Scientific Writing — Stage 1: Find the Story

Overview

A research repo accumulates parallel and abandoned experiments. Before any outline or prose, establish ONE linear story: the thesis, the experiments that support it, and an explicit record of what was dropped. Output: docs/writing/<story-slug>/story.md in the consuming repo.

Do not write the story from your own assumptions. Map the repo, then converge with the user.

Method backbone — Steps 1–2

This stage is Steps 1–2 of the fifteen steps (detail: scientific-writing/fifteen-steps.md):

• Step 1 — main point. The thesis is one short declarative sentence (subject + verb) that scopes the phenomenon and gives focus. Every later paragraph must illuminate it. It need not become the final title.
• Step 2 — narrative arc. The arc answers what did we want to learn, why, what did we do, what do we know now, and follows the key transitions (question→answer, theory→evidence, data→interpretation). Prefer a two-part structure — the most memorable and publishable shape. Confirm which two-part scheme fits when you converge with the user (e.g. experiment 1 + experiment 2; simple model + complicated model; experiment + model to explain it; model + experiment to test it; observed pattern + model to explain it).

Process

1. Map the research cloud. Dispatch Explore subagents (read-only) to inventory:

experiments/scripts, results/, figures/, notebooks, git log, and CLAUDE.md. Use superpowers:dispatching-parallel-agents if there are independent areas. Goal: a list of every experiment/result thread, each tagged candidate-main vs possibly-dropped.

1. Converge with the user, ONE question at a time. Confirm: the working title, the single thesis, which threads are IN, which are DROPPED. Prefer multiple-choice. Do not proceed to writing until the user confirms the thesis and the in/out split.
2. Write story.md (template below).

story.md template

# <Working title>
slug: <kebab-case-slug>

## Thesis
<one or two sentences — the single claim the paper makes>

## Narrative arc (linear)
Two-part scheme: <which Step-2 scheme, e.g. "experiment + model to explain it">
1. <beat> 2. <beat> 3. <beat> …  (the logical order the paper will argue in)

## Beat → artifacts
| Beat | Supporting artifacts (scripts / results / plots / notebooks) |
|------|------|
| <beat> | `path/to/script.py`, `results/x.csv`, `figures/y.svg` |

## Dropped / out of scope
- <thread> — why dropped (user-confirmed)

Red flags — STOP

• Writing story.md before mapping the repo → map first.
• Asserting a thesis the user hasn't confirmed → ask.
• A beat with no backing artifact → either find it or it's not a beat.
• An abandoned experiment silently included → it goes in Dropped.

Next stage

Once the user approves story.md, invoke scientific-writing-outline.