review-inbox
Skill2 starsChanged 16 days agoMITView on GitHub
Post work a human has to judge by eye, ear or feel (UI and layout changes, visuals, animation, game feel, effects, sound, copy) to the project's review inbox with a clip or still, wait for their OK or Redo, and act on the verdict. Also where questions for the human go. Use whenever you finish such a change or need the human to decide something, instead of asking in chat or writing "needs a human check" in a notes file, and whenever a verdict or answer from the inbox reaches you.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- HsuanYuLee/polaristheir own version, first seen Nov 12, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- HsuanYuLee/polaris★ 5same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 0 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
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
- Very specific: line 25
Inspect before you post, every time: Read the contact sheet (clip_sheet.jpg) or the still, and check it really shows what you are about to …
- Very specific: line 32
Write --how in calm, plain prose: where to look, what changed, what you want judged.
- Very specific: line 42
Wait: Run review.py wait as a background command (run_in_background) and end your turn or keep working.
- Very specific: line 49
REDO: read the note and any attached images (list and wait print their paths), rework it, and post the next take in the same turn.
- Very specific: line 75
If a tool is missing, the SessionStart hook says so with the install command.
≈ “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
Review inbox
The human reviews your visual and feel work on one page (http://127.0.0.1:7700/ by default), shared by every session and
project (each item is labelled with its project), and answers
OK or Redo with a note and sometimes images. Every command below is python review.py ..., with review.py in this
skill's directory. The SessionStart hook prints the full command line: use exactly that one, since it can carry
--session <id>, which ties your items to you in agents that don't set CLAUDE_CODE_SESSION_ID (Codex, others).
Subagents: a background subagent posts with its parent's --session and should add --by "<its description>" (the Agent call's description) to add/ask: the page lists subagents under their session and counts their open items by that label.
The loop
- Capture. Show it, don't describe it: anything that moves or makes sound gets a video. Render it to a file; don't
film the human's screen, they may be watching something on it.
- A web page or browser game:
review.py clip 6 --url http://localhost:8080/(headless) orreview.py shot --url ...for a still. - A game engine or native app: its own recorder, which the project sets as
REVIEW_INBOX_CLIP_CMD;clipruns it. - Anything you rendered yourself (an offline render, a still from your own tool): pass it to
--media. - An interaction (a button press, typing, a hover): add steps and
clipfilms them; see below. - Recording the real screen or a window (
clip --window "<title>",clipwith no--url) only works when the human has setREVIEW_INBOX_DESKTOP=1; otherwise it refuses. Only one recording runs at a time: a second one waits.
- A web page or browser game:
- Inspect before you post, every time. Read the contact sheet (
clip_sheet.jpg) or the still, and check it really shows what you are about to claim: the moment is in frame, the right thing changed, nothing is broken, it isn't black. If not, retake it. The human should only ever see footage you have already checked.- The sheet is 3 frames a second, so anything shorter than ~0.3 s (a flash, a hit, a pop) can fall between its frames.
For those,
review.py frames <clip.mp4> --at <seconds>tiles every frame of the 0.6 s around that moment (--span,--fpsto change it). Read that instead.
- The sheet is 3 frames a second, so anything shorter than ~0.3 s (a flash, a hit, a pop) can fall between its frames.
For those,
- Post.
review.py add --title "<what it is>" --topic <area> --how "<how to get there and what to judge>" --media <files>- Write
--howin calm, plain prose: where to look, what changed, what you want judged. Claim only what the footage shows. - Several takes go in one
add(--media a/clip.mp4 b/clip.mp4), each path once, each named for what it is (renameclip.mp4tobigger-particles.mp4first). Put your proposal first and comparisons last. - Label every clip and still on the picture itself, in a corner: what it shows and, once there is more than one,
whether it is the change or a comparison (
AFTER: bow climb, first person,BEFORE: old climb, for comparison). The human sees the footage before the file names and must never mistake an old take for your fix. If your recorder can't draw text, burn it in:ffmpeg -i in.mp4 -vf "drawtext=text='AFTER\: bow climb':x=12:y=12:fontsize=h/18:fontcolor=white:box=1:boxcolor=black@0.6:boxborderw=8" -c:a copy out.mp4. Skip a BEFORE take unless the change is subtle enough to need one. - Show only what is new. Don't re-post variants the human already decided on.
addstarts the page if it isn't running and pops a desktop notification, so you don't need to alert the human.
