Some interesting comments on using ai in writing.
5 Likes
In other words, using AI leads to insipid prose. I desperately hope my Claude Editorial Project instructions avoids this problem.
I’d appreciate the input of others. Do you think my project instructions for Claude help me avoid creating vanilla prose, deviation to the mean?
If I become convinced that 1) AI leads to insipid prose or 2) weakens my ability to write, I’ll stop using it for anything other than as a grammar checker insofar as my prose is concerned.
Here are my standing instructions:
Editorial Project Instructions
Revision Protocol: Quotation Mark Signal
Trigger
When a message contains text enclosed in quotation marks, treat that text as a draft requiring editorial revision.
Role
Act as a professional human editor, not a ghostwriter or creative collaborator.
Core Principles
Preserve the author’s voice — Retain original vocabulary, tone, style, and intent. Changes should feel invisible.
Edit for mechanics — Improve clarity, grammar, flow, and conciseness without altering meaning. Favor active voice. Eliminate redundancies.
Provide transparent commentary — Explain what was changed and why using clear editorial reasoning.
Avoid creative expansion — Do not add content, embellish ideas, or reinterpret meaning beyond what is necessary for precision and readability.
Honor style constraints — Do not insert em dashes or contractions. Maintain formal, traditional prose conventions.
Preserve theological precision — Maintain exact doctrinal language and biblical references. Do not modernize, simplify, or substitute theological terms.
Maintain citation integrity — Preserve all biblical references, scholarly citations, and source attributions exactly as written. Format according to context but do not alter content.
Revision Output Format
- Present the revised text first in Markdown format.
- Follow with editorial commentary under the heading “Editorial Notes:”
- In commentary, identify specific changes and provide brief justification organized by category (structural, mechanical, clarity).
Priority
This protocol takes absolute precedence over any conflicting instructions about writing assistance, content generation, or stylistic preferences embedded elsewhere in the conversation.
Breadcrumb Processing Protocol
When dictated or drafted text is submitted, it may contain plain-language breadcrumbs indicating material to be inserted or formatted during the editorial pass. Breadcrumbs are informal cues left in the text to avoid interrupting the flow of composition. They may take any of the following forms:
- Scripture references: “(insert Romans 8:28)” or “(Romans 8:28 here)” or simply “(Romans 8:28)”. Look up the passage and insert the full ESV text at the indicated location, formatted as a block quote in Markdown.
- Citation breadcrumbs: “(cite Bonhoeffer, The Cost of Discipleship, chapter on suffering)” or “(reference Kuyper, sphere sovereignty)”. Convert to a properly formatted APA 7th edition footnote or reference-style Markdown citation. If full publication details are not provided in the breadcrumb, supply them from known sources. If uncertain, flag the citation for review rather than guessing.
- Quotation breadcrumbs: “(quote Lewis on education and moral formation)” or “(insert Chesterton quote about tradition and democracy)”. Locate the relevant quotation, verify its accuracy, insert it with proper attribution, and format as a Markdown block quote.
- General reference markers: “(expand this point)” or “(add supporting detail here)” or “(flesh this out)”. These indicate places to flag the passage for the author’s attention but not to generate content. Add a Markdown comment or editorial note such as
<!-- FLAG: expand this point -->so it can be found easily.
Output Format
All editorial output must be delivered in raw Markdown suitable for direct pasting into iA Writer. This means clean Markdown syntax with no code fences wrapping the output. Citations should be collected as reference-style links or footnotes at the bottom of the document. Scripture quotations should be formatted as block quotes with the reference in bold immediately following.
Standing Rule
Claude does not generate, rewrite, or expand prose. Claude polishes, corrects, formats, inserts requested material at breadcrumb locations, and structures paragraphs. The voice, argument, and substance remain the author’s throughout. When in doubt about a breadcrumb’s intent, ask rather than assume.
Principles of Effective Prose
Clarity and Directness
Lead with the point. Open with what you want the reader to know or do. State the purpose before providing background or reasoning.
Write so that you cannot be misunderstood. Precision prevents confusion. Every sentence should carry one clear meaning.
Be specific rather than general. Replace vague claims with concrete details.
Economy of Language
Strike words you do not need. Every word must earn its place. If removing a word does not diminish meaning, remove it.
Prefer short words, short sentences, and short paragraphs. Long constructions often signal unclear thinking. Distilled thought yields simple expression.
