johnpfeiffer
  • Home
  • Engineering (People) Managers
  • John Likes
  • Software Engineer Favorites
  • Categories
  • Tags
  • Archives

Intention is all you need - Writing with LLMs

Contents

  • Why Write?
    • AI does not know anything
  • AI as a tool when writing
    • Challenges
  • How I actually write with AI
    • Anchored on Me
    • Feed It Lots
    • Less is More
    • Cross Pollinate
    • Multiple passes for different things
  • Go Write

Why Write?

Because you have something to say.

Technology is a tool, often helpful, sometimes not.

Did the word processor end writing as we know it? What about the typewriter? (Gosh I miss the days of quill and blotting ink!)

The bottleneck has never been typing speed or "output of words." It's knowing what you want to say, and why...

Large Language Models are incredibly powerful - based on the "Attention is all you need" paper that launched the current AI era.

Attention is a mechanism. Intention is the thing no model can supply, AI cannot answer for the Why.

AI does not know anything

Ask an LLM to summarize the last 10 Soccer World Cups and the most winning strategies.

It has never played football aka soccer, and it only knows the results and approaches from reading lots of data during model training.

Will its answer satisfy:

  • a coach looking for tactical advice?
  • a fan thinking about betting on outcomes?
  • a historian writing a book?

The "right answer" depends entirely on which of those people is asking, and only that person knows.

AI as a tool when writing

Let's just roll with AI = LLM for the conversation ;)

  1. Research Assistant: Get facts, grounding, existing works and opinions and trends. Clarify domain terms. Surface the most common objections and concerns. This saves real time and keeps you from getting lost going down a research rabbit hole. 🐇
  2. Draft Generator and Variations: Get past the empty blank page, or energize yourself by having something to disagree with and improve upon. Try a different tone — more formal, more casual? The draft is raw material, not the finished product.
  3. Tireless Editor: spelling, grammar, structure and narrative flow, distilling to concision. Tone, idioms, metaphors and analogies. (Note: do not ask the LLM to expand things - "more words" is never the problem.)
  4. Simulated Readers: ask for perspectives and viewpoints to become aware of blind spots. Virtual feedback that improves and hardens your piece before the real audience sees it.

Challenges

No free lunch, there's a real cost:

Volume: way more words and versions to read and consider. Too much research material to analyze — sometimes you just gotta move. And yes, hallucinations: confident, erudite... wrong.

Steering: don't let the AI drift from your topic or remove your key points. Watch for generic confidence, absolutes, and platitudes.

Authenticity: AI will generate average to mediocre text. Don't lose your voice! Avoid too much polish or "smoothing" — your rough edges are yours. (Every music fan knows that "sell-out" sound.)

Tone: You may want to be provocative or quirky, and the AI makes you banal and boring. Taking into account "other perspectives" too much can become inoffensive to everyone, interesting to no one.

How I actually write with AI

More than "categories of use", here are concrete techniques for getting better results.

Anchored on Me

Anchor with my voice: I provide the LLM with examples of my writing (even just links to my older blog posts) so it outputs closer to my style.

Find "pre-LLM" or write entirely new pieces without LLMs. The raw you, like word choice and phrasing, is distinct.

And I have to be explicit and instruct it to use my voice (otherwise, funnily enough, it will have my references and only use them in passing).

This matters more than it sounds — your voice encodes your priorities, what you emphasize, what you qualify, what you refuse to overclaim. Preserving voice is preserving intent.

Feed It Lots

Give it your raw material. Provide copious notes, bullet points, earned experience, and sources you want to incorporate. The more it has, the more you are carving from a boulder, rather than building with chewing gum and rubber bands.

Label the inclusion of your most precious bullet points as a fixed constraint: these are not to be lost. "If you need to cut length, cut elsewhere, but keep these. These are my claims; do not inflate them."

Less is More

Always ask for the AI to be concise. LLM outputs are notoriously generic and long-winded.

Say it up front: concise, cut filler. Say it at the end of your instructions too. Repetition in the instructions means "really do this thing".

Specifically ask for half as much, or even to convert a sentence or bullet point into "one best word", and you'll be pleasantly surprised.

Cross Pollinate

There are multiple good AI vendors, so I take a given draft or even just a couple of paragraphs from one model (like ChatGPT) and provide it as input to another (my friend Claude).

You can even think of it as a "council of advisors" or "mixture of experts". The other AI can act as a "critical reader" or very specific persona. No single model is an oracle or best at everything.

Another thing I've learned: these LLMs struggle when the conversation or material becomes too long. Just like a human, too many prior discussions and decisions leaves it confused about "task right now".

So I start a whole new conversation with the LLM and only give it my latest well-worked draft. Now it can focus.

Multiple passes for different things

I never ask for a "full rewrite" (╯°□°)╯︵ ┻━┻

I might as well throw away my day (after raging on the keyboard) while the whole piece is distorted and mangled.

I use distinct phases:

  1. Structure: review my structure and make suggestions. Sequencing and which sections should follow each other, is this keeping my arc or am I leaving something dangling?

  2. Substance: am I way off? Did I take a wrong turn or U-Turn, or lose the plot? Overall logic, bias, and tone?

  3. Reviewing carefully: going through a paragraph or two at most with detailed analysis and suggestions

  4. Micro-edits: give it a sentence or two and just get those tightened up

  5. Do it all over again. Great writing is incredibly hard work. Like an artist with a large palette and many different brushes - you choose what to apply when. And if the piece requires one last brush stroke.

Go Write

Try some or all of the things and see what you like and don't like. Or go back to your papyrus if that's your thing =)

Attention without intention is just pattern-matching. Be human, make mistakes, share, and learn.


  • « Cars not helicopters, or running a local LLM with MLX on a Macbook Pro
  • Building a desktop app with Golang and Wails »

Published

Dec 21, 2024

Category

leadership

~1041 words

Tags

  • ai 7
  • leadership 8
  • writing 1