Lab 02 · Ticket-to-Resolution Writing Pipeline

Work three messy ticket threads end-to-end with AI doing the first draft at every stage — pickup summary, L2 escalation note, resolution notes — and ship each artifact with your correction diff against the raw draft. The diff is what gets graded, because the diff is the job.

Arjun picks up tickets on a US-healthcare client desk in Pune, where every escalation note is read by an L2 engineer eight time zones away who was not on the call and will not re-read the thread. Since Copilot arrived in the ITSM tool, drafts are free — and so are the mistakes inside them: a summary that repeats the user's wrong guess as fact, an escalation note missing the one environment detail L2 needed, a resolution note that documents what the AI assumed rather than what actually fixed it.

This lab makes the correction pass explicit. You will generate every draft with AI, then mark exactly what you changed and why. If your diff is empty on all three threads, you shipped the model's mistakes.

Time & tier

~60 minutes hands-on · Core tier (maps to module M2). $0 — any free AI chat plus a text editor. Keep two files per thread: tN-draft.md (untouched AI output) and tN-final.md (your corrected version).

Setup

The three threads below are your inbound queue. Read nothing beyond the thread you are working.

Thread 1 — VPN, escalation candidate:

[T-2201] From: megan.hollis@client.example  Subj: VPN broken AGAIN
09:02 MEGAN: VPN is broken again. Third time this month. I have a
patient-billing deadline today. Fix it or give me someone who can.
09:15 L1 (previous agent): Asked for error message and screenshot.
09:40 MEGAN: It says error 720 something. Also my Outlook is slow but
that's probably unrelated. I already reinstalled the VPN client twice,
that's what your team told me last time and it worked.
10:05 L1: Suggested reinstalling the VPN client.
10:22 MEGAN: I JUST SAID I did that twice. Did you even read my message?
It connects for about 10 seconds then drops. My home internet is fine,
Netflix works all evening.
10:48 L1: Asked if she is using wifi or cable.
11:03 MEGAN: Wifi. Oh also - maybe relevant - I got a new router from my
ISP on Saturday. The old one died. Everything else works on it.
11:20 L1: Note - remote user, Windows 11 23H2, GlobalProtect 6.2.1.
Account not locked, no other VPN complaints in queue today. Escalating.

Thread 2 — printer, resolve it:

[T-2214] From: derek.wall@client.example  Subj: cannot print invoices
08:31 DEREK: The invoice printer in AP does not print from the ERP.
Prints fine from Word. ERP support says it is a network problem.
08:50 L1 (you): Asked which printer and which ERP screen.
09:04 DEREK: HP LaserJet AP-PRN-02, from the invoice batch screen.
Regular prints work. Batch prints just vanish - no error, nothing.
09:31 L1: Checked print server - AP-PRN-02 online, queue empty. Found
batch jobs from ERP arrive with paper size A4; printer tray configured
Letter-only, driver set to "hold mismatched jobs" - 14 jobs held.
09:47 L1: Changed ERP print profile to Letter, released held jobs,
all 14 printed. Derek confirms. Asked Derek to retest tomorrow's batch.
10:15 DEREK: All good. By the way the network here is still terrible,
whole floor agrees, you should look at that too.

Thread 3 — access, resolve it:

[T-2226] From: lina.ortiz@client.example  Subj: locked out of reports
14:02 LINA: I cannot open the finance reports site since the weekend.
Says access denied. My teammate Josh can open it fine. IT changed
something, I did not change anything.
14:20 L1 (you): Verified account active, password fine. Site permission
is via AD group FIN-REPORTS-RO.
14:41 L1: Lina is not in FIN-REPORTS-RO. Group membership audit shows
she was removed Saturday by the quarterly access recert job - her
manager never responded to the recert email, so access auto-expired.
15:05 L1: Re-add requires manager approval per policy. Sent approval
request to her manager, cc Lina, explained the recert rule.
15:38 Manager approved. L1 re-added Lina to FIN-REPORTS-RO, she
confirmed the site opens. Advised manager to respond to future recerts.

