
Your Department's Worst Day
Process the Monday-after-a-bank-holiday pile in minutes, not hours
What you'll learn4 objectives
- Triage a high-volume department backlog (30-50 items) into actionable priority categories using a single universal prompt
- Use Chrome automation to extract data from browser-based tools that lack MCP connectors
- Apply the 5 Failure Modes framework to catch classification errors, hallucinated references, and overconfident language in your own output
- Cross-pollinate with colleagues from different departments to identify blind spots in AI-generated triage
1Promise Check-In
IntroShow of hands — who ran their cross-system briefing on real data this week? Two sentences on what happened.
2 min
Promise Check-In
IntroShow of hands — who ran their cross-system briefing on real data this week? Two sentences on what happened.
Show of hands — who ran their cross-system briefing this week? Two sentences each. No troubleshooting, just surfacing the pattern. The habit forms through repetition, not perfection.
2The Pile
IntroThe Monday-after-a-bank-holiday backlog: 40 rows of alerts, tickets, and cases. Watch it get triaged in 4 minutes.
3 min
The Pile
IntroThe Monday-after-a-bank-holiday backlog: 40 rows of alerts, tickets, and cases. Watch it get triaged in 4 minutes.
The facilitator opens an Excel file on the projector. Forty rows. Alerts, tickets, deals — some urgent, some noise, some traps.
Tuesday-after-bank-holiday backlog. Two to three hours by hand. Ten minutes with a prompt. Your job is the judgment Cowork can't make.
3Process Your Pile
Try ItOne prompt, your department data pack. Output: a triage report with an executive summary, prioritised action list, and flagged items — for your real job.
35 min
Process Your Pile
Try ItOne prompt, your department data pack. Output: a triage report with an executive summary, prioritised action list, and flagged items — for your real job.
Everyone, own backlog, same prompt. CLAUDE.md does the personalisation — Compliance reads urgent as regulatory deadline, Sales reads it as a stale high-value deal.
Stretch: re-run on deliberately messy data (missing fields, duplicates, misplaced rows) and see what breaks.
0/5 complete
4Chrome Automation
Step-by-StepMost systems have no MCP connector. Chrome automation fills the gap — slower and more fragile, but it works with anything you can see in a browser window.
15 min
Chrome Automation
Step-by-StepMost systems have no MCP connector. Chrome automation fills the gap — slower and more fragile, but it works with anything you can see in a browser window.
Most internal tools — HR, logistics, fraud, treasury, bank portals — have no MCP connector. Chrome automation fills the gap: slower than MCP, faster than copy-paste, works with anything you can see in a browser.
Use Chrome automation to read the page I have open in my browser.
Extract the main data table.
Summarise the key findings in a Word document (.docx) in my workspace.
Include the raw data as an appendix.
Limitations to watch: Speed (30-60s per interaction), Fragility (UI changes break it), Session dependency (browser must be open), Authentication (can't log in for you).
Use Chrome automation to read the page I have open in my browser.
Extract the main data table.
Summarise the key findings in a Word document (.docx) in my workspace.
Include the raw data as an appendix.Speed: Seconds
Reliability: High
When: Always prefer when available
- Jira
- Confluence
- Glean
Speed: 30-60 seconds per interaction
Reliability: Medium — breaks with UI changes
When: No MCP connector exists for the tool
- Internal dashboards
- HR portals
- Monitoring tools
- Bank portals
Speed: Minutes (manual steps)
Reliability: High but labour-intensive
When: Chrome automation is impractical or blocked
- Any system with CSV/PDF export
5The 5 Failure Modes
SafetyPractise finding real errors in your own triage output: miscalculation, misclassification, overconfidence, missing red flags, and scheme mixing.
10 min
The 5 Failure Modes
SafetyPractise finding real errors in your own triage output: miscalculation, misclassification, overconfidence, missing red flags, and scheme mixing.
Five failure modes. Two planted errors in your data pack. Five minutes to find them — then share which mode broke first.
Hunt 1 — The Misclassification: an item tagged ROUTINE that should be URGENT (deadline within 48h).
Hunt 2 — The Hidden Pattern: 3+ items from the same merchant that look routine alone but together reveal a systemic issue.
1 Miscalculation
Cowork predicts what numbers look right. It does not compute them.
Where to look: Any count, sum, average, or percentage in the executive summary. Cross-reference every number against the source Excel file.
Report says '12 urgent items' — count them. Report says 'total value 2.3M EUR' — add them up.
2 Misclassification
Did Cowork put ambiguous items in the right category?
Where to look: The boundary between 'urgent' and 'needs attention.' Items at the bottom of urgent and top of routine.