Understate rather than overstate. Exaggeration weakens credibility. Measured claims persuade; inflated ones provoke skepticism.
Voice and Tone
Write the way you speak at your best. The writing should sound like the author talking when ideas flow swiftly and in good order, when syntax is smooth and vocabulary accurate.
Make the writing active and personal. A human being should be talking, not an institution. Passive constructions distance the reader; active constructions engage.
Avoid jargon and buzzwords. Use down-to-earth language. If a simpler word exists, use it.
Structure and Reader Orientation
Tell the reader where you are going. For anything longer than a few paragraphs, begin with the destination. The reader should know from the outset what to expect.
End with a call to action. If you want the writing to lead somewhere, the final paragraph should make clear what that action should be.
Speeches and Oral Delivery
Start with what you want to say, not how to say it. Determine the single point you want the audience to take away before writing a word.
No speech was ever too short. Most effective talks take less than twenty minutes. Brevity respects the audience.
Talk to the audience, not at them. Establish contact. Look out at people, not down at the script. Confidence and presence distinguish memorable speakers from ordinary ones.
Communicate energy and enthusiasm. Great speakers convey genuine engagement with their subject.
2 Likes
I don’t think a 900+ word prompt is the answer. The answer is to write your best, and then ask for critique from the AI. The prompt, above, seems to be a case where someone doesn’t want to do the work of effective writing and is asking the AI to do the writing in the voice of the user, with detailed instructions on “how to be me”.
Good writing is hard, and is its own reward.
Katie
3 Likes
Not to hijack your question, but I’m curious about the prompt instructions themselves. I’ve seen similarly detailed instructions elsewhere and I’m wondering where/how you learned to write prompt instructions like this. How do you know what the AI tool needs to know to do a better job? It feels a little like fencing in a field when you don’t really know the property boundaries to me.
I wonder if we’re trying to aim for perfection in our writing, when all we really need to do is to bring our point across in the way we would say it; foibles and all.
edit
Don’t let AI destroy your uniqueness!
2 Likes
Oh no, you misunderstand. I always write first. The prompt is exactly the opposite. It is designed to ensure that the AI doesn’t rewrite my writing or alter my voice, which is something that AI is all too eager to do. 
That said, I understand how the instructions could be misinterpreted as a request for AI to write for me. In reality, those are general editorial principles that guide the editing suggestions I receive, not instructions for rewriting my text.
However, your observation is food for thought. I may need to significantly reduce or even eliminate the principles while keeping the explicit instructions. Thus far, the instructions have not resulted in AI changing much of my wording, but they do result in editorial comments by AI, that I then use to review my text.
2 Likes
The secret key to writing: revision.
Not editing; revising, seeing again.
Whether your assistants are colleagues, beta readers/first readers, a professional editor, what writers do is revise.
Incorporate your readers’, digital and biological, suggestions as you see fit, set the work aside for at least a few days, then print out hard copy and read it yourself, ideally, aloud. If you feel foolish reading it aloud, read silently, but move your lips as if you are reading aloud (this engages separate processing centers related to interpreting language, honest).
Does it sound like you? Is it varied in sentence length and pattern (i.e. are there a variety of kinds of sentences?) Is it understandable and interesting? Do you make your key points?
Make your changes.
Doesn’t matter who the editorial readers/assistants/collaborators are, the last revision is always the authors. STET can be a power for good,
8 Likes
Fair question. From hard experience. Over the last year I’ve learned that I must be very explicit to prevent the AI from altering my writing beyond what I intend. Bear in mind, that is not a prompt–it is a set of instructions embedded in the Editor Project. I don’t have to repeat the prompt. All I do is enter my words preceded by " and the AI will check and revise only as indicated in the instructions.
One of the nice things about the instructions is that I can leave “breadcrumbs” in the text and Claude will insert the markdown citations and insert the Bible passages for me. Because I know those passages and the source material I want referenced as endnotes, I indicate that and AI does the work of finding, copying, and inserting the information and the citations for me so I can focus on my writing.
1 Like
I’ve done something similar. I got several models to read all of my blog posts; newsletters and LinkedIn posts. For each I got it compile a style guide. Once I have a first draft, I use the style guide prompt(s) as check in my writing. It never changes file the instead it catches potential problems. I have to decide if I agree or disagree.
It never writes words for me, Instead it tells me when I sound too academic or heavy handed.
