look
Skill2 starsChanged 5 days agoMITView on GitHub
Inspect the shared Chrome viewport — DOM queries, computed styles, box model to diagnose; one screenshot to confirm a visual change landed, otherwise only when asked. Headless mode gives a coding agent its own private browser to check that a UI works — navigate, click, type, collect errors — without touching the user's window.
- Reads credentials
- Sends data out
In this repository24 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- 1passwordSkill
- brainSkill
- brainstormSkill
- bridgeSkill
- claude-switcherSkill
- clear-cacheSkill
- commitSkill
- cross-artifactSkill
- ctoSkill
- diskSkill
- killSkill
- parseSkill
- pbarSkill
- planSkill
- pruneSkill
- pushSkill
- reportSkill
- researchSkill
- swarmSkill
- vibe-auditSkill
- video-editorSkill
- webSkill
- wikiSkill
The file
We split this skill into the rules it gives an agent
12 rules. About 0 are common, 7 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 19
Before any command, resolve the port from the registry:
- Very specific: line 33
If $LOOK_PORT is empty, the server isn't running — launch it (see below).
- Distinctive: line 43
Check if a server is already running for this project:
- Distinctive: line 58
If this fails, launch the cobrowser automatically — don't ask the user:
- Distinctive: line 65
Run this command in the background — it's a long-running process.
- Very specific: line 81
If /look is called with a URL argument (e.g.
- Very specific: line 95
All commands are POST requests to http://127.0.0.1:${LOOK_PORT} with a JSON body.
- Distinctive: line 125
Use it for two things only: confirming a visual change you just made (one capture of the affected area, looked at before you say it works), …
- Distinctive: line 172
no entry in the registry.
- Distinctive: line 259
When the user says "look at X":
- Very specific: line 284
Use status to get current viewport width, then check what CSS values apply at that width by inspecting the element.
- Very specific: line 316
If .claude/kit.json has a rules."look" entry, read it and apply it as an additional instruction for this skill.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Look — Chrome Viewport Inspection
Inspect elements in the shared Chrome browser that both you and the user are looking at. The user drives the browser (navigating, setting mobile viewport via Chrome's device toolbar). You inspect programmatically via a Puppeteer server.
DOM-first to diagnose, eyes to confirm. Use inspect or dom for exact values when diagnosing a known problem — computed styles answer "why is this 40px too wide" better than a picture does. But a measurement only reports on the elements you chose to measure, and a layout regression usually lands somewhere else. So after making a visual change, take one screenshot of the affected area and look at it before calling the change done. Otherwise, screenshot only when the user asks.
Multi-profile support
Each project gets its own Chrome instance with an isolated profile. Instances self-register in ~/.claude-chrome-registry.json with their profile name, port, and PID. The profile name defaults to the project directory basename (e.g. my-project).
Resolve the port for this project
Before any command, resolve the port from the registry:
LOOK_PORT=$(PROFILE_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)" node -e "
const fs = require('fs'), path = require('path'), os = require('os');
const reg = JSON.parse(fs.readFileSync(path.join(os.homedir(), '.claude-chrome-registry.json'), 'utf8') || '[]');
// This project's window only: the repo root's name, never another entry (it is another project's browser).
const profile = path.basename(process.env.PROFILE_ROOT);
const entry = reg.find(e => e.profile === profile);
if (entry) { try { process.kill(entry.pid, 0); console.log(entry.port); } catch { console.log(''); } }
else console.log('');
" 2>/dev/null)If $LOOK_PORT is empty, the server isn't running — launch it (see below). Use $LOOK_PORT in place of a hardcoded port in all curl commands.
Never drive another project's window. The registry lists every project's cobrowser on this Mac, and other
sessions are often working in them. Look up only this project's own entry (the repo root's directory name, as
above): no fallback to the first entry, no port from memory or from another project. An empty $LOOK_PORT means
launch this project's own, from the repo root. Never stop, kill or restart a browser server or Chrome you did not
launch: find yours by its pid in the registry or the stdout line it printed, and kill that pid only (SIGTERM).