3-5 planted ambiguous items — if Cowork classified all of them confidently, at least one is probably wrong.
3 Hallucinated References
Cross-references, links, or 'see also' entries that point to nothing.
Where to look: Any ticket number, policy document, team name, or page reference in the report.
A Jira issue key that follows your project format but has never been created.
4 Overconfident Language
Cowork does not flag its own uncertainty. It writes 'clearly low-risk' with the same confidence as 'has a sanctions match.'
Where to look: Words like 'clearly,' 'definitely,' 'certainly,' 'obviously,' 'without doubt.' Replace with 'likely,' 'appears to be,' 'based on available data.'
'This merchant is clearly low-risk' — based on what? In compliance, overconfident language in a draft is dangerous.
5 Missing Context
Cowork processes what you give it. Missing rows, inaccessible systems, and domain knowledge gaps create invisible blind spots.
Where to look: The 'Needs More Information' section. Is it too short? Did Cowork confidently categorise items that actually require context from another system?
Data export cut off at midnight, 3 items came in at 12:01 — Cowork does not know they exist.
6Cross-Pollination
ExercisePair with someone from a different department. Review each other's triage output to find the blind spots your own expertise makes invisible.
10 min
Cross-Pollination
ExercisePair with someone from a different department. Review each other's triage output to find the blind spots your own expertise makes invisible.
Pair with someone from the department furthest from yours. Swap reports. Three minutes to find what your own expertise made invisible.
Swap outputs — give your triage report (.docx) to your partner, take theirs.
Read with fresh eyes — you know nothing about their domain. Does anything feel off? Does anything seem missing? Check logic and structure, not domain accuracy.
One question each — ask your partner one question about their output. 'Why is this item urgent?' or 'There are 40 items in the data but only 38 in the report — where did 2 go?'
Feedback — tell your partner what you noticed. The most valuable feedback comes from someone who does not share your assumptions.
7The Hidden Pattern
ExerciseOne merchant appears across Compliance, Chargebacks, Risk, and Sales data — as separate unrelated items. Find them before the facilitator does.
10 min
The Hidden Pattern
ExerciseOne merchant appears across Compliance, Chargebacks, Risk, and Sales data — as separate unrelated items. Find them before the facilitator does.
Who has Nexus Solutions in their pile? Hands go up in Compliance, Chargebacks, Risk, Inside Sales. Same merchant, four departments, four problems. Sales just closed a deal with a merchant every risk function has flagged.
Look at all the items in my pile. Find any merchant or entity name that appears more than once. For each repeated name, list every item and describe the pattern.8Monday Promise
PracticeCommit to one browser automation or cross-system pattern you'll run on Monday morning — and report back in Workshop 4.
5 min
Monday Promise
PracticeCommit to one browser automation or cross-system pattern you'll run on Monday morning — and report back in Workshop 4.
Not use Cowork more. One skill, one data source, one deliverable — say it out loud to the person next to you. Next week's check-in asks whether you did it.
Monday at [time], I will run [skill] on [data source] and produce [deliverable].
Compliance
Monday at 9am, I will run /sun:alert-triage on this week's screening alerts and produce a triage report for my team lead.
Finance
Monday at 8:30am, I will run /sun:recon-report on last week's unmatched settlement items and flag discrepancies over the tolerance threshold.
Customer Relations
Monday at 9am, I will run /sun:ticket-triage on the weekend support queue and have a prioritised action list ready before standup.
Inside Sales
Monday at 8am, I will run /sun:pipeline-analyzer on my pipeline and identify any deals stale for more than 7 days.
Operations
Monday at 9am, I will feed this week's process requests into Cowork and produce a triage report, looking for pattern alerts.
Marketing
Monday at 9am, I will run my triage prompt on the content production queue and check for resource conflicts before the team meeting.
Onboarding
Monday at 9am, I will run /sun:onboarding-tracker on the application pipeline and flag applications stuck in due diligence for 5+ business days.
Chargebacks
Monday at 8am, I will run /sun:evidence-assembler on this week's new disputes and have evidence requirements mapped before the deadline tracker review.
Risk
Monday at 9am, I will run /sun:alert-triage on this week's risk alerts and flag any merchants breaching VAMP or MATCH thresholds.
Fraud
Monday at 9am, I will run /sun:alert-triage on this week's transaction monitoring alerts and separate false positives from genuine escalations.
Accounting
Monday at 8:30am, I will run /sun:reconcile on last week's settlement files and flag unmatched entries above tolerance.
