If you outsource your first draft to AI, you quietly wreck your paper before it even starts.
I tried letting AI ‘help’ with a first draft once.
I spent the next week arguing with a version of the paper I didn’t even agree with.
AI can’t decide what matters in your dataset.
It doesn’t have your TASTE.
So you end up editing a polished draft you don’t fully believe.
That’s not manuscript writing.
That’s cleanup.
Here are 5 steps to write your sh*tty first draft. (And NOT outsource it to AI).
1️⃣ Critical components of a sh*tty first draft
Your job is to get the spine on the page.
Not prose.
Not citations.
Not IMRaD perfection.
Just the 5 pillars.
The 5 pillars
- Research problem
- Gap
- 2–3 key findings
- One core message
- Key implications
If those exist, you are already halfway done.
Because most “stuck” manuscripts are not stuck on writing.
They are stuck on judgement.
And good judgement is important now more than ever (in the age of AI).
This was the reason why we made a deliberate choice to ask for drafts in Research Boost. If you use AI to generate this first draft, it is going to be generic.
What this looks like in real life
A lot of people think their paper is “messy” because they haven’t written well yet.
No.
It’s messy because they haven’t chosen what to emphasize, and what to kill.
The first draft is where you choose:
- Which two results carry the paper.
- Which results are interesting but distracting.
- Which covariates matter, and which ones are decoration.
- Which audience you’re actually talking to.
AI can generate text.
It cannot make those trade-offs for you.
2️⃣ The template I use (copy/paste)
This is the exact block I paste into a blank doc.
I do not negotiate with it.
I fill it in like a form.
Problem: <what’s broken in the world/clinic>
Gap: <what we still don’t know or have been thinking wrong about>
Key findings:
- <Finding 1 headline + rough number>
- <Finding 2 headline + rough number>
- <Finding 3 optional>
Core message: <one sentence you’d say to a colleague>
Implications:
- <practice / field / next study>
That’s it.
A few notes that make this work
- Keep only 2-3 key findings. Not every p-value. Not every subgroup. Not every sensitivity analysis. Write them as declarative headlines:
- “In ___, ___ was associated with ___ (rough number).”
- “The effect was strongest in ___.”
- “The finding held up after ___.”
- Core message is one sentence. This is the sentence you’d say when a colleague asks, “So what did you find.” If you can’t say it simply, you haven’t decided it yet. And that is exactly what the first draft is for.
- The gap often gets clearer after you see your results. That’s normal. The mistake is pretending the gap is fixed from day one, then forcing the paper to match an old story. Your job is consistency in the final story, not purity in the first week.
- Implications are not a formality. This is where your expertise lives. AI loves generic implications. “Future studies are needed.” Reviewers are asking, “Who cares.” Answer like a real clinician-scientist, not like a placeholder. Whether you use Research Boost or any other AI writer, inputting at least these components is a must. These are what give the manuscript its rigor and individuality.
3️⃣ Turn the 5 pillars into IMRaD
Once the pillars exist, IMRaD is mostly assembly.
Not saying easy.
But straightforward.
INTRODUCTION
I write 3 things.
- Problem
- Gap
- One line: “So we asked…”
That’s it.
If your introduction is 5 pages and still feels vague, it usually means the gap is not crisp.
Or your core message is not decided.
The intro is not the place to “sound academic.”
It’s the place to make the reader trust that you know exactly what question matters.
METHODS
Bullets only.
Yes, bullets.
I write Methods like I’m leaving instructions for future-me who will forget everything in six months.
- Cohort and setting
- Inclusion and exclusion
- Exposure and outcome definitions
- Time window
- Covariates
- Analysis plan
- Sensitivity analyses
In a sh*tty first draft, your Methods exist to preserve decisions.
Not to impress anyone.
RESULTS
Paste key outputs early.
Ugly is fine.
I literally paste:
- Table shells
- Rough figure drafts or ideas
- Primary model estimates
- Key subgroup result
- One sensitivity analysis that matters
Then I annotate in plain language underneath.
“Signal looks consistent.”
“Wide CI, interpret cautiously.”
“Probably confounded by X.”
This is where your thinking shows up.
And this is the part AI cannot reconstruct later if you skip it.
DISCUSSION
I use 4 blocks.
- Restate the core message
- Explain the findings in plain language
- Situate them against what’s known and what’s uncertain
- Tie directly to implications
No philosophical wandering.
No literature-tourism.
Every paragraph should serve the spine.
ABSTRACT comes last
The abstract is compression.
You can’t compress what you haven’t decided.
If you draft the abstract first, you end up reverse-engineering a story you don’t believe.
4️⃣ The anti-stuck move: placeholders
This is the part that makes the first draft actually happen.
The moment my brain slows down, I type one of these and keep moving.
<ADD CITATION><CHECK NUMBER><DEFINE WINDOW><EXPLAIN WHY THIS MATTERS><LIMITATIONS HERE><WRITE THIS BETTER LATER>
Placeholders are not laziness.
They are strategy.
They protect momentum.
Because momentum is the only thing that beats perfectionism.
A first draft is not a performance.
It’s a thinking space.
Treat it like one.
5️⃣ When AI is allowed to help
Only after the spine exists.
Then AI becomes useful in the right way.
Not as the author.
As your brainstorming partner and editorial help who never gets tired.
Here’s what I’ll use it for after I have the 5 pillars:
- Tighten clarity without changing meaning
- Flag over-claims and causal language creep
- Suggest reorganizations that improve flow
- Generate alternate headings
- Point out where a reader might get lost
And here’s what I do NOT let it do:
- Decide the core message
- Choose which findings matter
- Invent implications
- Write my interpretation from scratch
Because those are taste decisions.
And taste is the job.
Pro Tip: The best practice when working with AI is to break it down into smaller chunks. So instead of asking the AI to help with the full manuscript, work with one section at a time (similar to the workflow we implemented in Research Boost).
Try this today
Open a blank doc.
Write the worst 300 words of your manuscript.
No formatting.
No citations.
No polishing.
Just force these 5 pillars onto the page:
- Problem
- Gap
- 2–3 key findings
- Core message
- Implications
If you do that, you will feel the shift.
You will stop “working on writing.”
And start working on the paper.
Send this to the co-author who keeps polishing sentences instead of deciding the point.
PROMPT OF THE WEEK
The problem solver
Prompt:
I’m stuck with [clearly describe the research problem] (e.g., low recruitment, missing data, unclear outcome, slow IRB, underpowered sample, reviewer pushback, stalled manuscript).
Act as an experienced clinical research advisor + operator.
First: ask 10 sharp diagnostic questions to find the true root cause (not symptoms). Prioritize PICOT/positioning, feasibility & data source, definitions (exposure/outcome/timing), bias/confounding, analysis plan, operations/constraints, and publication strategy.
After I answer: deliver:
• The core reason this problem exists (grounded in my answers)
• 3 actionable solutions, ranked by impact vs effort
• A step-by-step execution plan for the highest-leverage solution (what, order, why)
• Guardrails/systems to prevent recurrence
• A 30-day success definition with measurable outcomes
Keep it practical, decisive, operator-level (no fluff). Focus on leverage + speed.P.S. I built Research Boost on these same principles for clear academic writing using proven writing frameworks to keep all workflows researcher first. It will not write for you. It will write with you – intelligently, transparently, ethically.
Try it FREE: https://researchboost.com/