Writing, part 1 — the process

This blog post is on the writing process — that is, how to begin writing, how to alternate between writing / outlining / reading / requesting feedback, etc. Note that this is not about the content itself, so I’ll have no advice on how to craft better sentences, how to organize paragraphs, etc. I don’t think I’m a great writer, but I’ve seen consistent improvements both in myself and collaborators when applying the process I outline here.

I plan to cover content in Part 2 (primarily focused on writing papers), with suggestions and outline templates for specific sections (e.g. Introduction, Abstract) or paragraphs (e.g. how to describe results), as well as some sentence-craft.

Preliminaries

Writing is more than information transfer

Some people assume writing is merely transferring information from your mind to the page, later to be transferred to other people’s minds (through reading). While we certainly want to communicate (i.e. transfer information) in the end, the process is only this simple for trivial and small-scale writing. Instead, most writing helps you discover and organize information you didn’t have when you first started. This seems to be a universal experience, even for professional writers (perhaps more so):

First, I do not sit down at my desk to put into verse something that is already clear in my mind. If it were clear in my mind, I should have no incentive or need to write about it. We do not write in order to be understood; we write in order to understand.
C.S. Lewis

Writing about something, even something you know well, usually shows you that you didn’t know it as well as you thought. Putting ideas into words is a severe test.
Paul Graham, ‘Putting ideas into words’

Assuming writing is just information transfer leads to delayed writing, where you wait until you ‘have it all figured out’. This is a mistake, as writing is both a great tool at figuring things out and for discovering the limits of what you thought you knew (as Paul Graham argues above). If you expect to merely transfer information, discovering such limits can be extremely frustrating, rather than being welcomed as a great benefit of writing. A good writing process should assume the writer does not have all of the information ahead of time, and should help them discover and organize it.

Writing happens at multiple levels

Good writing involves organizing words into sentences, sentences into paragraphs, paragraphs into sections, etc, with each one of those ‘levels’ requiring its due attention.

The error of neglecting outlines: thinking of writing as just “putting words on a page” with little regard to higher level outlines often leads to text resembling a stream of consciousness. This sometimes works as literary device (e.g. Virginia Woolf’s “Mrs Dalloway”, Marilynne Robinson’s “Gilead”), but its value lies precisely in highlighting how incoherent and meandering our minds are most of the time, often with no logical sequence between units. Unfortunately, such properties are not desirable for most other writing, which requires clearer structuring at higher levels. Sitting down and writing a flow of sentences is a great process if you want to capture unedited thoughts (e.g. journaling), but not so much for writing papers, essays, blog posts, etc.

The error of neglecting sentence-craft: another extreme says writing is mostly about ideas or outlines, and treats sentences and words as mere formalities, downplaying the importance of word choice, illustration, rhythm, flow, etc. However, individual sentences with the same meaning (and even the same words) can vary dramatically in quality. Take the following sentence pairs as examples, reading them out loud if needed:

It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife.
Jane Austen, Pride and Prejudice
Paraphrase: Everyone knows rich single men want to be married

As men are not able to fight against death, misery, ignorance, they have taken it into their heads, in order to be happy, not to think of them at all.
Pascal, Pensees 168
Paraphrase: People ignore bad things they can’t control, so as to be happy.

Language models are Few-shot learners
Tom Brown et al, title of the GPT-3 paper
Paraphrase: Some language models can do few-shot learning

The wages of sin is death
King James Bible, Romans 6:23
Paraphrase: Death is the wages of sin

The same principle is true for larger units — writing is more than just instantiating outlines. It also involves sentence-craft, a good turn of phrase, rhythm, etc. A lot of academic text unfortunately reads like an outline, and sometimes is structured like one (i.e. bullet points everywhere), which makes for very unpleasant reading (regardless of whether the ideas are good or not). Even if one does not care about beauty as an end in itself, text that is boring or unpleasant to read is harder to understand (we’ll deal more with this in Part 2).