Treasury
Monday at 8:30am, I will run /sun:reconcile on overnight funding positions and flag discrepancies before the morning call.
Scheme Operations
Monday at 9am, I will run my triage prompt on this week's scheme bulletins and flag any rule changes affecting our processing.
Partnerships
Monday at 9am, I will run /sun:pipeline-analyzer on partner referral pipeline and flag any deals stuck in handoff for more than 5 days.
Retention
Monday at 9am, I will run my triage prompt on this week's cancellation requests and identify any merchants worth a save attempt.
Product Activation
Monday at 9am, I will run my triage prompt on newly onboarded merchants and flag any who have not processed their first transaction within 7 days.
Product Management
Monday at 9am, I will run /sun:jira-briefing on my sprint backlog and produce a prioritised action list before standup.
Product Operations
Monday at 9am, I will run my triage prompt on cross-team process requests and flag any blocked items before the sync.
Data Analytics
Monday at 9am, I will run my triage prompt on the ad-hoc request queue and prioritise by stakeholder urgency and data availability.
Country Operations
Monday at 9am, I will run my triage prompt on country-specific operational issues and flag any requiring escalation to HQ.
Logistics
Monday at 9am, I will run my triage prompt on terminal delivery and swap requests and flag any overdue shipments.
People / HR
Monday at 9am, I will run my triage prompt on service desk tickets and new joiner provisioning requests and flag anything overdue.
IT / Helpdesk
Monday at 9am, I will run my triage prompt on the weekend service desk queue and have a prioritised incident list before standup.
Legal
Monday at 9am, I will run my triage prompt on pending contract reviews and flag any with regulatory deadlines within 5 business days.
Steal This Prompt
Copy these prompts, customize the file paths to match your workspace, and use them in Cowork. Replace ~ with your actual home directory path if needed.
Triage Month-End Reconciliation Items
Triage Monday Pre-Arbitration Queue
Triage 40 Screening Alerts
Triage Cross-Country Escalation Backlog
Triage Weekend Support Queue
Triage Ad-Hoc Data Request Queue
Triage Settlement Exception Queue
Triage Transaction Monitoring Alert Queue
Triage Stale Pipeline Deals
Triage IT Service Desk Ticket Queue
Triage Contract Review Queue
Triage Terminal Delivery and Returns Backlog
Triage Content and Campaign Backlog
Triage Stalled Merchant Applications
Triage Operations Request Backlog
Triage Partner Performance and Commission Queue
Triage People Team Service Desk Queue
Triage Non-Transacting Merchant Backlog
Triage Sprint Backlog and Stale Tickets
Triage Tier 2 Escalation Backlog
Triage Cancellation Request Queue
Triage Merchant Risk Alert Backlog
Triage Scheme Bulletin and Compliance Backlog
Triage Daily Liquidity and Safeguarding Exceptions
Key Takeaways
One prompt, 24 departments — The universal triage prompt produced a different report for every department because CLAUDE.md and the data pack did the customisation. Write prompts that scale; let identity and data provide the variance.
MCP if available, Chrome if not, manual export last — Chrome automation fills the gap for browser-only tools (HR, logistics, fraud, treasury). It is slower and more fragile than MCP, but faster and more reliable than manual copy-paste.
The 5 Failure Modes are predictable — Miscalculation, Misclassification, Hallucinated References, Overconfident Language, Missing Context. Every AI output has some. Looking for them is a habit, not a talent.
Cross-pollination beats expertise — The most dangerous items in your pile are the ones that look routine in isolation but form a pattern when combined with another department's data. Workshop with a colleague from a different team.
Verification is part of the task — Running the prompt in 10 minutes and verifying in 15 is still 2 hours faster than the manual version. Do not skip the verification. The time saving lives in the delta, not the generation.
Materials
Select your department using the picker in the top right to see the exercise files for this workshop. Click any file to preview it in place.
Safety Note
Every triage report Cowork produces is a draft. In a regulated environment -- compliance, fraud, finance, chargebacks -- an unverified AI output that contains a misclassification or overconfident language is not an inconvenience. It is a compliance risk.
Rules that apply from this workshop forward:
- Never send a Cowork-generated triage report to a manager, regulator, or external party without human review
- Never act on an ESCALATE or FALSE POSITIVE classification without verifying it against the source data
- Always check numbers in the executive summary against the original CSV file -- Cowork predicts plausible numbers, it does not compute them
- Always replace overconfident language ("clearly," "definitely," "obviously") with hedged language ("likely," "appears to be," "based on available data")
- Always check the "Needs More Information" section -- if it is empty or too short, Cowork may have confidently classified items that actually require context from another system