Feature walkthrough

Drive a feature through every state that matters, screenshot each one from the real app, and assemble them into one page a reviewer can read start to finish. The output is a published Artifact — a private URL the author can share.

This is a demo substitute, not a test report. If the question is "does it work", run the tests. If the question is "is this the right product", this is the shape that answers it, because the reviewer sees the screens rather than a description of them.

⚠️ Dev database only — never production

Every step below writes: seeding a scenario needs contributor rows, invites, lapsed flags and half-finished states that no real account should ever be put into, and the restore at the end deletes rows. Doing that against production would corrupt real users' data.

Before the first write:

1. Confirm the app you are driving points at a dev database, and that the query tool you seed with points at the same one. Check the connection your .env actually resolves to — do not trust a log line that says which database it is.
2. Re-check after any long gap. The postgres-query skill's target can be re-pointed between sessions, and a "dev" fingerprint taken an hour ago is not evidence about the query you are about to run.
3. Never pass a writable flag to a production target. If you cannot tell which target you have, stop and ask.

Also assume the screenshots are going to be shared. Anything on screen — usernames, avatars, collection names, item thumbnails — leaves with the report. Prefer dev accounts whose content is unremarkable, and read the captures before publishing rather than after.

Setup

Copy .env.example to .env in this directory and adjust if your ports differ. It is gitignored; every developer supplies their own. Nothing in it is secret.

Screenshots default to a temp directory, outside the repo. If you point SHOTS_DIR somewhere else, keep it outside the checkout — the captures show dev accounts and their content, and this repo is public. /shots/ and **/walkthrough-shots/ are gitignored as a backstop, not as permission to write there.

You need three things running:

• the app and, for anything behind sign-in, the auth hub — see the dev-server skill
• the browser-automation server, which gives you one session per user so you can hold several signed-in accounts at once
• the postgres-query skill for seeding and for verifying the restore

Signing a session in as a given user is covered by browser-automation and the dev-login notes in memory; it is deliberately not restated here.

The loop

1. List the states before you touch anything. Write them down. Every state a screenshot has to prove, including the empty, blocked and refused ones — those are usually the interesting half and the easiest to forget. Note which are one-way.

2. Seed the scenario. Multi-user features need states the UI cannot reach on its own (someone already holding a role, an invite sent six days ago, a lapsed subscription). Seed those directly, and keep a note of every row you create — that note is your restore script.

3. Capture, in dependency order. This is the step people redo. A pending state has to be shot before it is accepted; a blocked state before the block is lifted. Get the order wrong and you re-seed the whole scenario. Sequence the one-way transitions first, then everything reversible.

4. Build the page. A template with {{shot-name.png}} placeholders, then:

node .claude/skills/feature-walkthrough/build-report.mjs report.template.html shots/ report.html

5. Publish and hand it over. See below — the sharing step has an order to it.

6. Restore, and verify the restore. Delete what you seeded, put changed columns back, then query for the rows again and confirm they are gone. Report anything you deliberately left behind and why.

Publishing, and the sharing handshake

Publish the built file with the Artifact tool. It is the right vehicle for this: one scrollable page, screenshots inline, a URL that keeps working, and it updates in place.

Artifact(file_path: "report.html", favicon: "🤝", description: "<one line for the gallery card>")

Keep the file path and favicon stable across redeploys. Same path means the same URL, so a reviewer who already has the link keeps seeing the current version; a changed favicon reads as a different page to anyone who kept the tab open.

Artifacts are private until the author shares them, and an agent cannot share on their behalf. That ordering catches people out, so do it deliberately:

1. Publish, and give the author the link.
2. Say plainly that it is private until they open it and use the page's share menu.
3. Only after they confirm it is shared, send the link onward.

Announcing the link to a reviewer before that step gives them a URL they cannot open. If you are sending the message yourself, wait for the confirmation.

To update a walkthrough from a later session, pass the existing artifact's url — publishing without it creates a second artifact rather than updating the first, and the reviewer's link goes stale. Use action: "list" to find the URL if you do not have it.

If an artifact isn't the right vehicle

It usually is, but the alternatives have real trade-offs:

The strongest combination is the artifact as the canonical page, linked from the PR, with a short chat message pointing at it.

Do not commit the report or its screenshots to this repo. It is public and permanent: the screenshots carry dev usernames, avatars and content, and the images bloat the tree. Publish them, don't check them in.

Writing capture scripts

Run them with:

node .claude/skills/feature-walkthrough/capture.mjs <session> script.js

The browser-automation server executes a chunk and throws away its return value, so a script that inspects the page has no way to tell you what it saw. capture.mjs wraps the script, writes its return value to script.js.out.json and prints it. Return whatever you want to assert on.

Scripts get page plus these, already defined:

Call dismissOverlays() after every navigation, before clicking anything. A sticky announcement sits above the page and swallows clicks; Playwright retries for 30 seconds and then fails with subtree intercepts pointer events, which reads like a broken selector and is not one.

Two more that cost time:

• Assert what the screenshot is supposed to show. Return the text alongside the shot. A screenshot of the wrong state looks fine until someone reads it.
• Re-running a script does not reset the page. The session keeps its position from the last chunk, so a script that assumed a starting URL will capture whatever is on screen instead — which is how the wrong modal ends up in a report. Navigate explicitly at the top of each script.

Screenshot craft

• Element shots beat viewport shots wherever the DOM can scope it: no clip maths, and the framing survives a layout change. Reach for a clip only for composites the DOM does not group, like a sidebar plus its header.
• Set a fixed viewport for the whole run so shots share a scale.
• Keep them tight. The artifact ceiling is 16 MB with images inlined as base64, and a page of full-page screenshots hits it fast. build-report.mjs fails the build rather than letting a rejected publish surprise you.
• Pair the before and after for anything you changed. Two narrow shots side by side carry a permission split better than a paragraph.

The page itself

Load the artifact-design skill before writing the template.

What makes these read well:

• Lead with what the feature is in two sentences, then the walkthrough in order. A reviewer should be able to stop after the first screen and still know what it does.
• Caption every screenshot with whose view it is. "Manager's edit form" beats "edit form" — half the value of a permissions walkthrough is who is looking.
• Explain the non-obvious rule, not the mechanics. Say why a row counts as shared; don't narrate that there is a sidebar.
• Put the open questions in their own section and make them answerable. A reviewer who can reply "option B" is worth more than one who has to write an essay.
• State the deploy prerequisites — anything that has to be done by hand before it ships.

Writing the message that carries it

A link with no framing gets opened last. Say what it is in one line, then what you want back — "three things I'd like your call on, last section" beats "let me know your thoughts".

If it goes out over chat, use the discord-bridge skill and read its voice.md first, so the message sounds like the person sending it rather than like a release note. A walkthrough written in memo voice under someone's own name is worse than one they wrote in a hurry. If no voice.md exists, read a few of their own messages in that thread before drafting — and weight organic messages over any an agent already sent, which otherwise trains the profile on itself.

Sending on someone's behalf is outward-facing: confirm the recipient before sending, and report exactly who received it.