@angusthenontechnical scores vending sites as night headcount times dead hours, and the agent in his own frames reports the job ads never state headcount

The artifact is a 55.70 s portrait clip (ffprobe: format.duration=55.703991, video stream 55.700000 s), 1080x1920 SDR, mov,mp4,m4a,3gp,3g2,mj2 container, HEVC video (codec_name=hevc, yt-dlp vcodec=h265, format_id=bytevc1_1080p_1146428-1) at 1,091,516 bps and r_frame_rate=30/1 across 1,671 coded frames, AAC audio at 48,164 bps across 1,203 frames, 1,146,449 bps overall, 7,982,727 bytes on disk. Posted at epoch 1792609750, which is 2026-10-21T19:09:10Z. Uploader angusthenontechnical, uploader_id 6783510017912865797, channel name "Angus the Nontechnical"; the audio credit in metadata is track: original sound, artist: Angus the Nontechnical, so no licensed music is attached. The description is 90 characters: "I have been on vending machine TikTok thought this was a fun one #vibecoding #buildwithai ". At the capture time recorded by the mtime of metadata.json, 2026-10-26 00:41:53 local, the counters read 65,200 views, 6,089 likes, 48 comments, 660 reposts and 3,532 saves. I read all 28 extracted frames (frame_001 through frame_028, 2 s apart) and the full 229-word, 13-segment transcript. The frames move through seven distinct surfaces: at 0 s and 2 s a white lower-third card reading "How to exploit a ghost town" over stock footage of a hand fanning banded one-dollar bills next to a snack machine; at 4 s and 6 s an UpFlip sales page with an orange "ENROLL NOW" button, a red ellipse around "THE most important module", the heading "MODULE 3: Finding, evaluating, and purchasing the most profitable vending locations", body copy promising a "30-30 rule" and routes that "easily generate over $100k/year before you even get your first machine", and a teal pill reading "or $997" circled in red; at 12 s a dark agent client whose breadcrumb reads "Threads > DEMO - The 2am vending map", carrying a prompt that instructs the model to "Actually search and read their public job ads, using available Apify or browser tools where possible. Extract employer, exact facility address, job title, shift hours, work arrangement and physical/desk work, with source URLs and short evidence quotes. Use Indeed plus official careers pages; if blocked, say so and use official pa[ges]", a plan list with "Grainger + Dollar Tree official careers pages read" checked and "Post compact result" pending, a status line reading "Spinning 7m 9s", an input placeholder "Steer this thread…" and a model chip reading "Opus 5.5 Medium"; at 14 s an Indeed search for "warehouse night" in "joliet, il" returning "Sortation Center Warehouse Associate, Amazon Warehouse, $24 an hour" and "Warehouse Associate Nightshift, Berkots Super Foods, Joliet, IL 60433, $18 - $22 an hour"; at 16 s through 22 s Google Maps, first a results list for "warehouses near Joliet Illinois", then an 800 m circle annotated "800 m radius / Straight-line search area" around "Dollar Tree Distribution Center", then a Ryder Systems - Joliet card at "3451 S Chicago St, Joliet, IL" closing 3:30 PM; at 28 s two stacked Maps cards both reading "2401 S Chicago St, Joliet, IL 60436", one captioned "Restaurant closes 9 p.m." and one "Convenience store open 24 hours" with the site foodnfuel.net and phone (815) 724-0066, under a monospace footnote "Check every food option"; at 32 s through 40 s a flowchart of white cards with monospace footnotes "Verify both inputs first" and "Illustrative numbers", a black banner reading "Estimate based on job board**", and the arithmetic N night headcount times 6 dead hours resolving to "480 / worker-hours"; at 42 s the agent's finished answer, "I collected 3 facilities and 6 night-shift ads with quotes and source links. The full tool-call log is in the file. Everything stops before scoring, so Jev can label next", bulleted findings for "Grainger, 701 Grainger Way, Minooka", "Dollar Tree DC, 300 Dollar Tree Lane, Jo[liet]" and "Panduit, Lockport", a caveats paragraph beginning "Things to know: Indeed's search pages loaded but [employer] pages were blocked (401), Exa is out of credits", and a file chip "night-shift-ads.md updated now"; at 44 s and 46 s a phone email thread titled "Night shift food", from "me Oct 21" "to Minooka facilities", opening "Hi there, I saw your Minooka warehouse ad for the 7 p.m. night shift" and signed "Thanks, Angus"; and at 48 s through 54 s the creator alone in a room delivering the follow CTA. The spoken argument is that a paid vending course sells location selection as its core secret, that an AI agent can reproduce that selection by reading night-shift job ads, mapping the employers, counting food options within 800 m and their closing times, and ranking sites by night headcount times hours without food, and that the operator then cold-emails the winner. The transcript's final segment ends at 55.600 s against an audio stream of 55.703991 s, closing 0.104 s inside the audio with no segment timestamped past the end; the narration is speech-dense throughout and shows none of the Whisper end-of-audio fabrication pattern, so I quote it directly below.

