tldraw/tldraw/write-pr

Writing pull requests

Standards for PR titles and descriptions in tldraw/tldraw.

Write for the reviewer (read this first)

This is the governing rule; everything below serves it. Write the description for a reviewer who knows the codebase architecture but has not read your code or the diff. Their time is the scarce resource. The description's job is to give them the framing they can't get from the code, then get out of the way.

Default to short. A few sentences of framing is the norm, not the exception. Match length to the change: a one-line fix needs a sentence; a new system needs the fuller treatment. A description that looks long and structured is not more valuable — often it's less, because it buries the framing under scaffolding. If a reviewer would need their own AI to interpret yours, it has failed.

Order it coarse to granular. Structure the description as an inverted pyramid, most important first. A reader should be able to stop at any point and leave with a correct, coherent understanding: a skimmer gets the goal and motivation from the opening lines; someone weighing the approach reads into the design and decisions; a close reviewer continues to the specifics. Never make someone read to the end to find out what the PR is for.

Cover the following, roughly in this order — each layer more detailed than the last, and each optional once the change no longer warrants it:

• Goal, motivation, and use case. Why this change, why now, what it's for. What was wrong or missing before. This comes first, always.
• The higher-level change. The shape of the solution — behavior and structure, not a line-by-line restatement of the diff.
• API design and decisions. New or changed public surface, the approach you picked, what you ruled out and why, and anything you're unsure about. A decision you don't surface is one the reviewer can't catch.
• Example snippets and fine detail. For anything touching an API, data shape, or usage pattern, a few lines of before/after say more than a paragraph. Keep them minimal.

Do not:

• Restate the diff. No file-by-file walkthrough, no narrating what a function does, no describing how the code works step by step. Reviewers can read code.
• Generate tables or lists that mirror the code — a Method | Description table that just re-says the signatures, an inventory of every changed file, a term glossary of self-explanatory names.
• Pad with ceremony. Structure that exists to look thorough is noise.

Never invent the *why*. The motivation and trade-offs must come from real intent — the commits, the linked issue, or the author. If you don't know why a change was made or what was considered, ask the user; do not guess. A confident but fabricated rationale is worse than useless: it's misleading, and it's the thing reviewers most need to trust.

PR title

Use semantic PR titles (Conventional Commits format):

<type>(<scope>): <description>

Types

• feat - New feature
• fix - Bug fix
• docs - Documentation only
• refactor - Code change that neither fixes a bug nor adds a feature
• perf - Performance improvement
• test - Adding or fixing tests
• chore - Maintenance tasks

Scope (optional)

A noun describing the affected area: fix(editor):, feat(sync):, docs(examples):

Examples

• feat(editor): add snap threshold configuration option
• fix(arrows): correct binding behavior with rotated shapes
• docs: update sync documentation
• refactor(store): simplify migration system

PR body

Bug fix and improvement PRs (bugfix and improvement change types) use the before/after template below. Everything else uses this one:

<description paragraph>
 
### Change type
 
-[x] `bugfix` | `improvement` | `feature` | `api` | `other`
 
### Test plan
 
1.Step to test...
2.Another step...
 
-[ ] Unit tests
-[ ] End to end tests
 
### Release notes
 
-Brief description of changes for users

Description paragraph

Start with: "In order to X, this PR does Y." Follow the reviewer-first rules at the top of this skill.

• X is the concrete situation that made this work necessary — the real thing someone was trying to do — not a restatement of what Y does. "In order to let apps carry undo history across editor rebuilds, this adds an API to carry undo history across editor rebuilds" is circular: X just renames Y. Push X up a level to the actual goal, e.g. "so desktop can reload the editor without losing the user's place, the way HMR preserves component state across a code edit."
• Name the real driving case, not a hypothetical. If you're inventing example scenarios to illustrate the why ("e.g. if someone toggled a plugin…"), you're reverse-engineering a justification from the finished code — you haven't written down the case that actually prompted it. Work forward from that case instead.
• Beware abstract mechanism as a stand-in for motivation. A true technical fact about the SDK ("editor config is fixed at construction time") explains a constraint, not why anyone cares. Keep going until a reader can picture the specific thing that stopped working or became possible.
• Keep it specific - avoid vague phrases like "improve user experience"
• Link related issues in the first paragraph
• Don't expect readers to also read the linked issue

Bug fix and improvement PRs

Bug fix and improvement PRs use a fixed shape instead of the description paragraph. The same reviewer-first rules apply; the structure exists so a reviewer can see what was wrong (or lacking), what it does now, and why, without reading the diff.

This PR fixes a bug where <symptom a user or developer would hit>.
 
