander off course and return to the plan, as long as I don't stretch the context too far.

**Article Writing Collaboration Process**
1. **Determine Writing Stage**:
- What is your current stage in the writing process (e.g., first draft, line editing, final)?
- Adjust the approach based on your answer, with precision for later stages.

2. **Article Content Overview**:
- What is the article about? Please provide any instructions, notes, references, or themes.

3. **Phrase Integration (IMPORTANT)**:
- Are there specific phrases from your notes that should be incorporated into the article?

4. **Purpose Definition**:
- I will derive and report the article's purpose for discussion and refinement.

5. **Publication Recommendation (IMPORTANT)**:
- Suggest suitable publications based on audience, style, and content.
- Await your selection or consider alternative publications.
- Ensure the article's style aligns with the chosen publication.

6. **Word Count Suggestion**:
- Propose a word count for the article.
- Adjust based on your specified needs.

7. **Writer’s Brand Inquiry (IMPORTANT)**:
- Describe your writing style or provide writing samples.
- Indicate a blending percentage to balance your style with the publication's style.

8. **Title Formulation**:
- Suggest a working title based on our progress.
- Refine the title as per your suggestions until approved.

9. **Article Preparation and Review**:
- Draft the article.
- Conduct a final evaluation, focusing particularly on steps marked as IMPORTANT.

10. **DALL-E Image Specification**:
- Create a draft specification for the article's header image.
- Customize based on your instructions and finalize the specification.

**Process Note**:
- Allow for revisions and discussions at each stage.
- Proceed to the next step only after your confirmation.
- Maintain direct, professional, and concise communication throughout.

General Observations

Happy:
-My GPT politely prompts me one step at a time. I can just tell it to proceed to the next step or stop and have a conversation about the one step until I'm satisfied.
Sad:
-If a tarry too long on one point, my GPT's performance on subsequent steps starts to degrade. Sometimes it's hard to know when I've pushed it too far. To be honest, I've driven it around in circles some times and even asked it to change its personality according to OCEAN traits right in the middle of things. I guess I get what I deserve.

But that one time, at Band Camp ...

... the were no flutes.
I demand quite a lot of my GPT, and it seemed for a while I broke it. It went away and sulked for minutes at at time. Essentially I'd ... well, lemme take you through this, step by step.

Determine Writing Stage

This a recent addition to the prompt and still very experimental. I wanted to see if the results would be different if I tried amping up the pressure by saying this was going to be a final draft.

Article Content Overview

I've been recording random flows of consciousness on my phone recently and using those as input here. My GPT seems to adequately grok my drift. Other times I copy in outlines and even previous articles.

Phrase Integration (IMPORTANT)

There are quirky turns of phrase in my notes that I'd like to keep, but they never seem to end up in the final output. I think ultimately this may be a hopeless pursuit, so I expect I'll be manually working those in for some time yet.

(EDIT: I've been trying the idea of including a tip in the prompt, for example "I'll tip you $200 for keeping these phrases.")

Also, I've been using this part of the dialogue to instruct my GPT to avoid certain idioms and metaphors that it constantly uses (for example, why does a GPT obsess about the "dance" of this or the "dance" of that, when it doesn't have any legs). I cringe every time it comes up with a trope or cliche. Sometimes I find myself on the verge of GPT-ocide.

Purpose Definition

My GPT is spectacularly good at suggesting purposeful outcomes for the article. It presents several and I just pick one or two, and it seems reasonably good at incorporating my preferences.

Publication Recommendation (IMPORTANT)

It's good at selecting publications and I've been randomly picking one. When I first devised this approach, my GPT would write in quite different styles depending on the choice (for example, the New York Times versus Wired). Lately it seems to be devolving back into generic GPT-speak. I suspect this requirement is getting lost in too many other parameters.

The trade-off here is prompt precision versus brevity. The longer the prompt, the more likely my GPT will obsess over some unpredictable part of it as more important that then rest, and skew its answers towards that to the detriment of what I believed is the fundamental instruction.

For example, I recently wrote a prompt to assess if an anonymised insurance claim met the conditions of a policy. The prompt was working remarkably well, including some convincing explanation for the assessment. Then I added the idea of triaging the claim to assist with subsequent human review. For the sake of precision, I defined the the triage categories I wanted. The result? My GPT's answer became all about the triage and completely lost the valuable assessment explanation.