"Vibe coded" is the hashtag, and no code appears in any of the 28 frames

At 0.00 s the video opens with "turns out these vending machine businesses can be vibe coded", and the description carries #vibecoding and #buildwithai. Vibe coding describes generating software you do not write yourself. Nothing in this clip produces software. Across all 28 frames the surfaces are an UpFlip marketing page, an agent chat client, Indeed's web search, Google Maps place cards, a motion-graphic flowchart built in post, and a phone email client. There is no terminal, no editor, no file tree, no repository, no diff, no build output and no running program. The one artifact the agent produces is a Markdown file, shown in the attachment chip at 42 s as "night-shift-ads.md", which the agent itself describes as a log: "The full tool-call log is in the file."

What the frames show is an agent doing read-only research with browser and search tools, then writing notes. The prompt at 12 s is explicit about it, asking the model to "search and read their public job ads" and "Extract employer, exact facility address, job title, shift hours". The gap matters because the narration's payoff at 50.00 s, "go have Claude feed some truckers", implies an automated pipeline, while the evidence on screen is one chat thread that ran for a few minutes and produced a memo a person still has to act on.

"Scrape job boards" versus the agent's own report of a 401 and an exhausted search API

At 12.80 s the narration says "you just need your AI to scrape job boards of all places". The agent's finished answer at 42 s contradicts the verb twice over. Its caveats paragraph reads "Things to know: Indeed's search pages loaded but [employer] pages were blocked (401), Exa is out of credits", and its plan checklist at 12 s shows the item that did complete as "Grainger + Dollar Tree official careers pages read". The prompt had already anticipated this, instructing "Use Indeed plus official careers pages; if blocked, say so and use official pa[ges]". So the run read Indeed search result pages, hit an HTTP 401 on the pages behind them, exhausted its third-party search credits, and fell back to employers' own careers sites. That is retrieval with a documented failure, not scraping.

The permission picture is also narrower than the word "scrape" suggests. Indeed's live robots.txt blocks named scrapers outright, including Scrapy, JobdiggerSpider, ia_archiver and MyCentralAIScraperBot, and disallows /jobs, /viewjob, /q- and /l- for the training crawler group that contains GPTBot, CCBot, anthropic-ai and ClaudeBot. It separately allows retrieval agents including GoogleBot, ChatGPT-User, PerplexityBot and Claude-User. An assistant fetching a page on a user's behalf sits in the allowed group; a crawler harvesting listings in bulk sits in the blocked one. Indeed's terms at indeed.com/legal are also narrower than secondary write-ups claim: the clause I could verify, section A.3.5 "Limits", states "Use of any automation, scripting, or bots to automate the Indeed Apply process outside of Indeed's official vendors and tooling is prohibited", which governs automated applying rather than reading listings. I found no verifiable blanket anti-scraping clause on the page I fetched, so I assert none.

