---
title: "Claude Cowork Marketing Team Demo: Opus 5.5 Planned, Sonnet Wrote"
date: "2026-09-23"
video_id: "7688558521356504334"
tiktok_url: "https://www.tiktok.com/@theviableedge/video/7688558521356504334"
creator: "@theviableedge"
category: "AI"
tags: ["claude cowork", "opus 5.5", "ai marketing agents", "claude sonnet", "owala"]
excerpt: "A Claude Cowork demo credits Opus 5.5 with a marketing team, but Sonnet wrote the captions and QA passed a draft calling Owala's real Instagram fictional."
---
A TikTok from Adam at Viable Edge (@theviableedge) says "We just built an entire marketing team within Claude Cowork using Opus 5.5." Opus 5.5 did design the workflow, but the run's own log shows a model labelled `sonnet` wrote the caption. The governance step blocked a true award claim, while the QA reviewer gave a "Clean pass" to a draft that called Owala's real Instagram account fictional.

The build uses Owala, the water bottle brand, as a practice client. The draft's header reads "Unofficial demo, not affiliated with Owala" and "Ship target: none (practice build, local draft only)."

## What is Claude Cowork?

[Claude Cowork](https://claude.com/product/cowork) is Anthropic's desktop agent for knowledge work. It reads and writes local files, which is how this build stores its brand context and campaign folders.

## How the marketing team is set up

The orchestrator asks a few setup questions before it builds anything:

- Which specialist roles report to it. The creator checks four: content and repurposing, social and distribution, email and lifecycle, and research and analytics.
- How strict output governance should be. The recommended option is "Draft-only, you approve," ahead of "Tiered autonomy" and "Mostly autonomous."
- How to treat marketing claims found on the brand's site.

In the resulting architecture, a human request goes to the orchestrator, which sends "work order + context packet" to the specialists. Drafts go through a pass or revise loop, then an "approval packet" lands in an approval queue for the human. Each draft carries a paper trail of eight context files (a context card plus brand files for audience, offers, voice, claims and channels, and campaign files) and two cited source pages from Owala's site.

The video description promises "7 specialist agents." The setup screen shows four checked roles plus one free-text option, so seven never appears on screen.

## Did Opus 5.5 write the captions?

No. The log attributes every step that produced text to `sonnet`, in lowercase, with no version number:

> Dispatched  Sep 22  2:19 PM
> Social · sonnet
> Instagram caption 1 of 3 (Amor Y Vida)

> Draft written  Sep 22  2:22 PM
> Social · sonnet
> Instagram caption drafted

Anthropic released Opus 5.5 on September 22, 2026, the same date stamped on every event in the log. [Anthropic's model page](https://www.anthropic.com/claude-opus-5-5) lists it at $4 per million input tokens and $20 per million output tokens, with Terminal-Bench 4.0 at 66.4% and GDPval-AA v2.1 at 1846 Elo.

The current Sonnet that day was Claude Sonnet 5, released June 30, 2026 at $2 per million input and $10 per million output. Anthropic said Sonnet 5.5 and Haiku 5.5 would follow "in the coming weeks," so a Sonnet 5.5 dispatch on September 22 was not possible, which [TechCrunch](https://techcrunch.com/2026/09/22/anthropic-releases-opus-5-5-with-lower-prices-and-fable-level-performance/) also confirms. The log prints no version, so the exact Sonnet model is not knowable from the video. Opus 5.5 built the architecture. A cheaper model wrote the words.

## Was the blocked TIME award claim false?

No. Claude asks how drafts should treat a claim on Owala's site that it shows "TIME Best Inventions." The creator picks "Block until verified" and says:

> Block until verified. This is governance in action. We're not just taking claims that are found online.

TIME's [Best Inventions 2023 collection](https://time.com/collections/best-inventions-2023/6324400/owala-freesip/) has an Owala FreeSip entry in the Fitness category, and Owala's [own blog post](https://owalalife.com/blogs/water-cooler/time-s-list-of-the-best-inventions-of-2023) links to it. One lookup confirms it. Blocking until verified is a sensible default, but it only pays off when a claim is false or stale, and this one is neither.

## How accurate is the drafted caption?

The draft announces the Amor Y Vida colorway, dropping 10/6 on the 32oz FreeSip, made with Yollocalli of Chicago's National Museum of Mexican Art and inspired by Día de los Muertos. It ends with a "Notify me" call to action.