Prerequisites
Check if a server is already running for this project:
LOOK_PORT=$(PROFILE_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)" node -e "
const fs = require('fs'), path = require('path'), os = require('os');
const reg = JSON.parse(fs.readFileSync(path.join(os.homedir(), '.claude-chrome-registry.json'), 'utf8') || '[]');
// This project's window only: the repo root's name, never another entry (it is another project's browser).
const profile = path.basename(process.env.PROFILE_ROOT);
const entry = reg.find(e => e.profile === profile);
if (entry) { try { process.kill(entry.pid, 0); console.log(entry.port); } catch { console.log(''); } }
else console.log('');
" 2>/dev/null)
curl -s http://127.0.0.1:${LOOK_PORT}/status 2>/dev/nullIf this fails, launch the cobrowser automatically — don't ask the user:
# Launch puppeteer server for this project
node scripts/puppeteer-server.cjsRun this command in the background — it's a long-running process.
Then wait for it to be ready and capture the port:
sleep 2
LOOK_PORT=$(node -e "
const fs = require('fs'), path = require('path'), os = require('os');
const reg = JSON.parse(fs.readFileSync(path.join(os.homedir(), '.claude-chrome-registry.json'), 'utf8') || '[]');
const profile = path.basename(process.cwd());
const entry = reg.find(e => e.profile === profile);
if (entry) console.log(entry.port); else console.log('');
" 2>/dev/null)
for i in 1 2 3 4 5 6 7 8 9 10; do curl -s http://127.0.0.1:${LOOK_PORT}/status > /dev/null 2>&1 && break; sleep 1; doneIf /look is called with a URL argument (e.g. /look http://localhost:3000), pass it to the server:
node scripts/puppeteer-server.cjs <URL>To launch with a specific profile name:
node scripts/puppeteer-server.cjs --profile my-other-project http://localhost:3000Commands
All commands are POST requests to http://127.0.0.1:${LOOK_PORT} with a JSON body.
inspect — Computed styles + box model (use this most)
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"inspect","selector":".hero-section"}'Returns: computed styles (display, position, width, height, padding, margin, font-size, line-height, overflow, flex properties, gap, z-index, colors), bounding rect, tag/class/id, child count, text content.
dom — HTML structure + children
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"dom","selector":".hero-section","children":true}'Returns: outerHTML, attributes, child elements (tag, class, id, text). Use children: true to see immediate children.
screenshot — Viewport or element capture (to confirm a visual change, or when the user asks)
# Full viewport
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"screenshot"}'
# Specific element
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"screenshot","selector":".hero-section"}'Returns: { "path": "/tmp/puppeteer-screenshot-<ts>.png" }. Read the file to view it.
Use it for two things only: confirming a visual change you just made (one capture of the affected area, looked at before you say it works), and when the user asks. Diagnosis stays DOM-first.
eval — Run JS in page context
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"eval","expression":"document.querySelectorAll(\".card\").length"}'Returns: { "result": "6" }. Use for queries that don't fit inspect/dom.
status — Current page state
curl -s http://127.0.0.1:${LOOK_PORT}/statusReturns: URL, title, viewport width/height, device pixel ratio, scroll position, document height, profile name, port. Use to confirm what viewport the user has set.
viewport — Pin a width, to check a responsive layout
# A phone-sized viewport
curl -s -X POST http://127.0.0.1:${LOOK_PORT} \
-d '{"command":"viewport","width":390,"height":844,"isMobile":true}'
# Hand it back to the real window — ALWAYS do this when finished
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"viewport","width":null}'The browser runs with defaultViewport: null, so the page normally follows the real window
and there is no way to hold a fixed width without this. window.resizeTo does not do it —
it moves the outer frame while innerWidth stays put, so a layout check against it is
measuring nothing.
height defaults to 844 and deviceScaleFactor to 2; isMobile also turns on touch. A
width under 200 is refused.