The course price on screen is $997, the narration says "a thousand dollars", and UpFlip's live page sells something else

At 4.16 s the narration asks "Have you ever seen one of these courses and wonder what do I actually get for a thousand dollars?", and the frame at 4 s circles a teal pill reading "or $997" in red. Rounding $997 to "a thousand dollars" is fair. What that $997 buys is not established by anything on screen or by any source I could reach.

UpFlip's live course page, The Vending Bootcamp, names Adam Hill as the instructor and displays no price at all. It structures the curriculum as Steps 1 through 8, not Modules, and the closest item to the "30-30 rule" the frame promises is a "30/30 Service Area Worksheet" filed under Step 3. The page's headline revenue figure is "$58k/month", not the "$100k/year" and "$1M/year" the captured frame shows. The sales page the video filmed is therefore not the page UpFlip serves today, either because it is a different funnel page or because the copy was rewritten. The checkout page that search results point to, course.upflip.com/vending-checkout, did not resolve at all (DNS ENOTFOUND), so I cannot cite it. Search results attach $997 to several different UpFlip and Hill Vending items, including a yearly community membership and an in-person training, and attach $2,999 to the main course. I could not trace $997 to a first-party page, so what the badge prices remains unverified. The instructor the frame calls "Adam" is Adam Hill, named on UpFlip's own bootcamp page, so that attribution holds even though the current page does not carry the "biggest regret" sentence.

The score is night headcount times dead hours, and the agent says the ads do not contain the headcount

At 32.64 s the narration says "count the number of people that work there every night", at 36.00 s "multiply it by how many hours there is no food", and at 37.20 s "whichever one has the highest score is the best location". The flowchart spells it out as N night headcount times 6 dead hours equals 480 worker-hours, and the video is honest about the status of those numbers in its own burned-in text: "Verify both inputs first" under the inputs, "Illustrative numbers" under the result, "Estimate based on job board**" in a banner at 32 s.

The deeper problem is that the first input is not available from the stated source. The agent's answer at 42 s says, of the Panduit listings, "None of the ads I read gives a stre[ngth]" and, in the caveats, "the open ads says nothing about how [many people work there]". The prompt at 12 s had flagged the risk in advance, instructing the model to infer "headcount from vacanc[ies]" and then "stop bef[ore scoring]". So the ranking variable N is the one number the retrieval step reports it could not obtain. A job posting states a shift window and a wage; it does not state how many people are on the floor at 2 a.m. Counting open vacancies measures hiring velocity, and a fully staffed facility posts nothing at all. The on-screen 480 is arithmetic, not a measurement, which is what "Illustrative numbers" says in six-point type while the voiceover calls it a score.

The agent's handoff line names the gap as deliberate: "Everything stops before scoring, so Jev can label next." Jev is a real model, not a mangled transcription. TypeSafe's documentation describes System One models as "a class of AI models built to make fast, structured decisions that software can use directly", returning "typed answers and probabilities" through choice, score and boolean primitives rather than prose, and names Jev as "TypeSafe's flagship model and the first System One model". Handing scoring to a typed-decision model is a coherent design. The clip never shows that step running, so the ranking the narration promises is demonstrated nowhere in the 55.7 seconds.

The 800 m food check lands on one street address that is open 24 hours

At 23.68 s the narration says "we wanted to look within 800 meters of each business see how many food options are there", and the frame at 26 s draws the circle around the Dollar Tree Distribution Center labelled "800 m radius / Straight-line search area". The label is more careful than the narration: straight-line radius is not walking or driving distance, and for a facility ringed by rail yards, Route 66 and the Des Plaines River the two diverge sharply. A worker on break has minutes, not a 1,600 m round trip through a truck gate.