Owala's [Amor Y Vida page](https://owalalife.com/color-drop/amor-y-vida) matches every factual element. It describes the Día de los Muertos inspiration and the Yollocalli partnership, lists a 32oz stainless steel FreeSip at $39.99, and shows a "Notify me" button because the bottle is not yet for sale. The caption tracks Owala's copy closely, so the QA step is checking a restatement more than new claims.

The agent did catch one real problem. The [Color Drop index](https://owalalife.com/color-drop) shows the launch as "10/6 at 10am MT," while the product page shows "Drops 10/6 at 10:00 AM MST." MT and MST are not the same thing in early October, when Mountain Daylight Time is in effect. The agent flagged this as an open question instead of picking one.

## Why the QA "Clean pass" is a problem

The draft's account line reads "Instagram, @owala (fictional posting account for this demo; not[...]", and its assumptions list repeats that the account is fictional. [@owala](https://www.instagram.com/owala/) is Owala's real, verified Instagram account, with roughly 1M followers. The line is cut off on screen, so it may only mean the demo never posts there. Either way, the draft labels a live brand handle as fictional.

The QA reviewer returned:

> QA review · PASS  Sep 22  2:24 PM
> QA Reviewer
> Clean pass

The rubric's claims and facts meters are full. So the claims check blocked a true award and passed a false description of the posting account. The whole chain, from work order at 2:17 PM to the approval queue after the 2:24 PM review, took about seven minutes for one caption.

## Bottom line

Opus 5.5 designed this Cowork marketing workflow, and Sonnet wrote the caption. The governance rules caught a real time zone inconsistency on Owala's site and blocked a true TIME award. The QA reviewer passed a draft that called a live 1M-follower account fictional.

## Sources

- [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) - Release date, pricing and benchmark figures for Opus 5.5.
- [Anthropic releases Opus 5.5 with lower prices and Fable-level performance](https://techcrunch.com/2026/09/22/anthropic-releases-opus-5-5-with-lower-prices-and-fable-level-performance/) - The release date and the pending Sonnet 5.5 and Haiku 5.5.
- [Owala FreeSip: The 200 Best Inventions of 2023](https://time.com/collections/best-inventions-2023/6324400/owala-freesip/) - The TIME award the workflow blocked.
- [Owala FreeSip Water Bottle Makes TIME's Best Inventions 2023](https://owalalife.com/blogs/water-cooler/time-s-list-of-the-best-inventions-of-2023) - Owala's post linking the TIME listing.
- [Amor Y Vida](https://owalalife.com/color-drop/amor-y-vida) - Size, price, MST drop time, "Notify me" and the Yollocalli partnership.
- [Color Drop | Owala](https://owalalife.com/color-drop) - The same drop listed as "10am MT."
- [Marketing plugin | Claude](https://claude.com/plugins/marketing) - Anthropic's first-party marketing plugin, which has no orchestrator or approval queue.
- [Claude Cowork](https://claude.com/product/cowork) - Anthropic's desktop agent that works with local files.
- [Owala (@owala) on Instagram](https://www.instagram.com/owala/) - The live account the draft calls fictional.


