<- all tokdocs
2026.10.15@itsjakethewizard

Claude Code's /loop Exists, but the Form Shown Self-Paces Between One Minute and One Hour and Can End Itself

AI
Watch on TikTok

View on TikTok ->

The /loop command in this video is real and shipped in Claude Code 2.1.71 on 2026-03-06, but the form typed on screen (no interval) runs in self-paced mode, where Claude picks each wait between one minute and one hour, prints the reason, and can end the loop on its own once it judges the task done, so "wait a few minutes and then do it again until I tell you to stop" misstates both the cadence and the stop condition. The artifact is a 76.54-second vertical clip, 1080x1920, in an MP4 container (ffprobe reports mov,mp4,m4a,3gp,3g2,mj2), HEVC Main video tagged hvc1 at 1,740,415 bps and 30 fps across 2,295 frames, HE-AACv2 audio at 44,100 Hz stereo and 48,109 bps, 1,794,280 bps overall, 17,167,672 bytes on disk. TikTok's metadata puts the post at 2026-10-15 18:04:50 UTC from uploader itsjakethewizard, channel nickname "Jake | The Marketing Wizard". At capture on 2026-10-18 the counters read 64,200 views, 2,786 likes, 52 comments, 1,233 reposts, and 4,823 saves. I read all 38 extracted frames (2-second intervals) and kept 210 transcript words across 16 Whisper segments; the audio ends at 76.59 seconds, and the 8 segments from 75.70 seconds onward are Whisper tail output with invented text ("free bit moreила", "AI Seaten and FAIL"), which I discarded. Every frame shares one setup: a bearded presenter in black-rimmed glasses and a gray patterned short-sleeve shirt, seated behind a silver broadcast microphone on a boom arm, with framed film posters on the wall behind him. Frame 1 carries a title card reading "AUTOMATE YOUR SEO: THE CLAUDE CODE LOOP". Bold one-word captions track the narration (ONE, OPEN, GIVE, LOOP, SLASH, TECHNICAL, CHANGELOG, PDF, MINUTES, AGAIN, DEPLOY, CONVERSATION, and the closing "THAT'S WHERE I WOULD START WITH AI"). Frames 2 through 17 hold a dark terminal graphic titled "claude code - my-website" (rendered here with a hyphen) that fills in line by line: "✻ Welcome to Claude Code", "> /loop", "SEO-analyze a page of my website.", "Everything to change: technical, text, tracking.", "Write a changelog, save it as a PDF.", and "Mark the page done. Go to the next one." Frames 20 through 25 retitle the window "claude code - /loop running" and print "✓ /home → changelog-home.pdf", "⏳ waiting a few minutes...", "✓ /services → changelog-services.pdf", and "✻ looping until you say stop". Frames 27 through 29 show a white checklist card headed "YOUR WEBSITE" with EVERY PAGE, EVERY PRODUCT, EVERY SERVICE, and OPTIMIZED ONE BY ONE ticking green in sequence. Frame 32 reads "YOU CAN DO IT PAGE BY PAGE" over "RESULT".

/loop shipped in March 2026, seven months before this video

The video says "open up Claude code, okay, give it your website, and then say slash loop." The command exists. The Claude Code changelog records it under version 2.1.71: "Added /loop command to run a prompt or slash command on a recurring interval (e.g. /loop 5m check the deploy)". The npm registry's publish time for 2.1.71 is 2026-03-06T22:47:03Z. The video went up on 2026-10-15, so the feature was seven months old by then. The current release as I write is 2.1.303, published 2026-10-18, and 20 lines of the changelog between those versions mention /loop, most of them fixes to how loops survive backgrounding, resume, and process restarts.

Anthropic documents the command on the scheduled tasks page rather than on the slash-command list, which does not mention it. That page describes /loop as a bundled skill and summarizes it as "the quickest way to run a prompt on repeat while the session stays open." I also pulled the skill definition out of the Claude Code 2.1.278 binary installed on this machine. Its description string reads "Run a prompt or slash command on a recurring interval (e.g. /loop 5m /foo). Omit the interval to let the model self-pace."

The prompt as typed runs self-paced, so the cadence and the stop condition both differ from the narration

The on-screen prompt never gives an interval. It reads "/loop" followed by "SEO-analyze a page of my website." and three more lines. The docs cover that case in a three-row table: with a prompt and no interval, "Your prompt runs at an interval Claude chooses each iteration." The rule for choosing it: "After each iteration it picks a delay between one minute and one hour based on what it observed: short waits while a build is finishing or a PR is active, longer waits when nothing is pending. The chosen delay and the reason for it are printed at the end of each iteration."

So "wait a few minutes" is one possible outcome inside a one-minute to one-hour range that Claude sets per iteration. A fixed cadence requires an interval token. /loop 10m followed by the prompt creates a cron job, and the docs add that a /loop 5m job starts each run "4 minutes 45 seconds after the previous run", while other intervals run late "by up to half the gap to the next run, and by at most 30 minutes" because Claude Code adds jitter to spread API load across sessions.