The example the video then uses undercuts the "dead hours" count. Both Maps cards at 28 s carry the same address, "2401 S Chicago St, Joliet, IL 60436". One is captioned "Restaurant closes 9 p.m." and described as a "Local, counter-serve fast-food chain, for fried & grilled sandwiches, hot dogs & burgers". The other is captioned "Convenience store open 24 hours", with the website foodnfuel.net and phone (815) 724-0066. Those are the same building: the address resolves to Route 66 Food N Fuel, a 24-hour truck stop reachable at (815) 724-0066, which houses a Pop's Italian Beef counter. At the one site the video walks through, the number of hours with no food available is zero, and the 9 p.m. closing it shows is a counter closing inside a store that never closes. Google Maps lists that counter as closing at 9 p.m. on weekdays while several listing aggregators show 8:30 p.m., a reminder that the closing times this method depends on are crowd-maintained and inconsistent.

The monospace footnote in that frame, "Check every food option", is the right instruction, and the video's own screenshot shows why: counting place cards without resolving them to physical sites turns one address into two data points with opposite answers.

Every named facility, employer and model in the clip is real and correctly spelled

Grainger's Minooka facility is at 701 Grainger Way, Minooka, IL 60447, matching the agent's bullet exactly. Grainger's own careers site lists "Warehouse Associate - Night Shift" and "Warehouse Associate - Day & Night Shift" at Minooka with posting dates of Oct 7, 2026, which is precisely the claim the agent made at 42 s: "It's live [on Grainger's] own site, posted Oct 7." The Indeed card at 12 s shows "$21.75 - $22.95 an hour" for that role, and listings for the Minooka night shift describe 7:00pm-5:30am and 8:00pm-6:30am windows on a four-day, ten-hour schedule with a $1 per hour overnight differential, matching the agent's "7pm-5:30am or 8pm-6:30am". I sourced the titles and dates to the careers search page I fetched; I did not open an individual job page, so the exact hour windows rest on aggregators rather than first-party text.

Dollar Tree's Joliet distribution center is at 300 Dollar Tree Lane, Joliet, IL 60436, again matching the agent's bullet, though Dollar Tree's own page for the site returned HTTP 504 and the address rests on independent directories. Panduit's Lockport plant is confirmed on Panduit's own contact page at 16530 W. 163rd Street, Lockport, Illinois 60441, with headquarters separately at 18900 Panduit Drive, Tinley Park, IL 60487. Two of the three facilities therefore sit outside Joliet even though the Maps query at 16 s reads "warehouses near Joliet Illinois": Minooka is in Grundy County and Lockport is roughly ten miles north. The area the method searched is wider than the query string implies.

The model chip reads "Opus 5.5 Medium". Anthropic's live documentation lists Claude Opus 5.5 with API ID claude-opus-5-5, a 1M-token context window, 128K max output, $4 per million input tokens, $20 per million output tokens, and a default effort of medium, so the chip is accurate on both counts. The prompt's reference to Apify and the caveats' reference to Exa both name real tools; I did not fetch Exa's pricing or account pages, so its credit exhaustion is the agent's claim rather than a checked fact.