## Video description and verification notes

**The audio names four things: Claude Cowork, Opus 5.5, an orchestration model, and Owala. The frames name twelve more that the narration never says out loud, including the ledger ID `Q-20260922-1`, the draft fingerprint `709d22b15be7`, the colorway "Amor Y Vida", the exact file paths `context/brand/claims-and-compliance.md` and `campaigns/2026-09-color-drop-ig/work-order.md`, and the single lowercase word `sonnet` in the chain of custody. The video is 100 seconds at 576x1024, uploaded 2026-09-23 by Adam @ Viable Edge, and at capture it had 1,564 views, 69 likes, 2 comments, 17 reposts, and 61 saves. Fifty frames at 2-second intervals walk through the whole build. Frames 1 to 4 sit on a Claude desktop home screen with a yellow title card reading "Claude Opus 5.5 Built My Marketing Team in Cowork" over a sidebar entry "Marketing team orchestration model / Needs attention · 23 minutes ago" and a mode toggle showing "Chat" and "Cowork". Frames 5 to 7 show the thread running: "Working on it...", then "Thinking... · 5s", then "Reading memory" and "Using memory... · 8s". Frames 8 to 12 show a multi-select card, "Which specialist roles should report to the or[chestrator]", with four options getting checked one by one: "Content + repurposing / Newsletter, YouTube scripts/talking points", "Social + distribution / Drafts, scheduling, and engagement via Ze[...]", "Email + lifecycle / Kit broadcasts, sequences, nurture, taggin[g]", "Research + analytics / Competitor/market intel, KPI digests, perf[...]", plus a fifth free-text row "Something else". Frame 13 shows the governance card "How strict should output governance be?" with "1 Draft-only, you approve (Recommended)", "2 Tiered autonomy", and "3 Mostly autonomous". Frames 15 to 20 show a rendered architecture diagram: Human, request, Orchestrator, then "work order + context packet" to "Specialists", a "pass / revise" return edge, and "approval packet" to "Approval Queue" to "hum[an]". Frames 22 to 25 capture the brand setup, where an assistant reply reads "I'm missing two things from your answers: the br[and name]" and lists "Brand name: I'll use it to name the folder, Do[cuments...]". Frames 26 to 30 hold on the claims card: "Claims: the site shows 'TIME Best Inve[ntions]' ... should drafts treat them?" with option 1 "Block until verified (Recommended) / QA rejects any draft that uses eith[er]". Frame 31 lists four decisions and their answers: Approve, No schedule, Yes set it up, and "Mirror off (Recommended)", under a status line "Ran 5 commands, loaded a skill, read a file · 1 note" and "Using a skill... · 6m 35s". Frames 32 and 33 show an empty "Team Ledger" for Owala reading 0 AWAITING YOU, 0 BLOCKED, 0 IN PROGRESS. Frames 36 to 44 show the filled queue and one draft opened with badges "REVIEWED" and "QA PASS", a "Ship target: none (practice build, local draft only)" line, and a paper trail counting "8 CONTEXT FILES" and "2 SOURCES". Frames 45 to 48 show "PENDING CONTEXT PROPOSALS" naming `2026-09-22-color-drop-pricing.md` and `extraction-notes.md`. Frames 49 and 50 close on a white card: "Get the Cowork Plugin Free Link in Bio".**

## The model in the title built the workflow; a different, unnamed model wrote the words

The narration frames the whole thing as an Opus 5.5 result. The transcript line is:

> We just built an entire marketing team within Claude Cowork using Opus 5.5.

The paper trail in frames 41 and 42 contradicts that for the actual drafting. Every step where text got produced is attributed to `sonnet`, in lowercase, with no version number:

> Dispatched  Sep 22  2:19 PM
> Social · sonnet
> Instagram caption 1 of 3 (Amor Y Vida)
> Tools allowed: read files
> Context packet: 8 files

> Draft written  Sep 22  2:22 PM
> Social · sonnet
> Instagram caption drafted

Opus 5.5 was released on September 22, 2026, the same date stamped on every chain-of-custody event in the frames. Anthropic's model page lists it at $4 per million input tokens and $20 per million output tokens, with Terminal-Bench 4.0 at 66.4% and GDPval-AA v2.1 at 1846 Elo. At that moment the current Sonnet was Claude Sonnet 5, released June 30, 2026 at $2 per million input and $10 per million output. Anthropic said Sonnet 5.5 and Haiku 5.5 would follow "in the coming weeks," so a Sonnet 5.5 dispatch on September 22 was not possible. The frame does not print a version, so which Sonnet ran the caption is not determinable from the video. The architecture is the Opus 5.5 output. The captions are not.

## The award claim the system blocked is true, and one search confirms it

The pivotal governance moment is a question Claude asks about Owala's own marketing copy. Frame 26 shows it verbatim:

> Claims: the site shows 'TIME Best Inve[ntions]' ... should drafts treat them?
> 1  Block until verified (Recommended)
>    QA rejects any draft that uses eith[er]
> 2  Allow as shown on site
>    OK to repeat word for word, citing

The narrator picks option 1 and says:

> Block until verified. This is governance in action. We're not just taking claims that are found online.

The claim itself holds up. TIME's Best Inventions 2023 collection includes a dedicated Owala FreeSip entry in the Fitness category, on time.com, describing the stainless-steel bottle's two-way spout. Owala's own blog post links to it. So the policy blocked a statement that is accurate and confirmable from the publisher's primary page in a single lookup. The rule is cheap insurance rather than a save. It is worth naming the distinction: "block until verified" only pays off when the claim is false or stale, and this one is neither.