Perhaps more importantly, an outline can seem good on paper, until you write the concrete sentences that instantiate it (remember, writing helps you discover the limits of what you know). Thus, writing ‘top down’ (first overall idea, then sections, then paragraphs, then sentences, then words) can paradoxically lead to bad outlines at every level, if one does not revise the outlines multiple times while writing. Authors should not write exclusively top down, nor exclusively bottom up, but alternate between the two.

Balance: writing happens at multiple levels, and each level needs attention

The movie producer and all-around mensch Stuart Cornfeld once told me that in a good screenplay, every structural unit needs to do two things: (1) be entertaining in its own right and (2) advance the story in a non-trivial way.
George Saunders, A Swim in a Pond in the Rain

The sentiment expressed here applies equally well to writing, with a small adjustment: each structural unit needs to (1) be good in its own right, and (2) advance the higher level unit in a non-trivial way. That is, each sentence has to be good in its own right, and advance the paragraph in a non-trivial way. Each paragraph has to be good on its own (and thus sentences have to flow well), and advance the section in a non-trivial way, and so on.

The author must either be able to do this intuitively, or deliberately work on crafting good individual units (sentences, paragraphs) as well as making sure units at every level flow well and fit together. A good writing process makes this easier to do, or at least forces the writer to do it. I have yet to meet someone who does this well and intuitively (if that’s you, you can stop reading).

A sketch of a writing process

I call it a sketch because I do not expect anyone to follow this exactly as written. However, the ideas in each step should be useful, regardless of how much or how little of the process you incorporate into your own writing.

1. The initial dump

This is the stage where you collect ‘raw data’ for your piece. Maybe you’ve already run experiments and selected some related work, so you add them here. You will also need to dig into your own mind for arguments and ideas such as ‘why I think the method I’m proposing is important’, ‘what this result means’, etc. There is no point in trying to be exhaustive (you don’t yet know everything you want to write), but it is worth spending some time eliciting as much information as you can.

I call this a ‘dump’ because I dump the data (results, ideas, facts, quotes, etc) into bullet points (not actual prose), with no organization or evaluation whatsoever (selection comes later). For this blog post, I had bullets as diverse as ‘’Common errors from grad students’, ‘What is writing?’, ‘Distinguish between writing and outlining, emphasize the importance of the latter’. For papers, I usually do a selective initial dump, focusing mostly on the introduction and results, with less material for the other sections (although still some material), resulting in bullets like “we care about our method because of X”, “our main results are X, Y, Z”, ‘don’t forget to mention argument W’, as well as a few experimental results.

2. Outline, top down and bottom up

Sidenote: what outlining is
Outlining is writing recursive summaries at different levels of abstraction: you can outline the whole text, sections, paragraphs, etc. As an example, here is an outline of the beginning of this blog post, at the highest level:

If we expand each bullet until we get individual paragraphs, we get:

Of course, the true outline is expanded even further, such that each paragraph also has an outline.

When writing an outline, I suggest explicitly asking (1) how the different parts relate to one another to form a sequence, and (2) how each part contributes to the whole. These questions apply at any level: if you’re outlining a paragraph, the ‘parts’ are sentences, while the ‘whole’ is the paragraph (conveniently, you have a summary of the ‘whole’ in the higher level section outline, where the parts are paragraphs). This process is excellent for organizing your thoughts and discovering new things about what you’re trying to write, and also for producing text that has a good flow.

I think the most common mistake I’ve observed when writing with collaborators is a lack of attention to outline. Sometimes they don’t even have any outline I can look at (much less outlines at different levels), and I have to do some reverse outlining (given a written text, produce an outline) in order to help improve the text. These outlines often reveal sentences and paragraphs (or even larger chunks!) that don’t follow naturally from the preceding ones (i.e. the structure is bad), and a lot of material that does not really contribute to the whole, and thus can be cut.

