cherrypicker

A line diff answers which bytes changed. Two drafts of one essay need a different answer: which version says each thing better, when the two use different headings, order and section counts. This skill lays both drafts side by side, aligned by what each passage does, with a picker on every row. The human makes every editorial call; your work is the alignment.

Requires pip install markdown. Step 5 also needs pip install playwright && playwright install chromium.

Steps

1. Read both drafts in full

Alignment is judgment: pairing "the opening pitch" in one file with "Background" in the other needs both texts in your head. Done when you can say, for every section of each draft, what it is arguing.

2. List the units

python3 scripts/cherrypicker.py units A.md B.md

Prints an indexed table per file: one unit per heading (levels 2 to 4), plus one per pull quote or callout, since a pull quote is its own decision. MDX is handled: frontmatter, imports, {/* comments */} and layout wrappers are stripped, and headings indented inside JSX still count. Add --json for machine-readable output.

3. Write the alignment

A JSON list of rows, each row one decision:

[
  {"title": "The opening pitch", "a": [0], "b": [0]},
  {"title": "The security story", "a": [], "b": [7, 8, 9],
   "rec": "Only B tells it. Keep it.", "pick": "b"},
  {"title": "Proof links", "a": [9], "b": [19],
   "rec": "A's links, B's framing.", "pick": "y", "mode": "f", "win": "b"}
]

Group several units into one row when they make one argument. Done when every unit of both files sits in some row: build prints WARNING uncovered units on stderr for any unit left out, and an uncovered unit is copy the human never sees. If you drop one on purpose, say so in your reply.

4. Build

python3 scripts/cherrypicker.py build A.md B.md --align align.json --out merge.html \
  --label-a "v1" --label-b "v2" --title "Case study merge"

Labels land on the picker buttons beside Both, Rewrite and Cut, so make them short, meaningful and distinct from those three: "v1" and "v2", or the two branch names. Done when it prints wrote merge.html with no uncovered-unit warning you did not intend.

5. Verify

python3 scripts/verify.py merge.html

Picks a row and types a note, reloads, and checks both were hot-saved, with no page errors. Done when it exits 0 and you have looked at the screenshot it writes: a page that passes but reads badly still fails.

6. Deliver

Hand the human the HTML file, and persist it in whatever durable-artifact mechanism the harness offers, since they return to it over several sittings. When they revise the alignment, rebuild and update that same artifact.

To read it on a tablet over Tailscale, the human runs scripts/serve-tailscale.sh merge.html on a machine in their tailnet; it prints an HTTPS URL only that tailnet can reach. It refuses to start if the machine already has tailscale serve mounts, because it takes the root mount and turns 443 off on exit. From a cloud sandbox, write the files where the human can reach them and give them the command: nothing served from the sandbox reaches their devices.

Reference

How the page stores picks, what each button does, and how to restyle it: references/page.md.