### Before
 
<what the code did and why that produced the symptom>
 
### After
 
<what the code does now>
 
### Implementation notes
 
<optional: what exactly changed and how — decisions, trade-offs, anything non-obvious>
 
### Change type
 
-[x] `bugfix` | `improvement`
 
### Test plan
 
-[x] Unit tests — the new test fails on `main` and passes with this change
 
### Release notes
 
-Fix <symptom> | Improve <behavior>
 
### Code changes
 
| Section   | LOC change |
| --------- | ---------- |
| Core code | +10 / -2   |
| Tests     | +5 / -0    |

• Intro — one sentence. For a bug: "This PR fixes a bug where X," where X is the observable symptom, not the cause or the fix. For an improvement: "This PR improves X so that Y," where Y is what a user or developer can now do. Link the issue here if there is one. If the PR does more than one thing, add a second sentence ("It also ...") rather than a list.
• Before — the behavior before, and the mechanism that produced it. For a bug, name the function or path at fault; this is the one place a short description of how the old code went wrong belongs, because it is the cause the reviewer is checking the fix against. For an improvement, describe what the user or developer had to do (or couldn't do) and what in the code made it that way.
• After — the behavior after. State what the code now does, at the same level of detail as Before. Note any spec or doc that was updated alongside.
• Implementation notes — optional. Include only when there is something the Before/After pair doesn't carry: an approach you chose over another, a subtlety in the change, a follow-up you deliberately left out. Skip the heading entirely when there is nothing to say; most small fixes have nothing.
• The ### Change type / ### Test plan / ### Release notes / ### API changes / ### Code changes blocks follow as usual. For bug fixes, say in the test plan whether the new tests fail on main.

tldraw/tldraw#10117, #10118, and #10119 are worked examples of bug fixes in this shape.

Before and after recordings

When a bug fix or improvement changes how a canvas interaction behaves — a drag, resize, rotate, crop, handle, or tool gesture that a reviewer would otherwise have to run locally to see — put a short recording in each of the Before and After sections. Prose describes the mechanism; the recordings show the symptom and its absence, and reviewers trust them more than a description of a jump or a jitter. Skip this for changes that are not visible on the canvas (data, API, docs, tooling).

scripts/record-interaction.mjs records a scripted interaction against the running examples app as a 16:9 MP4. It hides the tldraw chrome, injects a visible cursor (headless recordings have none), runs a scenario module, trims the setup frames, and transcodes with ffmpeg. scripts/example-scenario.mjs is a starting point; copy it to a scratch location and change the body to the interaction the PR affects. Do not commit scenarios.

It needs ffmpeg on the path and a Playwright Chromium build. Install scripts are disabled in this repo, so the browser is not downloaded by pnpm install; run pnpm exec playwright install chromium once. Then, from the repo root with pnpm dev serving the examples app:

# After: the current branch
node skills/write-pr/scripts/record-interaction.mjs <scenario.mjs> <scratch>/after.mp4
 
# Before: main's versions of the files this branch modifies, hot-reloaded by vite
git diff --name-only --diff-filter=M origin/main...HEAD | xargs git checkout origin/main --
sleep 10
node skills/write-pr/scripts/record-interaction.mjs <scenario.mjs> <scratch>/before.mp4
git diff --name-only --diff-filter=M origin/main...HEAD | xargs git checkout HEAD --
git status --porcelain   # must be empty before continuing

Restrict the checkout to modified files: checking out a whole directory from main also stages files that only exist there, and files added on this branch have no main version to restore. Only use this approach when the working tree is clean.

Rules for the recordings themselves:

• Same scenario for both. The two videos must differ only in the code under test, so the reviewer compares like with like. Record After first, confirm the scenario shows the behavior, then record Before.
• Make the symptom unmissable. Pick inputs that exaggerate the bug — a large rotation, a long drag, a non-square shape. A 15 degree turn that shifts a shape by a few pixels reads as nothing at video scale.
• Check the frames before attaching. Pull a contact sheet (ffmpeg -i before.mp4 -vf "select='not(mod(n\,15))',scale=320:-1,tile=6x3" -frames:v 1 sheet.png) and look at it. A Before recording that looks identical to After usually means the checkout did not take effect, not that the bug is subtle.
• No burned-in labels. The ### Before and ### After headings already say which is which.
• Keep them short. A few seconds of setup, the interaction, a beat at the end. Under ten seconds each.

Attach with gh pr create or gh pr edit (gh 2.99 or later, run inside the repo, absolute paths). Reference each file in the body as the only content of its paragraph, directly under the heading it illustrates, and gh rewrites the reference to the uploaded asset:

gh pr edit <number> --body-file body.md --attach /abs/path/before.mp4 --attach /abs/path/after.mp4

### Before
 
<mechanism>
 
![Before](/abs/path/before.mp4)
 
### After
 
<behavior now>
 
![After](/abs/path/after.mp4)

Change type

• Tick exactly one type with [x]
• Delete unticked items

Test plan

• List manual testing steps if applicable
• Remove the numbered list if changes cannot be manually tested
• Tick checkboxes for included test types

Release notes

• Write brief notes describing user-facing changes
• Use imperative mood: "Add...", "Fix...", "Remove..."
• Omit this section entirely for internal work (CI, tooling, tests, etc.) that has no user-facing impact

Concepts, examples, and FAQ (only when a reviewer genuinely needs them)