## The approved caption is a close paraphrase of Owala's product page, and the time zone the agent flagged is a real site inconsistency

Frame 37 shows the drafted caption. The legible portion reads:

> A new Color Drop is coming, and it's about celebrating who — a[nd]
> Amor y Vida drops 10/6 at 10am MT on the 32oz FreeSip, inspir[ed by] traditions of Día de los Muertos. Made in partnership with Yolloc[alli] of Chicago's National Museum of Mexican Art — this one-of-a-ki[nd] Chicago's young Latino artists sharing their own connection to t[he]
> A vibrant reminder to cherish the moments — and people — that[...]
> Don't miss it. Tap "Notify me" so you're first in line. Link in bio.
> #ColorDrop #AmorYVida #Owala #FreeSip #NotifyMe

Owala's live product page for Amor Y Vida describes the bottle as "Inspired by the joyful colors and traditions of Día de los Muertos and made in partnership with Yollocalli—a youth arts initiative of Chicago's National Museum of Mexican Art," calls it a vibrant reminder to celebrate life and love by cherishing the moments that matter most, lists it as a 32oz FreeSip in stainless steel at $39.99, and carries a "Notify me" button because it is not yet purchasable. Every factual element in the draft matches. The phrasing tracks the source copy closely enough that the QA pass is checking a restatement rather than a claim.

The one thing the agent caught is real. Frame 39 shows an open question:

> Exact time zone abbreviation: [site says] ... MT," p[...] MST" — used "MT" here to [...]

Fetching both Owala pages reproduces that discrepancy: the Color Drop index renders the launch as "10/6 at 10am MT" while the Amor Y Vida product page renders "Drops 10/6 at 10:00 AM MST". MT and MST are not interchangeable in early October, when Mountain Daylight Time is in effect. The agent noticed a genuine inconsistency on the brand's own site and escalated it instead of picking one.

## "Clean pass" was awarded to a draft that calls a real 1M-follower account fictional

Frame 37's Platform / Account block reads:

> Instagram, @owala (fictional posting account for this demo; not[...]

and the Assumptions list in frames 38 and 39 repeats it:

> Fictional posting account @owal[a] used only to give the demo[...] actually checked or posted to.

@owala is Owala's real, verified Instagram account, with roughly 1M followers and a bio reading "⭐️ Where Boring Ends. 🙌 FreeSip® (yes, we patented that) 👜 We're in the bag biz". The line is truncated in frame, so the intended meaning may be "not actually posted to during this demo" rather than "this handle does not exist." Either reading leaves the same problem: the draft labels a live brand handle "fictional" while the same draft's header says "Unofficial demo, not affiliated with Owala." Frame 42 shows what the reviewer returned:

> QA review · PASS  Sep 22  2:24 PM
> QA Reviewer
> Clean pass

The QA rubric in that frame shows bar meters for brief, channel, cla[ims] and [fac]ts, all filled. A claims-and-facts check that blocked a true TIME award and passed a false description of the posting account is inverted on exactly the axis the video is selling.

## The numbers the frames print and the audio never says

The narration describes the paper trail in general terms: "Here's a list of the context files that were used. External sources cited." The frames print the specifics.

The draft carries the identifier `Q-20260922-1` and the version fingerprint `709d22b15be7`, and frame 43 shows the approval dialog gating on that exact string: "Approve this exact version (fingerprint 709d[22b15be7]". The paper trail tab is labeled "Paper trail 5" and counts "8 CONTEXT FILES" and "2 SOURCES". The eight files carry tier labels T0, T1, and T2:

> T0  context/context-card.md
> T1  context/brand/audience.md
> T1  context/brand/offers.md
> T1  context/brand/voice.md
> T1  context/brand/claims-and-compliance.md
> T1  context/brand/channels.md
> T2  campaigns/2026-09-color-drop-ig/work-order.md
> T2  campaigns/2026-09-color-drop-ig/research/next-color[-drop...]

The two sources cited are `https://owalalife.com/color-drop (fetched 2026-09-22)` and `https://owalalife.com/color-drop/amor-y-vida (fetched 2...)`, matching the two pages that actually carry the drop date, size, partnership and CTA. Tools used is a single chip: "read files". The chain of custody runs Work Order at 2:17 PM, Dispatched at 2:19 PM, Draft written at 2:22 PM, QA review PASS at 2:24 PM, then Queued for approval, so one caption moved from order to approval queue in about seven minutes.

Frame 45 shows the self-learning claim the narrator makes at the end. The pending proposals are `2026-09-22-color-drop-pricing.md`, annotated "Result: N/A (research finding, not a post-p[...])", and `extraction-notes.md`, reading "Sources fetched sequentially via WebFetch. https://owalalife.com". The pricing finding is real and relevant: the page lists $39.99. It was captured as a context proposal and kept out of the caption.