Outlining, top down and bottom up
In this step, I try to assemble the initial dump into an outline. I usually have a rough idea of what sections I wanted to include to begin with (top down outlining), but new sections or lower level units tend to emerge as I cluster data from the dump (i.e. I try to put different bullets together, bottom up).

While I outline, I inevitably discover that I need new data, not included in my dump. Thus, I alternate between outlining and gathering more data. A word of warning: it’s easy to procrastinate by ‘gathering more data’ for an inordinate amount of time. You can always read or search for one more thing, and it’s easy to convince yourself that you should. A good heuristic against this predicament is making a plan and defining what ‘done’ looks like. That is, you explicitly list what information you are looking for (e.g. ‘a citation that addresses problem X’), and what sources you plan on consulting.

Once I have units that look reasonable on their own, I move them around (and change them) until there is a sequence that makes sense, given the two questions in bold above. I try to do this at various levels, such that I end up with a high level outline of sections, and an outline of each section with a summary of what each paragraph will have (I also write outlines for some of the most important paragraphs).

3. Write (and rewrite) each section

This step is where you start writing the actual sentences that go into the text. For each section, follow substeps 3.1 and 3.2 below:

3.1 Write a bad first draft: If you followed the process until now, you will be writing from an outline, which should be easier than trying to outline and write at the same time. Even so, writing is usually the most painful stage (because it’s the most concrete), and thus ‘write a bad first draft’ is good advice. That is, rather than getting stuck trying to craft great writing right away, you should silence your internal evaluator and just write a first draft. A bad first draft almost always reveals flaws in your outline, and missing data you need to go gather. Outlines are summaries, which by construction abstract (i.e. hide) things away, and thus the presence of such flaws is almost inevitable. It’s better to find these outline flaws and fix them before you get too attached to a carefully crafted sentence that doesn’t fit the outline anymore. You can read this as a preemptive application of ‘Murder your darlings’, another good piece of writing advice (it’s easier to murder sentences before they become darlings).

3.2 Read, re-outline, re-write

I can’t understand how anyone can write without rewriting everything over and over again.
Leo Tolstoy

If the internal evaluator should be silenced when writing a first draft, it shines when reading and rewriting. It’s much easier to evaluate text that is already written, versus text you are in the process of writing. Such evaluation inevitably reveals flaws, and thus you alternate it with bouts of (re)writing.

There is a temptation to focus exclusively on local edits (typos, word choice, crafting isolated sentences etc), neglecting higher level questions like ‘how do these sentences / paragraphs fit together?’. The fix is to also alternate with reverse outlining (going from what you wrote to an outline) and verifying if the new outline is better and / or needs adjustments. Again, ask yourself (1) how the different parts relate to one another to form a sequence, and (2) how each part contributes to the whole. Just as revising text is easier than writing text from scratch, revising an outline is much easier than writing one from scratch.

As you do this, remember the reader does not have your outline
If you did a reasonable job outlining, your parts will already be ordered in a logical sequence, and will contribute to the overall whole. However, you still have to translate that labor into actual text. Here is George Saunders again:

For most of us, the problem is not in making things happen … but in making one thing seem to cause the next. This is important, because causation is what creates the appearance of meaning.

“The queen died, and then the king died” (E. M. Forster’s famous formulation) describes two unrelated events occurring in sequence. It doesn’t mean anything. “The queen died, and the king died of grief ” puts those events into relation; we understand that one caused the other. The sequence, now infused with causality, means: “That king really loved his queen.”

George Saunders, A Swim in a Pond in the Rain

Saunders is talking about fiction, but I think what he says applies well to most writing. You want to add hints to help the reader grasp your sequence and your overall whole, e.g. using transition words (e.g. ‘thus’, ‘therefore’, ‘instead’, ‘in contrast’, etc), and making explicit connections to prior material. After I get someone to read my text, I often notice that I get so familiar with my outlines that I don’t make such connections clear to the reader. Remember, the reader does not have your outline.