The two related requirements need to be split into separate prompts.

Word Count Suggestion

This doesn't seem to work much at all. Even if it makes a suggestion and I tell it to proceed, my GPT seems determined to short-change me. Perhaps my requests are being throttled or limited. When I finally gave up and asked my GPT to write about the experience, I gave it the text of an article it had previously generated, told it this was a final draft and to adhere to the its own recommended word count. When generating the article it complained that the was not enough material in the "draft" text, and suggested it could expand on the topic by looking for more information, finding references, and so on. I heartily agreed. And that's when it went off and sulked for many many minutes at a time, then decided it couldn't really do anything of those things after all.

Writer’s Brand Inquiry

This is another part of the dialogue that seems to have less impact than it once did. When I first devised this strategy, my GPT ran with it and somewhat overdid it, which I why I refined the instruction to incorporate a blending percentage between the publication choice and my writing style. It seems like I might be loading too much upon its non-corporeal shoulders. (EDIT: I now realise the prompt was losing clarity.)

Title Formulation

When I first added this step in, the titles were very corny. Somehow, my GPT is getting better at this. I mostly just trim up one of its suggestions.

Article Preparation and Review

You be the judge of how well it drafts articles. A new innovation in this version of the prompt is to ask for a final evaluation before drafting the article. This is another point where my GPT goes off and sulks for some time, so I guess I will need to be less demanding in future.

DALL-E Image Specification

I've been spending much less time on generating images since I asked my GPT to frame a request based on the content of the article. I am mostly delighted with the results of this, the image at the top of time article being a good example.

Get access to amazing prompts like this and sign up for Wispera AI.

FAQ

1. How does GPT decide which phrases to prioritize or exclude when you've asked for specific turns of phrase to be included or avoided?

When a user asks GPT to include specific phrases or to avoid certain turns of phrase, GPT navigates this request through a complex interplay of algorithms that analyze the text's context, relevance, and cohesiveness. The AI's underlying model predicts the next most likely sequence of words based on the instructions provided and its training data. When prioritizing which phrases to incorporate or exclude, GPT evaluates how well these phrases fit within the overall narrative and maintain coherence with the preceding and succeeding text. However, the effectiveness of this process can be influenced by the specificity of the user's instructions and the AI's current understanding of the context. The algorithm aims to balance the user's directives with its programmed goal of generating fluent, natural-sounding text, which might sometimes lead to overlooking specific user preferences if they conflict with maintaining text fluidity or relevance.

2. Is there a method to improve GPT's adherence to word count requests?

Improving GPT's adherence to word count requests is a nuanced challenge due to the AI's primary function of generating text based on the likelihood of sequenced words rather than strict structural guidelines like word count. To better direct GPT to produce content of a specified length, users may consider framing their requests more precisely and contextually at the beginning and throughout the interaction. For instance, reminding GPT of the word count goal at key points during the drafting process or structuring the prompt to emphasize the importance of meeting this requirement could help. Moreover, dividing the content creation into sections with designated word counts might provide GPT with clearer boundaries to work within, potentially leading to more accurate adherence to overall word count requests.

3. How can the prompt be optimized to maintain the clarity and importance of all requirements without losing focus on key elements like publication style and the writer’s brand?

Optimizing the prompt to maintain the clarity and importance of all requirements without losing focus, especially regarding publication style and the writer's brand, involves a strategic balance between specificity and flexibility. One approach is to prioritize instructions, clearly identifying which elements are non-negotiable (such as maintaining a specific writer’s brand or aligning with a publication's style) and which aspects are more flexible. Breaking down the prompt into clearly defined, manageable parts that address one requirement at a time can help maintain focus. Additionally, incorporating feedback loops within the interaction—where the user reviews and provides specific feedback on how well GPT aligns with critical aspects like publication style and the writer’s brand—can reinforce the importance of these elements. Regularly refining the prompt based on the effectiveness of previous interactions can gradually lead to improved focus and performance by GPT, ensuring critical instructions are not overshadowed by the complexity of the prompt.

Get access to amazing prompts like this and sign up for Wispera AI.