## Key Takeaways

- The narration credits Opus 5.5 for the whole build, but the ledger attributes both the dispatch and the draft to `sonnet`. Opus 5.5 designed and stood up the workflow on the day it shipped. It did not write the captions in this run.
- The blocked claim, Owala's TIME Best Inventions 2023 listing, is verifiable in one lookup against time.com. Blocking until verified is a defensible default, not a catch in this instance.
- The agent flagged a real inconsistency on Owala's site: the index page shows "10/6 at 10am MT" and the product page shows "10:00 AM MST".
- A "Clean pass" QA review approved a draft describing @owala, a live account with about 1M followers, as a "fictional posting account". The claims-and-facts check missed the one factual error in the draft while stopping a true statement.
- Transcription and visual gaps: Whisper wrote "Owalla" for Owala and "time's best inventions" for the on-screen "TIME Best Inve[ntions]". The colorway "Amor Y Vida", the price context file, the fingerprint `709d22b15be7`, the ID `Q-20260922-1`, the tiered context paths, the word `sonnet`, the "Ship target: none (practice build, local draft only)" line, and the "Unofficial demo, not affiliated with Owala" disclaimer appear only in the frames. Whisper also rendered "draft-only human approved" from an on-screen option that reads "Draft-only, you approve (Recommended)". The video description claims "7 specialist agents"; the frames show four checked roles plus one free-text entry, so seven is not visible anywhere on screen.
- Unverified: the plugin itself. The narrator says "Let's use our marketing plugin" and "You can grab this plugin for free. Just use the link in my bio," and frame 50 shows "Get the Cowork Plugin Free Link in Bio". The TikTok profile page and bio link would not render through WebFetch, so the plugin's identity, publisher, and license are unconfirmed. Anthropic does publish a first-party Marketing plugin at claude.com/plugins/marketing, but its documented commands are `/draft-content`, `/campaign-plan`, `/brand-review`, `/competitive-brief`, `/performance-report`, `/seo-audit`, and `/email-sequence`, with no orchestrator, work orders, approval queue, or Team Ledger, so the plugin in this video is a different, third-party one. Also unverified: which Sonnet version ran the drafts, since the frame prints no version; the truncated option text "engagement via Ze[...]" and "Kit broadcasts"; the full wording of the caption beyond the right edge of the 576px frame; and whether the Team Ledger sparkline axis actually stops at 9/14 or continues past the cropped right edge.

## Resources

- [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) - Confirms the September 22, 2026 release date, the `claude-opus-5-5` model ID, $4/$20 per million token pricing, and the Terminal-Bench 4.0 and GDPval-AA benchmark figures.
- [Owala FreeSip: The 200 Best Inventions of 2023](https://time.com/collections/best-inventions-2023/6324400/owala-freesip/) - Confirms the TIME award claim that the workflow blocked until verified.
- [Owala FreeSip Water Bottle Makes TIME's Best Inventions 2023](https://owalalife.com/blogs/water-cooler/time-s-list-of-the-best-inventions-of-2023) - Owala's own post linking the TIME listing, the on-site claim the agent refused to repeat unverified.
- [Amor Y Vida](https://owalalife.com/color-drop/amor-y-vida) - Confirms the 32oz FreeSip, $39.99, "Drops 10/6 at 10:00 AM MST", the "Notify me" CTA, and the Yollocalli and National Museum of Mexican Art partnership.
- [Color Drop | Owala](https://owalalife.com/color-drop) - Shows the same drop rendered as "10/6 at 10am MT", the inconsistency the agent flagged, plus Book Club, Midnight Manor, Coastal Christmas 11/10 and Autumn Ridge Road 9/1/26.
- [Marketing plugin | Claude](https://claude.com/plugins/marketing) - Anthropic's first-party marketing plugin, whose command list contains no orchestrator or approval queue, ruling it out as the plugin shown.
- [Claude Cowork](https://claude.com/product/cowork) - Confirms Cowork is Anthropic's desktop knowledge-work agent that reads and writes local files, matching the Documents/Claude paths on screen.
- [Anthropic releases Opus 5.5 with lower prices and Fable-level performance](https://techcrunch.com/2026/09/22/anthropic-releases-opus-5-5-with-lower-prices-and-fable-level-performance/) - Independent confirmation of the release date and that Sonnet 5.5 and Haiku 5.5 were still pending at the time of this recording.
- [Owala (@owala) on Instagram](https://www.instagram.com/owala/) - The live brand account the draft describes as fictional.