Why do it section by section?
I don’t move to the next section until I feel the outline of the current section is ‘good enough’. Changes in the first sections typically have a big impact later on, and thus you want to improve their outlines before errors propagate down. You should avoid spending too much time with sentence-craft at this point (as tempting as it is), otherwise you may create darlings that are harder to murder later.

At the end of this step, you should have an initial written version of the whole piece. Hooray :)

4. Top-down reading and rewriting

This step involves reading, rewriting, and re-outlining at multiple levels. I recommend doing it top down, starting with a cursory reading of the whole piece with an eye to how larger sections flow from one another, and contribute to a coherent whole (don’t stop to fix typos or individual sentences). You’ll often need to write a reverse outline here if your original outline is out of date. Then, read each section, and ask if the paragraphs connect well to one another and contribute to the overall whole of the section (re-outline and rewrite if they do not). Finally, do a final reading with an eye to detail and sentence-craft (you will inevitably have done some of this already). It’s often good to let a day pass in between each of these readings.

A potential danger in this step is avoiding big cuts or changes even when they are necessary. Cuts are painful because they mean throwing away work, but they just can’t be avoided. Big changes are painful due to fearful anticipation, a fear that says ‘if I make this change in outline at this point, I will have to change everything’. While this feeling may seem really plausible (it does every time), it is almost always false once you actually try making the change.

5. Ask for feedback

It’s almost impossible to avoid blindspots in your writing, so you want to have other people read and critique your work. I always feel my writing is great immediately after I write it, much worse the next day, and terrible after other people read it and talk to me about it.

After getting feedback, it’s tempting to just make minor writing changes, without considering whether the outline needs revision. This is a mistake, as outline problems tend to be much more severe. I suggest going through step 4 again, with the feedback you received. When writing a paper with collaborators, I actually ask for feedback on the outline at the end of step 2, before I write anything in prose.

Another temptation when you get negative feedback or confusion is to just dismiss it (perhaps with mean thoughts such as ‘how can this reader be such an idiot, the text is so clear!’). While the reader may very well be at fault, the author usually has their share of the blame, and thus you should ask yourself what caused the misunderstanding, and how you can avoid it.

You don’t want to overfit to idiosyncratic readers, so it’s important to ask multiple people to read your work. If only one reader gets confused at a certain point, maybe it’s their fault. If multiple people get confused, it’s definitely your fault. This also helps you put feedback in context, e.g. if you get one reader who says ‘I hate all your quotes’ you may be tempted to just remove them, but this feedback is softened if another reader says ‘I love this quote’. Pay attention to see if overall patterns emerge. Again, don’t shy away from what seem to be big changes in outline — they typically are smaller than they seem. When I wrote the first version of this blog post, it had a very different outline, which multiple initial readers complained about. Ironically, I could not avoid a fearful anticipation of having to change everything, even though I wholeheartedly agreed with the feedback.

Comments on the process

Here is an outline (😊) of the process, for reference:

I think something like this process forces the writer to think, outline, write, and read at multiple levels. Losing sync between the text and the outline is almost inevitable with so much rewriting, which is why I stress reverse outlining in steps 3–5. The process can also be used for individual paper sections rather than a whole piece (at which point steps 3 and 4 get merged).

The process also lends itself to the oft-quoted advice to write ‘bird-by-bird’, i.e. instead of feeling overwhelmed with the seemingly gigantic task of writing a complete piece, you break it down and commit to doing one thing at a time. If you have outlines at multiple levels, the writing task is already broken down into pieces, and intermediate tasks are also modular (e.g. one ‘bird’ could be ‘fix the outline of section 5’).

Tooling: I really like workflowy for outlining, as it makes it easy to toggle sections, move them around, and look at the outline at different levels of abstraction. I unfortunately have not found a tool that does it all well, so I put workflowy side by side with whatever word processor I’m using to write the text (Overleaf, Google Docs, or Word).
I recently started using wordtune to help me craft sentences at the very end of the process. When I’m struggling with a sentence, I think I get a good suggestion roughly 1/3 of the time.

