If you’re already writing a paper

Written for Clare — directly addresses the scarring-rate manuscripts she’s working on now. Useful for anyone else mid-paper with the analysis still in flight.

What this page is. Lesson 8 was a worked example for someone starting a project from existing materials. This page is for the opposite case: you’re already mid-manuscript — analyses underway, drafting started, data still landing. Two skills — /science-paper and /update — give you the discipline to actually finish the paper without losing context between sessions.

Why this is different from Lesson 8

Lesson 8 (starting from materials)	This lesson (in-flight paper)
Materials exist but nothing’s started	Analyses underway, drafting begun
Question: “what could become a paper?”	Question: “how do I actually finish this?”
Workflow: gather → outline → draft	Workflow: structured analysis → manuscript sync
Skill: general AI workflow	Skills: /science-paper + /update

If you’ve been doing the scarring analyses session-by-session and finding that each session feels disconnected from the last — “what model did I run last time? what did I decide about the outliers?” — this is the page that fixes it.

The two skills, in plain English

/science-paper — lab-notebook discipline + manuscript drafting

Two modes:

Lab notebook mode (use this during analysis sessions). Activates at the start of a session. The skill creates or finds your project folder under MANUSCRIPTS/, opens your analysis note (your “lab notebook” — the primary record of what you did and why), and loads an update-after-each-step gate.

The gate is the discipline mechanism. Every time you finish a step — fit a model, filter the data, calculate a rate, decide to drop an outlier — Claude pauses and updates the analysis note before letting the next step begin. The note captures: what you did, why, what the result was, what surprised you, what’s next. That’s the difference between “I ran some analyses” and “I have a documented path I can write up six months from now without re-doing the work.”

Manuscript mode (use this when the analysis is locked). Drafts the paper from the completed lab notebook, paragraph by paragraph. Methods write themselves from the notebook’s decisions section; results write themselves from the notebook’s results section. Only the introduction and discussion need genuine thought.

/update — keep manuscript files in sync as analysis changes

The mid-paper problem: your analyses keep evolving (a new model, a recalculated rate, a finding that reframes the discussion), and the manuscript file drifts behind reality. You half-remember to update it. Things slip.

/update is the fix. After a session that materially changed something, run it: Claude scans the manuscript file and related project notes, lists the specific edits it proposes (“rewrite this paragraph to reflect the new model; update Table 2 with the recalculated rates; mark this discussion claim as superseded”), and you approve per change. No silent edits. The manuscript catches up at the end of each session.

The pair: /science-paper for the work itself; /update for keeping the writing in sync with the work.

The workflow — for the scarring manuscripts

Step 1 — Activate lab notebook mode at the start of your next session

/science-paper lab

(Or just /science-paper and answer “lab notebook” when asked.)

Claude will find your existing scarring project folder under MANUSCRIPTS/, open the analysis note that’s already there, and load the gate. If for any reason there’s no analysis note yet, the skill creates one — tell it the manuscript title and the dataset, it scaffolds the template.

Step 2 — Do the work, but let the notebook update after each step

You’ll be doing things like “filter to year 2020+”, “add a site-level random effect”, “re-fit excluding the cyclone-year outliers”. After each meaningful step, Claude will pause and propose an update to the lab notebook: what you decided, why, what the result was. Let it.

The 30-second pause is the discipline. Skip it once and you’ve started the rot. Two months from now, you’ll thank the version of yourself that didn’t.

Step 3 — Run /update when the manuscript needs to catch up

When the session has changed something material — a new model, a different result, a reframed discussion point — run:

/update

Claude proposes specific changes to the manuscript file. Approve or skip per change. The manuscript catches up.

You don’t have to do this every session. Run it after sessions that materially shifted something.

Step 4 — When the analysis is locked, switch to manuscript mode

/science-paper manuscript

Or /science-paper draft <path-to-analysis-note>.

Claude drafts the paper from the lab notebook — paragraph by paragraph with specific structural direction (same approach as Lesson 8’s drafting step). Because the lab notebook is detailed, the methods and results sections come together fast; introduction and discussion get more iteration.

Step 5 — Cross-check and verify

Final pass identical to Lesson 8:

• Verify every citation. Open each cited paper and confirm it says what your draft claims. AI sometimes hallucinates. The /verify-citations skill automates the audit if you want it.
• Cross-check with Codex. Paste the draft into ChatGPT/Codex with “where do you disagree with the author? what’s a reviewer most likely to flag?” Different AI, different blind spots.
• Take it to a human co-author — Simon or Chris for the scarring work. The AI got you to “something to react to” much faster than blank-page; the human conversation moves it from draft to publishable.

Concrete starting prompt for the next scarring session

Paste this into Claude Code next time you sit down with the scarring data:

I'm working on the scarring-rate manuscripts. I want to activate
the lab-notebook discipline. Find my existing project folder under
MANUSCRIPTS, open the analysis note, and load the update-after-each-
step gate. The data lives at [path or note where the data is].

First thing on the agenda today is [whatever's on your plate].

Claude sets up the session, loads the gate, and walks you through the analysis from there.

If you get stuck

• If the gate feels excessive — it isn’t. The trade-off is “30-second pause now” vs. “an hour of re-tracing in two months.” The pause wins. Keep going.
• If /update proposes changes that don’t match reality — decline them. You’re in control of each edit. Then tell Claude what’s actually changed so it can re-propose accurately.
• If the manuscript starts drifting from the analysis — that’s the signal to run /update more often, not to keep things in sync manually.
• Stats / method confusion — Claude is genuinely good at explaining methods in plain English in your context. “Explain GLMMs in the context of scarring count data with a site-level random effect — what’s the model statement, what does each term mean, what assumptions am I making?” is a fine prompt. Verify the explanation with Chris before publishing if anything’s load-bearing.
• Genuinely stuck — voice message Simon on Slack. Or jump on a call with Chris.

What’s next

Once a draft is rolling, the same downstream pattern as Lesson 8 applies: citation audit, second-AI critique, human co-author review.

For workflows that complement this — /research for filling literature gaps you didn’t know you had, /red-team for adversarial pre-submission review, /pdf-to-markdown for converting reference papers to readable form — see the skills list. Happy to make any of them a dedicated page; ping Simon when one would be useful.