"Until I tell you to stop" misses three documented ways a loop ends without the user. In self-paced mode, "Claude can also end the loop on its own once the task is complete" by calling ScheduleWakeup with stop: true. If an iteration ends without rescheduling or stopping, Claude Code "schedules one fallback wakeup about 20 minutes later and ends the loop when that iteration doesn't reschedule either." Every recurring task also expires 7 days after creation. The bundled skill text in the 2.1.278 binary gives the model the same instruction: stop when "the task is complete, further iterations can't make progress, or the user asked you to stop", and "Stopping is the loop's normal ending".

Two more constraints matter for an SEO crawl that could take hours. Tasks "only fire while Claude Code is running and idle. Closing the terminal or letting the session exit stops them firing." A self-paced loop "isn't restored" on --resume, so a closed laptop means typing /loop again. For a job meant to finish when every page is done rather than when a timer fires, the docs point to /goal, which re-runs turns until a separate model judges a stated condition met. Its comparison table lists /loop as the option whose next turn starts when "A time interval elapses."

The terminal in the video is an illustration, and its last line contradicts the docs

The dark window titled "claude code - my-website" types out the prompt one phrase at a time in sync with the narration and never shows a reply from Claude. The "running" view prints checkmarks, an hourglass emoji, and the line "✻ looping until you say stop". A real fixed-interval /loop responds by confirming "the cadence and job ID", per the docs, and the 2.1.113 changelog entry says a self-paced iteration displays as "Claude resuming /loop wakeup". Neither appears in the graphic. The line "looping until you say stop" states the opposite of the documented self-paced behavior, in which Claude is instructed to stop once the task is complete or further iterations can't make progress.

The filenames changelog-home.pdf and changelog-services.pdf are props as well. The tools reference lists every built-in tool, and none writes a PDF. Write "Creates or overwrites files", and the only PDF handling on the page is on the read side, where page-range reads of long PDFs "render pages with pdftoppm from poppler-utils". Getting a PDF out of this loop means Claude runs a converter through Bash, which has to be installed on the machine first, and Bash is marked as requiring permission on the same page, so each conversion prompts unless the session is in auto mode or the command is allowlisted.

"Mark the page done" lives in the session transcript unless Claude writes it down

The prompt's bookkeeping step, "Mark the page done. Go to the next one.", has no storage of its own. Scheduled prompts fire "between your turns" inside the same session, so Claude knows which pages it already covered only as long as that conversation context survives. Nothing in the docs persists that state across a closed terminal, and the self-paced loop itself is not restored on resume. A durable version of this workflow would have Claude keep a checklist file in the project so a restarted loop can read where it left off. The docs' own example of a goal condition, "Working through a labeled issue backlog until the queue is empty", is the same shape of task with the state held in an external queue.

The working directory also decides what "give it your website" means. The mock title bar shows a folder named my-website, which is the scenario where Claude can read templates and page source directly. For a site that exists only on a hosted CMS, Claude reaches pages through WebFetch, which the docs call "lossy by design" because an extraction prompt, rather than the raw HTML, is what Claude receives. The narration's "it will optimize each one individually" also overstates the typed prompt, which asks for "everything to change" and a changelog; the video itself defers applying changes to "a whole nother conversation."

Who is making the claim

The channel checks out. The @itsjakethewizard profile on 2026-10-19 showed the display name "Jake | The Marketing Wizard", 111.7K followers, 620 following, 636.7K likes, and the bio "Hire a wizard, not an agency! 2x Exits | Marketing Consultant, Author & Speaker Formerly thewizardmarketing". The video's caption summarizes the workflow accurately as far as it goes ("audits one page, writes a changelog pdf, marks it done and moves to the next page until you stop it"). The gaps are in what /loop does between and after those steps.

Key Takeaways

  • Verified: /loop is a bundled Claude Code skill, added in 2.1.71 (published 2026-03-06) and documented on the scheduled tasks page; the current release is 2.1.303 (2026-10-18).
  • Correction: with no interval, as typed in the video, Claude picks each delay between one minute and one hour and prints the reason, so "wait a few minutes" is one case inside a documented range; /loop 10m followed by the prompt is the form that fixes the cadence.
  • Correction: a self-paced loop can stop itself when Claude judges the task complete, falls back to a single 20-minute wakeup if an iteration neither reschedules nor stops, expires after 7 days, stops when the terminal closes, and is not restored on resume.
  • Clarification: the terminal window is a graphic; the documented confirmations (cadence and job ID, or "Claude resuming /loop wakeup") do not appear, and "looping until you say stop" contradicts the skill's own stop instruction.
  • Clarification: no built-in tool writes PDFs; Write creates text files, and PDF output requires a converter run through Bash.
  • Context: for "finish when every page is done", the docs' /goal command stops on a model-verified condition rather than a timer.

Resources

Published October 15, 2026. Writeup generated from a favorited TikTok.