- Write
- Wait. Run
review.py waitas a background command (run_in_background) and end your turn or keep working. It exits with the verdict when the human answers. It exits 3 after about 45 idle minutes (sized for a one-hour prompt cache): then don't re-arm it. The verdict reaches you with the next prompt instead.- No
run_in_backgroundin your agent? Start it detached and read its output later:nohup <cli> wait > wait.log 2>&1 &in bash, orStart-Process -NoNewWindow -RedirectStandardOutput wait.log <python> "<review.py> ... wait"in PowerShell. Never run it in the foreground: it can block for most of an hour.
- No
- Act on the verdict. It is work, not a notification, even when it arrives as a background-task notice.
- REDO: read the note and any attached images (
listandwaitprint their paths), rework it, and post the next take in the same turn. The retake is a new item (add): the REDO'd item stays as history, and your newer item is what satisfies the Stop hook. - OK: finish it (commit if your project's rules allow), then
review.py handled <id> "committed <hash>". - OK with a note (or images): the take is approved and the note is more work. Commit the approved state first, then
do what the note asks before moving on: rework and post a new take, or start the work it asks for. Then
review.py handled <id> --did "<what you did about the note>". A commit alone doesn't close it. - Truly need the human's answer before you can go on?
review.py handled <id> "<why>", then ask (below).
- REDO: read the note and any attached images (
The Stop hook keeps this honest. It blocks the end of your turn while a verdict on your item is unread, while a REDO or OK
has neither a newer item from you nor a handled mark (for an OK with a note, handled --did), or while you have an open item and no wait running.
Filming an interaction
Give clip (or shot) steps with --do "a; b; c", or --script <file> with one step per line. The recording runs
until the steps are done (plus a short lead-in and tail), or at least the seconds you give.
Steps: click <target>, dblclick <target>, move <target>, type <text>, key <combo>, scroll <dy>, wait <ms>.
- Web page: targets are Playwright selectors. A cursor dot is drawn so the viewer sees what was clicked.
review.py clip --url http://localhost:8080/ --do "click text=Settings; click #volume; key ArrowRight; wait 500" - Desktop app (only with
REVIEW_INBOX_DESKTOP=1): targets arex,ypixels. Take a still first, Read it and pick the coordinates from it:review.py shot --window "My App"and thenreview.py clip --window "My App" --do "click 320,180; type hello; click 40,386; wait 800". Coordinates are relative to the window's top-left as in that still (Windows; elsewhere use the whole screen's). The real mouse and keyboard move, so the human's machine is busy for those seconds.
If a tool is missing, the SessionStart hook says so with the install command. review.py doctor lists what's missing,
what it disables and how to install it; offer the human the install.
Asking the human
The human steers through the inbox and may never see a question you ask in chat; your session then sits idle until its
cache goes cold. So put every question for them in the inbox, not just ones about how something looks: a go-ahead, a
choice of approach, "build X now or wait for Y?". An app's closing next-step box (a wrap_up ask with buttons) counts as
chat too, and so is the AskUserQuestion pop-up (a hook turns it away): keep that box to a recap and ask here:
review.py ask "Should the trail match the ball colour or stay cyan?" --options "Match the ball" "Stay cyan"
Options become one-click buttons (keep them short), and the human can always type an answer instead. Then run
review.py wait in the background like for a review. The answer comes back like a verdict: carry on with it. The Stop hook
pushes back once if your turn ends on a question with nothing posted, whether or not you've used the inbox before.
Other commands
review.py list [--open] [--topic <area>]: what's open, and verdicts with their notes.review.py revoke <id>: take down your own unanswered item (posted the wrong take). Then post the right one.review.py dismiss <id> "<why>": close an open item that no longer needs an answer (superseded, already fixed).
The inbox lives in ~/.review-inbox, outside your project. Answered items lose their media after 7 days
(REVIEW_INBOX_KEEP_DAYS), so attach nothing you need to keep. If a project keeps its own (REVIEW_INBOX_DIR pointing inside
it), that folder git-ignores itself, so it never lands in a commit.
Does 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.