Key Takeaways

Correction: The clip is tagged #vibecoding and opens at 0.00 s with "these vending machine businesses can be vibe coded", but no code is written, run or shown in any of the 28 frames. The only artifact produced is a Markdown research memo, night-shift-ads.md, which the agent itself calls a tool-call log.
Correction: At 12.80 s the narration says the AI will "scrape job boards". The agent's own on-screen report at 42 s says Indeed's deeper pages returned 401, its third-party search credits were exhausted, and it fell back to Grainger's and Dollar Tree's official careers pages. Indeed's robots.txt blocks named scrapers such as Scrapy and MyCentralAIScraperBot and blocks training crawlers from /jobs and /viewjob, while allowing retrieval agents including Claude-User.
Correction: The ranking formula needs N, the night headcount. The agent reports that "the open ads says nothing about how [many people work there]" and that none of the Panduit ads give it. The 480 worker-hours on screen is labelled "Illustrative numbers" by the video itself while the voiceover presents it as a score.
Correction: The 800 m food check at 28 s shows a 9 p.m. restaurant and a 24-hour convenience store as two separate findings, but both cards read "2401 S Chicago St, Joliet, IL 60436". They are one building, Route 66 Food N Fuel, which is open 24 hours. The dead-hours count at that site is zero, not the gap the method assumes.
Clarification: The on-screen label says "800 m radius / Straight-line search area". Straight-line radius is not walking distance, and around a distribution center bounded by rail, highway and river the two differ enough to change the answer.
Verified: Grainger at 701 Grainger Way, Minooka, IL 60447, with night-shift warehouse associate listings posted Oct 7, 2026 on Grainger's own careers site; Panduit's Lockport plant at 16530 W. 163rd Street, Lockport, Illinois 60441, on Panduit's own contact page. Both addresses match the agent's bullets exactly.
Verified: Claude Opus 5.5 is current, with API ID claude-opus-5-5, a 1M-token context window, $4 per million input tokens, $20 per million output tokens, and a default effort of medium, so the "Opus 5.5 Medium" chip is accurate on both counts.
Verified: Jev is real and correctly named, not a Whisper mangling. TypeSafe documents it as "TypeSafe's flagship model and the first System One model", a class that returns "typed answers and probabilities" rather than prose.
Partial correction: UpFlip's live Vending Bootcamp page names Adam Hill as instructor but shows no price, uses Steps 1 to 8 rather than the frame's "MODULE 3", headlines "$58k/month" rather than "$100k/year", and lists a "30/30 Service Area Worksheet" rather than a "30-30 rule". The filmed page is not the page UpFlip serves today.
Unverified: The "or $997" badge at 4 s. UpFlip's bootcamp page carries no price, course.upflip.com did not resolve (DNS ENOTFOUND), and search results attach $997 to at least two different UpFlip and Hill Vending products.
Unverified: The agent's claim that "Exa is out of credits". Exa is a real commercial search API, but I did not fetch its pricing or account pages.
Context: Dollar Tree's own Joliet DC page returned HTTP 504 and Indeed's /legal/terms path returned 404 during this check. Nothing above is cited from a page that did not load.
Context: The transcript is clean. Its last segment ends at 55.600 s against an audio stream of 55.703991 s, with no segment timestamped past the end, so every quotation above is genuine speech.

Resources

The Vending Bootcamp, UpFlip establishes the instructor name Adam Hill, the Step 1 to 8 structure, the "30/30 Service Area Worksheet" under Step 3, the "$58k/month" headline, and the absence of any displayed price, all of which differ from the page captured at 4 s and 6 s.
Indeed robots.txt establishes the named scraper blocks (Scrapy, JobdiggerSpider, OmniExplorer_Bot, ia_archiver, MyCentralAIScraperBot), the /jobs, /viewjob, /q- and /l- disallows for training crawlers, and the allowance for retrieval agents including Claude-User.
Indeed legal terms establishes the only automation clause I could verify, section A.3.5 "Limits", which covers automating the Indeed Apply process rather than reading listings.
Grainger careers, Minooka establishes the "Warehouse Associate - Night Shift" and "Warehouse Associate - Day & Night Shift" listings with Oct 7, 2026 posting dates, corroborating the agent's "posted Oct 7".
Panduit contact and locations establishes the Lockport plant at 16530 W. 163rd Street, Lockport, Illinois 60441 and the Tinley Park headquarters at 18900 Panduit Drive.
Claude models overview, Anthropic establishes Claude Opus 5.5 as current, with API ID claude-opus-5-5, 1M-token context, 128K max output, $4 and $20 per million tokens, and medium default effort, matching the chip in the frames.
System One, TypeSafe docs establishes that Jev is TypeSafe's flagship System One model and that such models return typed answers and calibrated probabilities instead of prose.
Angus the Nontechnical on Substack establishes the creator's publication behind the "follow your boy" CTA at 52 s.