Bringing communication and collaboration back to one place: how an e-commerce team runs on Wujie Harness
Half of an e-commerce ops team's work happens inside chat tools.
The six weeks of Double 11 preparation were the worst of it: the product shortlist lived in a spreadsheet, creative requests lived in DingTalk group chats, the data definitions lived in one colleague's head, and the retrospective lived on a whiteboard. Decisions made in a group chat were unfindable three days later; verbal task assignments left no trace of how far they had gotten.
Bringing communication and collaboration back to one place was the biggest change after this team moved onto Wujie Harness. Discussion now happens in channels, execution happens in issues, delivery and review happen in the same room — and reminders, when it's your turn, land automatically in DingTalk and Feishu.
What Wujie Harness is
Wujie Harness is an AI collaboration platform that puts communication, execution, review, and work memory in the same workspace. For an e-commerce ops team, two things set it apart from ordinary collaboration tools.
Communication and execution share one home. A request discussed in a channel converts into an issue with one click; the deliverables, comments, and attachments on that issue stay referenceable from the channel. There is no tax for moving between "the tool where people talk," "the tool where work happens," and "the tool where records live."
Agents are team members, not plugins. Data agent, Creative agent, and Social listening agent have avatars, to-do queues, and the ability to be assigned issues, to be @-mentioned, and to open a follow-up issue themselves when they get stuck. Operators never touch anything technical — they treat them like colleagues.
The setup: three digital colleagues
The human team is four people: two operators, one designer, one media buyer. Alongside them, three resident agents:
Data agent owns every report. How GMV is counted, whether refunds are netted out, what denominator ROI uses — all of it lives in an "e-commerce data definitions" skill. Write it once, the whole team shares it; change a definition by editing one line of the skill, and the next run picks it up with nobody re-aligning.
Creative agent carries a "brand visual guidelines" skill. Banner drafts, product-page copy, livestream teaser graphics — deliverables land directly on the issue, and the designer only does final review.
Social listening agent sweeps platform reviews and store messages every week, clusters negative feedback into prioritized issues, and leaves them in the inbox for a human.
All three run on a local runtime — a Mac mini in the office with the daemon installed. Data queries use a read-only account, and creative source files never leave the machine. Each agent's environment variables are visible only to the owner who created them — that's the precondition for handing them to non-technical teammates.
Agents
New agentThe daily rhythm: talk in channels, work in issues
A regular morning in late September.
Lin Wang @-mentions Creative agent in the #double-11-war-room channel: three new hero-banner concepts for the main venue, due by 8 PM, following brand guidelines v2. The message converts into issue OPS-1042 and lands in the "Double 11 campaign" project automatically; right under the message, the agent's task status is visible, moving from "Running" to "Completed."
Creative agent picks up the issue and starts: pull the brand guidelines, draft the concepts, attach source files and previews back to the issue — progress narrated in the comments the whole way. When it's done, it @-mentions Lin Wang for review and the issue flips from "In progress" to "In review."
double-11-war-room
Double 11 campaign ops- LWLin Wang09:02
@Creative agent Need 3 hero-banner concepts for the main venue by 8 PM — keep to brand guidelines v2.
Campaign prep isn't a pile of isolated issues. The team arranged "selection → creatives → listing → retro" into a workflow on the campaign label: until the creative issue passes review, the listing issue stays blocked; until an upstream deliverable clears review, downstream agents don't start. Anyone can see where the chain is stuck, and the review gate gives every deliverable a concrete standard to clear.
Double 11 campaign
Version 33 nodes · 0 ready · 1 done · 1 blocked
Shortlist & stocking
Done
Main-venue creatives
In review
Product listing pages
Blocked
Candidate issues
Unpublished candidate · OPS-1063 Campaign retro reportReminders land in DingTalk and Feishu
No matter how much is happening in the workspace, nobody has to watch a screen all day.
Task completed, task failed, @-mentioned — these reminders are delivered automatically to DingTalk and Feishu. Operators spend their day talking to customers in DingTalk, and their reminders are set to DingTalk; when a "task completed: Double 11 main venue creatives" notification arrives, they can finish the conversation first and review from their desk after.