Ask before using it, and always reset afterwards. It changes what the user sees on their own screen — see the viewport rule below.
Headless — an agent's own browser
The shared window is the user's. When an agent needs to check that a UI works — while building, inside a batch, or as a swarm unit — it runs its own headless Chrome instead. That browser has:
- its own port;
- a throwaway profile, deleted on exit;
- no entry in the registry.
Several can run at once, and none of them can touch the user's window. Whether the UI looks
right — design, layout, wording, a design choice — is not the agent's call. It goes to the
plan's review block (/plan, Batches) with the page's URL, for the user to open in the
shared window themselves.
Start and stop
From the project root, start it in the background with its one stdout line going to a file:
HL_OUT="$(mktemp)" # or a file in your scratch directory
node scripts/puppeteer-server.cjs --headless > "$HL_OUT" # run in the backgroundfor i in 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20; do [ -s "$HL_OUT" ] && break; sleep 0.5; done
HL_PORT=$(node -e "console.log(JSON.parse(require('fs').readFileSync(process.argv[1],'utf8')).port)" "$HL_OUT")
HL_PID=$(node -e "console.log(JSON.parse(require('fs').readFileSync(process.argv[1],'utf8')).pid)" "$HL_OUT")The line reads {"headless":true,"port":57736,"pid":10072,"userDataDir":"…"}. The OS picks the
port; --port <n> asks for one exactly and exits if it is busy, rather than driving someone
else's server. Always stop it when done: kill $HL_PID. That removes Chrome and its profile.
Commands are the same POST requests as the shared window's, to http://127.0.0.1:${HL_PORT}. The
viewport is a fixed 1440×900 until you change it.
Commands only headless should use
These work in both modes, but on the shared window they would move the user's browser, which is never yours to do.
| Command | Body | Returns |
goto | {url, waitUntil?, timeout?} — waits for networkidle2, 30 s by default | {url, status, title}; clears the error log |
click | {selector}, {text}, or both (the text searched within each selector match) | {clicked, matches}, plus coveredBy when another element sat on top |
type | {selector, text, clear?} — appends, or replaces with clear: true | {typed, value, into} |
press | {key} — "Escape", "Enter", "Shift+Tab" | {pressed, focused} |
waitFor | {selector, hidden?, timeout?} — 10 s by default | {ok, ms}, or a 408 on timeout |
errors | {clear?} | {count, console, pageErrors, network} since the last goto |
- Failure is an
errorkey in the response. The new commands also answer with a non-200 status:- 400 — bad input;
- 404 — nothing visible matched;
- 408 — timeout;
- 409 — the element matched but can't be used, e.g. it can't take focus;
- 502 — navigation failed.
- Selectors are plain CSS.
click {text}matches visible text,aria-labelor a button's value, prefers an exact match, and clicks the nearest clickable ancestor. To target one overlay, scope it:{selector: "[role=dialog]", text: "Close"}. errorslists each failed request twice: once innetworkwith its URL and status, and once as Chrome's own "Failed to load resource" console line. Judge fromnetwork.- Set the viewport before
goto. ChangingisMobilereloads the page and loses its state, such as an open dialog.
Checking a changed page
For each changed page, at each width the project checks (desktop, and a phone width such as 390):
viewport, thengotothe page. Expect status 200 and the right title.errors: expect nothing new inpageErrorsornetwork.- Find the changed control the way a person would, by its text or role (
waitFor,click {text}). - Drive the change: open it, fill it,
press Escape,waitFor … hidden: true. If it saves something,gotothe page again and confirm it stuck. - Check for overflow:
{"command":"eval","expression":"document.documentElement.scrollWidth > innerWidth"}should befalse. - Take one
screenshotand look at it for breakage only: a blank page, overlapping or clipped controls, an error screen. A measurement only answers the question you asked. Record a look item for the review block:
- the URL;
- what changed;
- what to try;
- what right looks like;
- the widths;
- what this check confirmed.