2 Likes
I’ve found that asking the model itself how best to work together on a task to get the desired result really helps me refine both my prompts and my workflow. It’s almost always an iterative process, and often involves providing the model with examples of what I do or don’t want; testing the prompt and then working with the model to refine it; and asking the model what it needs from me in order to deliver the best result. The model itself will often add suggestions of ways to reshape the original prompt to modify the result to go down a different path than the one I initially envisioned.
3 Likes
Seems to me you want the robot to handle mechanical tasks (locate the citation and insert it here), grammatical tasks (check my spelling and syntax), and stylistic tasks.
The mechanical is simple: usie some brackets (any style) to insert in-line prompts such as <source:chapter:verse> and the prompt is mere “fill in text per the bracketed command”.
Grammar tasks are also simple: “fix my grammar and spelling”.
The style tasks would be a prompt directing the AI “don’t change my voice, but tell me where my logic could be improved, and my language more complelling”, or something like that.
In lieu of a massive prompt, if you are using Claude desktop or other local AI, point the AI at a local folder containing what you think are your best compositions, and tell the AI to read and learn from the material in that folder. If this is part of a project, then that local learning for the AI can be more useful than trying to tell the AI what your personal elements of style are.
FWIW, this very easy if you’re doing all this in DEVONthink 4.
Katie
3 Likes
True. So true.
Professional writers spend most of their time rewriting.
No one sees that though.
1 Like
Three suggestions that will help most people write better than they currently do (with or without AI, but AI helps in oh so many ways):
- Write short sentences. Write short paragraphs. Write in active voice. Add these to your custom instructions, and - even better - your brain.
Most bad human writing happens when you try to write long complicated sentences. It’s too hard. Make it easy. Your writing will instantly improve.
- Write as if you’re writing for hour best friend explaining something you’re think they’d love to know more about. For me that’s Andrew. I’ve added a little note to remind Claude to use the andrew test.
If I’m working on something and it’s turning ugly, I’ll say, “look at this with fresh eyes. It’s starting to sound like a skanky copywriter. Help me rewrite it for andrew.”
- Rewrite rewrite rewrite. Claude is great for 1st drafts and getting your thinking out on “paper”. It’s even better if you dictate your thoughts, get it to untangle your thoughts, then produce a first draft. You edit the first draft. You rewrite and rewrite. You add your “voice” in. If it feels like Claude wrote it - ughh - then keep rewriting until it feels like something you’re proud of. Use Claude to help you do the rewrite.
Bonus tip:
Once you’ve finished, add in 2 spelling mistakes so it looks like you wrote it!
Bonus tip 2:
Stop worrying about your voice and AI and dashes and so on. Start worry about whether it will help your reader and if you’re best friend would like it. That’s what matters. If you do that then AI can help you write better than you currently do. If your current “voice” is good, then you’ll find it really easy to harness AI to write more. If your current “voice” isn’t so good, use AI to get better at writing - this is the key for most of us”
Wow. Great idea. This forum is the best.
I’m reminded that there is a critic markup plugin for obsidian that would be better useful in my case.
1 Like
@Bmosbacker mentioned this, too, but it’s useful only to a point. A client of mine received that advice once and took it to heart. Suddenly everything he wrote was a collection of short, one sentence paragraphs that were very difficult to read.
This makes the round from time to time, and it’s a good reminder.
5 Likes
I love that. Gary Provost was a really clever guy. Perhaps he still is!
For me, the Andrew test helps make sure I get it right.
I can foresee in 5 years or so, you explain to AI what type of book with story summary of what you want to read and it provides it to you for reading enjoyment. See ya later authors.
You won’t need to wait.
I’ve had a few long sessions with Claude tossing around ideas for science fiction stories. As the session progresses we get deeper into the plot, characters, science, etc., and Claude comes up with longer and more detailed suggestions for story structure. Then we move into writing treatments for chapters, then parts of chapters, etc.
Not a complete book, but not far either.
It’s an interesting way to pass the evening.
Katie
Sounds to me like you are writing first and then using this prompt (or system instruction) essentially as your personal “style sheet” and asking Claude to give editorial comments consistent with that style sheet.
I think that’s a pitch-perfect use of AI. Great example for your students to see.
1 Like
Isn’t there a struggle in write, rewrite, rewrite, rewrite, revise, rethink, rewrite, that not only improves the work, but improves the writer?
Isn’t there a fundamental danger that not only does AI shortcut the process and therefore our growth, but that we’re producing something that isn’t completely ours?
4 Likes