Addendum: learn from good writing

The process above does not tell you how to write an outline, or how to go from an outline to actual text. I will offer some concrete suggestions and templates in part 2, but I think the best way to improve your overall skills is to pay attention to good writing.

In An Experiment in Criticism, C. S. Lewis contrasts two kinds of ‘readers’ (of books, art, music, whatever): those who merely use art and those who also appreciate it. Taking paintings as an example, the former look at paintings mainly as symbols, as mere substitutes for something else, such that you get comments such as ‘what a beautiful face’, and no comments about line, color, composition, etc. If instead you become like the other kind of reader, you will receive the content (e.g. understand a paper), but also notice the sentence-craft, paragraph flow, coherence of sections into a main whole, etc, and this in turn will help you improve.

Below are concrete suggestions for how to do this, but first, a warning. Lewis says bad work only admits the first kind of reading, because paying careful attention means noticing all the badness. I think this is true for papers as well, so you only want to do this with good writing.

Learn how good writers outline: simply write a reverse outline of a good piece (paper, blog post, whatever), and then ask how different parts connect to one another and how they add up to the whole. Of course, you can (and should) do this at multiple levels (paragraph, sections, larger sections).

Deliberate writing practice: you can borrow a page from Benjamin Franklin, who devised the following practice to improve his writing. He took pieces from a periodical he liked (‘The Spectator’), reverse-outlined them, waited a while until he forgot the actual writing, and then wrote his version from the outline. Writing from the same outline lets you compare your writing to good writing, and then learn from the differences. You can also invite someone else to do the same, and compare your writing to the original and to theirs (discussion really helps).

I once had an intern who was always too verbose in writing. To improve on this, we did a variation of Franklin’s technique: instead of writing from an outline, we read papers we liked (except for the abstracts), wrote our versions of the abstracts, and then compared them with the authors’ abstracts. It may have been a lucky coincidence, but the student’s skill in writing concise text improved very quickly.

Learn from collaborators
Most people write papers with other authors, and it’s often the case that another author has more writing experience than you (especially if you’re a grad student). While you can probably find better writers and do what I suggested above, it’s hard to find better writers who are willing to give feedback on your writing. The catch is that they will almost never give you abstract feedback, and will instead point to concrete problems (e.g. ‘this part is confusing’) or make concrete suggestions or edits. I was always amazed at how much Sameer would transform my writing whenever he ‘took a pass’ over anything I had written, and how good he was at suggesting big cuts with no quality loss. My writing improved by leaps and bounds once I started writing with him, but I think it would not have happened if I didn’t try very hard to extract the underlying patterns from his edits and suggestions. That is, paying attention to what he was doing, and why it made the outline or writing better.

Conclusion

I think the process above makes it a little easier to discover, organize, and write down what we want to say, but it is still hard, and a lot of work (at least for me). Someone (Samuel Johnson?) crafted a good sentence about this (I’m guessing they had to rewrite it a few times):

What is written without effort is in general read without pleasure.

I think we can improve this sentence further with some rewriting, but we’ll leave that for Part 2 :)

Acknowledgments

I don’t know if this blog post will be read with pleasure, but it was certainly not written without (significant) effort, and help from others.

Alex Cabrera, Irena Gao, Fereshte Khani, Sara Ribeiro, Tongshuang Wu and Yilun Zhou read a version of this blog post that I thought was reasonable, but was actually a bad draft. Thanks for helping me notice it, and for making suggestions that felt like I had to change everything.

Alex Cabrera, Adarsh Jeewajee, Scott Lundberg and Sameer Singh read version two, and helped me see that it was still bad in various ways. Thanks!

In hindsight, maybe I shouldn’t be dishing out writing advice… 🤷