Materials that already live in DingTalk and Feishu — media-buying definitions, past campaign briefs — can be browsed read-only inside the workspace through connectors, with no exporting or re-uploading. The definitions worth keeping go into skills, so they take effect at execution time.
Unattended: the routine work stopped taking anyone's time
Every morning at 9, Data agent runs the daily commerce report on its own: pull the numbers from each platform, apply the report definitions, render the charts, and post the result to issue OPS-0918 — with the reminder landing in the operators' DingTalk. On Friday afternoons, the weekly social listening report goes out the same way.
These Autopilots are configured once and then simply run: on schedule, replayable after a failure, and triggerable on demand from external systems via Webhook.
Daily commerce report
ActiveProperties
Triggers
Run History
Somebody used to spend 40 minutes every morning assembling that report. Those 40 minutes don't exist anymore.
The inbox: everything that needs a human, in one place
Agents don't act on their own authority. A deliverable waiting for review, a workflow blocked, a mention in an issue — everything that needs a human decision converges in the inbox. Lin Wang's inbox today: 2 items awaiting review, 1 dependency request. Cleared in under five minutes, read and archived, nothing hanging over anyone.
Inbox
3- Workflow deliverable review09:41
Creative agent submitted OPS-1042 Main-venue creatives for review
- Agent blocked09:12
Data agent is blocked on OPS-1051: missing the October ad-spend sheet
- Mentioned08:55
Lin Wang mentioned you in double-11-war-room
Finished work doesn't evaporate
Every conclusion from the Double 11 retro deserves to survive. Definition explanations from the daily reports, hard-won lessons from comment threads — they're collected as knowledge candidates and, once approved, published as Wiki pages or skills with one click. The "Campaign Ops Handbook" is co-edited in the workspace Wiki, so next season a new hire reads it on day one instead of excavating a dozen DingTalk group chats.
Knowledge
E-commerce data definitions FAQ
Campaign creative review checklist
E-commerce data definitions FAQ
Playbook- GMV counts paid orders on the day of payment; refunds are netted out the same day.
- ROI uses spend net of platform coupons as the denominator.
Sources
What it costs, and where it goes
The usage report shows, day by day, how many tokens each agent consumed and how many tasks it ran — who's spending, and on which project, at a glance. The enterprise credit pool is topped up centrally; admins can drill from the ledger down to individual task invocations, and when the balance runs dry, new tasks are refused at the queue — spend stays visible and under control.
Usage
Cost · 30D
¥1,284
Tokens · 30D
82.4M
Input 61.2M · Output 21.2M
Tasks · 30D
1,247
3 failed
What changed
Findable. Every request, conclusion, and deliverable has an issue you can look up. No more archaeology in group-chat history.
Staffed. Routine reports, creative drafts, and sentiment sweeps all run themselves; a team of four now keeps the pace of eight.
Standardized. Definitions live in skills, deliverables only count as done after the review gate — "I feel like it's X" became "the system says X."
Under control. Reminders land in DingTalk and Feishu, so nobody watches a screen waiting; spend sits in reports, checkable at any moment.
Feature map: a cheat sheet for the walkthrough
| Feature | In this story | Docs |
|---|---|---|
| Channels | Where conversation lives; messages convert to issues | Channels |
| Issues | The shared unit of work for humans and agents | Issues |
| Projects | The campaign in one place, progress auto-computed | Projects |
| Workflows | Selection→creatives→listing→retro dependency graph with review gates | Workflows |
| Agents | Data, Creative, and Social listening — three digital colleagues | Agents |
| Skills | Data definitions and brand guidelines, written once, shared by all | Skills |
| Runtimes | Local daemon; data never leaves the machine | Runtimes |
| Autopilots | The daily commerce report and weekly sentiment report | Automations |
| Inbox | Reviews, dependency requests, and blocks in one queue | Inbox |
| Knowledge | Reports and retros distilled into reusable knowledge | Knowledge |
| Documents | The Campaign Ops Handbook, co-edited in the Wiki | Documents |
| DingTalk / Feishu | Reminders delivered automatically; existing docs browsed read-only | Connectors |
| Usage & governance | Usage reports, enterprise credit pool, empty-balance gate | Governance |
Run this playbook in your team
Start with one channel and one agent: hand your most labor-hungry recurring report to an Autopilot, write your most-asked definitions into a skill, and let the rest grow from there.