Then
kill $HL_PID.
How to reach the app — its base URL, and how a fresh profile gets signed in — is project-specific;
read it from rules."look" (below).
Procedure
When the user says "look at X":
- Resolve the port — run the registry lookup to get
$LOOK_PORTfor this project - Always check status first — call the
/statusendpoint to confirm which page the user is on (URL), the current viewport dimensions, and scroll position. Do NOT assume you know which page or element the user is referencing — the status response is the ground truth. If the user's request references text or an element, confirm it exists on the page shown in the status URL before proceeding. - Inspect the element — use the
inspectcommand with the selector the user mentioned (or derive the selector from their description) - Reason from the data — use the computed styles, box model, and bounding rect to understand the layout. Report findings concisely: what the values are, what they mean for the layout issue
- If you need more context — use
domto see the HTML structure, or inspect parent/sibling elements to understand the layout context - Edit the source — make the CSS/HTML fix in the appropriate source file
- Re-inspect to verify — after the dev server hot-reloads (~1-2 seconds), inspect the same element again to confirm the computed values changed as expected
- Look at it — for a change to how something looks, take one screenshot of the affected region and check the whole area, not just the element you measured. Numbers that are right in isolation can sit beside a neighbour the change broke
Common Inspection Patterns
Overflow detection
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"eval","expression":"const el = document.querySelector(\".container\"); const rect = el.getBoundingClientRect(); JSON.stringify({width: rect.width, viewportWidth: window.innerWidth, overflows: rect.width > window.innerWidth})"}'Flex layout debugging
Inspect the flex container to see flex-direction, flex-wrap, justify-content, align-items, gap. Then inspect individual flex children to see their computed width, flex-basis, flex-grow, flex-shrink.
Spacing issues
Inspect the element — check padding-*, margin-*, gap in computed styles. Check parent and siblings if spacing comes from the container.
Media query state
Use status to get current viewport width, then check what CSS values apply at that width by inspecting the element. The computed styles reflect the active media queries.
Scope queries to the open overlay
Popovers, dialogs and menus usually portal their content to the end of <body>, and a page that renders several instances of a component puts the first one first in document order. A bare document.querySelector then reads a closed or different instance — and can produce a convincing false confirmation. Find the open overlay first and query inside it:
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"eval","expression":"const d = document.querySelector(\"[role=dialog], [data-state=open]\"); d ? d.querySelectorAll(\"button\").length : \"no open overlay\""}'Finding the right selector
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"eval","expression":"document.querySelectorAll(\".hero-section\").length"}'
curl -s -X POST http://127.0.0.1:${LOOK_PORT} -d '{"command":"dom","selector":".hero-section","children":true}'Rules
- No confirmation needed — execute immediately
- DOM-first to diagnose; one screenshot to confirm a visual change — inspect is the default for finding a problem; after changing how something looks, look at it before saying it works. Otherwise screenshot only when asked.
- Re-inspect after edits — always verify your fix by re-inspecting the element after the dev server reloads
- Report concisely — don't dump raw JSON at the user. Summarize the relevant values and what they mean for the issue
- Viewport is the user's — do not navigate the browser, and do not resize it on your own
initiative. The user controls Chrome directly.
goto,click,typeandpressare for your own headless instance only. In an end-of-session review the user opens each item's URL in this window themselves, and tells you which item they are on. The one exception is checking a responsive layout, which cannot be done any other way: ask first, use theviewportcommand, and hand the width back with{"width": null}as soon as you have the answer. Never leave a session pinned to a width the user did not choose.
Project overrides
If .claude/kit.json has a rules."look" entry, read it and apply it as an additional
instruction for this skill. Absent file or key means no overrides — that is the normal case.
It is where a project names what a headless agent needs: the base URL pages are reached through,
how a fresh profile gets signed in, the widths to check, and any page quirks.
jq -r '.rules."look" // empty' .claude/kit.json 2>/dev/nullDoes it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

