From Blank Page to First Draft with PACR's Writing Tools
Oct 15, 2026
A cardiology fellow I sat next to at a writing retreat last spring had forty-one PDFs open in her browser. I know the number because she asked me to count them. She wanted a witness. She'd been "about to start" a narrative review for three weeks, and she was beginning to wonder whether something was wrong with her.
Nothing was wrong with her. She'd read all forty-one papers, some of them twice. She had highlights in her PDF reader, a half-finished spreadsheet of study characteristics, a notes document with bullet points that no longer made sense to her, and a reference manager holding a different set of papers than the browser tabs. When she opened a blank document, everything she knew was spread across five places. None of it was where the writing had to happen.
She didn't have writer's block. She had a filing problem that looked like writer's block.
The blank page is usually a sorting problem
Most advice about starting a draft treats the blank page as a matter of nerve. Freewrite for twenty minutes, lower your standards, get something down. I don't think that's wrong for an essay. For a research manuscript, I think it's mostly wrong. You can't freewrite your way to an accurate account of what twelve trials found about a secondary outcome. If you try, you end up writing confident paragraphs from memory and then spending a week correcting them against the papers.
When I was supervising PhD students, the ones who stalled weren't usually the ones who'd read too little. They were the ones whose reading hadn't been sorted into an argument yet. Once the sources were grouped by what they showed, and the student could say in two sentences what the section was claiming, the writing usually moved quickly.
That's the problem we built PACR's writing tools around. PACR is an AI-native research workspace where finding papers, reading and summarizing them, outlining and drafting all happen in the same place. Context carries from one step to the next. The rest of this post walks through how I'd use it to get from a pile of sources to a first draft, with the author making the decisions at every step.
Step one: collect sources with your reasons attached
The fellow's spreadsheet had a column for sample size and a column for follow-up. It didn't have a column for why I saved this paper. That's the column that matters most when you sit down to write, and it's the one nobody records.
In PACR I'd start the project by running searches in the research search and exploration space and saving papers straight into the project, rather than downloading them to a folder and losing track of them. The habit I'd push hard is to write one line about each paper as you save it. It doesn't need to be polished. Something like:
"Largest cohort, but follow-up only 6 months; probably the anchor for the short-term section."
"Contradicts the meta-analysis on women over 70. Check whether it's subgroup noise."
"Methods paper. Cite for the definition, not the findings."
Those notes take seconds and save you hours later. They stay attached to the paper in the same workspace where you'll draft, so you don't have to rebuild your reasoning from memory three weeks on.
Step two: summaries you check, not summaries you trust
This is where most people's worries about AI in research come in, and I share some of them. Language models can produce summaries that sound fluent and are subtly wrong. They can flatten a hedged finding into a firm one, or attribute a subgroup result to the whole sample. Fabricated citations from general-purpose chatbots are well documented; if you want the evidence, PubMed has a growing literature on fabricated references from large language models.
So I don't use summaries as a replacement for reading. I use them as a first pass and a checklist, and I'm fussy about how I ask for them.
Ask narrow questions, not "summarize this"
A generic summary gives you the abstract back in different words. A useful summary answers the questions your section depends on. For a clinical review, I'd ask PACR something like: what was the population, what was the comparator, what was the primary outcome and how was it measured, what did the authors themselves list as limitations, and was anything reported only in a subgroup. Those are the details that cause errors in a first draft, so those are the ones I want surfaced.
Because the summary sits beside the paper it came from, checking it is one step, not a search through your downloads folder. Open the methods section and confirm the numbers. If the summary and the paper disagree, the paper wins, and you fix the note.
A rule I borrowed from a former student
Dana was one of my PhD students years ago. She now runs her own epidemiology group and supervises students herself. She told me recently that she has one non-negotiable rule for her lab: you don't cite a paper whose methods section you haven't opened. Not the whole paper, necessarily. The methods.
She came to this the hard way. As a postdoc she cited a study as showing an effect in adults, and a reviewer pointed out, politely, that the sample was adolescents. She'd taken the claim from another review's summary. It was the kind of mistake that doesn't sink a paper but does make a reviewer read everything else you wrote more skeptically.
I think her rule is exactly right, and it's the rule I'd apply to AI summaries. The summary tells you where to look. The methods section tells you whether to believe it.
Step three: let the structure come from the evidence
Once you have notes and checked summaries for your core papers, you can do what the fellow couldn't do with her forty-one tabs: look at all of them together and see what they say as a body of work.
This is the part of the process I think AI is best at, and the part people undervalue. Ask PACR to group your saved papers by finding instead of by author or year. Which studies support the main effect? Which show null results? Where do they disagree, and does the disagreement line up with design, population or follow-up length? You'll get a proposed grouping. Some of it will be obvious. Some of it will be wrong. Occasionally it'll show you a pattern you hadn't noticed, like all the null results coming from studies with under a year of follow-up, and that becomes a paragraph.
From those groupings, the outline starts to write itself. Each group becomes a candidate section, and each section gets a one-sentence claim that you write yourself. I'd keep that last part strictly human. The claim is the argument of your paper, and it should be yours.
If you're writing something with a formal reporting standard, bring that in now and not at revision. The EQUATOR Network collects reporting guidelines for most study types, and for systematic reviews the PRISMA statement effectively gives you a skeleton. Building your outline against the checklist from the start is far less painful than retrofitting it after a reviewer asks where your risk-of-bias assessment went.
A worked example: one section of a narrative review
Here's roughly how I'd take the fellow's review from tabs to a drafted section in PACR. The topic doesn't matter much; the sequence does.
Save the core papers into one project, each with a one-line note on why it's there.
Run targeted summaries on the eight or ten papers that will carry the section, asking about population, outcome definition, follow-up and author-stated limitations.
Spot-check every summary against the methods and results. Correct anything that's off.
Ask PACR to group those papers by what they found and flag where they conflict.
Write the section's central claim in one or two sentences, in your own words.
Ask for a paragraph-level outline that supports that claim using the grouped evidence, with each point tied to the papers it relies on.
Reorder, cut and add to the outline until it reflects how you'd explain the evidence to a skeptical colleague.
Draft the prose yourself from the outline, or ask for rough paragraph scaffolding that you then rewrite line by line.
Steps three, five and seven are where the author does the real work. Leave them out and you get a draft that's quick to produce and hard to defend.
Drafting with AI without handing over the pen
There's a version of AI-assisted writing that I dislike: you type "write the discussion section," paste in whatever comes back, and edit for tone. The result reads like every other AI-drafted discussion. It hedges in all the wrong places and claims nothing in particular. Reviewers are getting good at spotting it.
What I'd do instead is use the AI for the parts of drafting that are mostly mechanical and keep the parts that carry judgment. Turning an outline point and three checked summaries into a rough paragraph is reasonable to delegate, provided you rewrite it. Deciding how strongly to state a finding, which limitation matters most, or what the field should do next isn't something to delegate.
A second former student, Tomás, now an assistant professor running a small health services lab, described his approach to me in a way I liked. He treats AI-drafted text the way he treats a first draft from a new research assistant: useful raw material, never final, and his name is the one on the paper. He reads every sentence asking "would I say it this way, and can I point to where this came from?" If either answer is no, the sentence goes.
That second question is where having your sources, summaries and draft in one workspace pays off. When a paragraph is built from material already in your PACR project, you're checking it against papers you've saved and notes you've verified, not against whatever a general chatbot remembered from its training data.
Be transparent about how you used these tools. The ICMJE recommendations now address AI-assisted technology, and COPE has taken the position that AI tools can't be authors and that their use should be disclosed. Individual journals vary in what they ask for, so read the author guidelines before you submit, not after.
Why connected steps matter more than clever features
I've used plenty of good standalone tools over the years: a strong reference manager, a decent PDF annotator, a writing app I liked. The trouble was never any one of them. It was the gaps between them. Every time I moved from reading to notes to drafting, I lost a little context. Why did I highlight that sentence? Which version of the spreadsheet was current? Was this claim from the paper or from someone's review of it?
Multiply those small losses across a manuscript with sixty references and three co-authors and you get the forty-one tabs. You also get the discussion section that misstates a finding, and the reviewer who stops trusting you.
PACR's argument is that these steps shouldn't be separate. The paper you found in search is the paper you summarized, which is the paper your outline point depends on, which is the paper your co-author can open when they question your paragraph. Collaborators work in the same PACR workspace rather than emailing documents back and forth, so the reasoning behind the draft stays visible to everyone writing it.
None of that writes your paper for you, and I'd be suspicious of any tool that claimed it did. What it does is remove the sorting and re-finding work that sits between knowing your field and putting that knowledge on the page.
What I'd do this week
If you have a manuscript you've been "about to start," don't open a blank document. Open your sources. Pick the eight papers your argument really rests on, write one honest line about why each one matters, and try to state your central claim in two sentences. If you can't, that's useful information. The problem is upstream of the writing, and no amount of freewriting will fix it.
If you'd like to do that sorting in a workspace where the sources, summaries and draft stay connected, you can set up a PACR project and start with your own papers. Bring a real manuscript, not a test one. You'll find out quickly whether the workflow suits the way you think.
The fellow from the retreat, for what it's worth, closed thirty-three of her tabs that afternoon. She'd realized she was only going to cite eight of them in the section she was stuck on.