Guided steps

Generate the pickup summary for each thread. Prompt: "Summarise this support thread for an engineer picking it up cold. Three sections: Situation (2 lines), Already tried (bullets, only things actually done), Next action (1 line). Do not include anything not stated in the thread." Save all three as tN-draft.md. Expected output: three summaries of under 120 words each.
Audit each summary line-by-line against the thread. The classic AI failures here: listing something suggested as something tried, repeating the user's self-diagnosis ("network problem", "IT changed something") as fact, and dropping the detail that arrived late and casually. Thread 1 contains exactly such a detail — check whether your draft kept it. Mark every fix in tN-final.md with [was: ...] inline so the diff is visible.
Draft the L2 escalation note for Thread 1. Prompt the AI for an escalation note with sections: Environment · Symptom (precise) · Reproduction pattern · Evidence gathered · Ruled out · Not yet ruled out · User impact/tone. Expected output: a note an L2 network engineer could act on without reading the thread.
Correct the escalation note like the L2 who receives it. Check five things: is the error code exact and attributed correctly ("user reports 720-something" vs stating "error 720" as verified)? Is the new-router event in the timeline? Is "reinstall client" listed as ruled out — twice? Is the 10-second-connect-then-drop pattern captured? Is the deadline/frustration flagged so L2 calls her instead of emailing? Log each miss in your diff.
Generate resolution notes for Threads 2 and 3. Prompt: "Write a resolution note for the ticket record: Issue · Environment · Root cause · Resolution steps taken · Verification · Prevention/next step. Publication quality — another engineer should be able to resolve a repeat in five minutes." Expected output: two notes of 120–200 words.
Correct both resolution notes for cause-precision. Thread 2's root cause is not "network problem" and not merely "printer misconfigured" — it is a specific paper-size mismatch plus a hold-mismatched-jobs driver policy; the note must say why jobs vanished silently. Thread 3's root cause is not "user lost access" — it is a recert auto-expiry after manager non-response, and the prevention line belongs in the note. Check also that the AI did not invent steps (a favourite: "restarted the print spooler" — nobody did).
Build the diff summary. Create diffs.md: one row per artifact — artifact · # corrections · worst error the draft contained · category (invented fact / dropped fact / user-claim-as-fact / wrong emphasis / tone). Expected output: 5 rows (3 summaries, 1 escalation, 2 resolutions — count the escalation once).
Write the reusable prompt. Based on your corrections, improve the three prompts from steps 1, 3 and 5 so the next draft avoids your top two error categories (e.g., add "distinguish suggested vs done" or "quote error codes with attribution"). Save as prompts.md — this is your desk's template, and the lab's most reusable output.

Pro Tip

Read every AI draft in "attribution mode" — for each factual claim ask "which line of the thread says that?" A claim with no source line is either invented or upgraded from a user's guess to a fact; both go in the diff.

Verify

Six artifact files exist: t1/t2/t3-draft.md untouched, t1/t2/t3-final.md corrected, plus diffs.md and prompts.md.
Every final file contains at least one visible [was: ...] correction — an empty diff on any thread means re-audit it.
The Thread 1 escalation note includes the router replacement, the connect-then-drop pattern, and reinstall listed under ruled out.
Neither resolution note contains a step that nobody in the thread performed.
Thread 2's note explains why jobs disappeared silently; Thread 3's note names the recert mechanism and the prevention step.
diffs.md classifies every correction into one of the five error categories.

Self-grading rubric (pass = 7/10): dropped-fact catches (0–3, router detail alone is worth 1) · claim-vs-fact discipline (0–2) · invented-step catches (0–2) · root-cause precision in both resolution notes (0–2) · improved prompts address your actual top error categories (0–1).

▸Answer key — the corrections your diff should containoptional deep dive — click to expand

Thread 1 summary/escalation — expected catches:

The router. Megan mentions a new ISP router casually at 11:03 after the emotional peak; AI summaries reliably drop it or bury it. It is the single most probable root cause (new router defaulting to settings that break GRE/PPTP-era error 720-class connections or aggressive NAT/session timeout matching the 10-second drop). It must appear in Situation or Evidence, prominently.
"Reinstalled twice" is ruled out, not next-step. Drafts often list "reinstall VPN client" under next actions because the previous agent suggested it at 10:05 — the user had already done it twice. Listing it again is exactly what enraged her.
Error code attribution. She said "error 720 something". The note should say user-reported error ~720 (unverified) — stating "Error 720" as confirmed fact is a claim upgrade.
Pattern precision. "Connects ~10 seconds then drops" is diagnostic gold (tunnel establishes, then dies — points at NAT/keepalive, not credentials). Generic "VPN not working" in a draft is a wrong-emphasis correction.
Netflix ≠ ruled out. "Home internet is fine, Netflix works" rules out total connectivity failure only — a draft listing "home network ruled out" is wrong; the new router is the prime suspect.

Thread 2 resolution — expected catches: root cause is A4 batch jobs vs Letter-only tray, with driver policy holding mismatched jobs silently (hence "vanish, no error"). "Network problem" (Derek's/ERP-vendor's claim) must not appear as cause. Watch for invented steps — spooler restarts, driver reinstalls, test pages nobody printed. Prevention line: fixed the ERP print profile and released 14 held jobs; note the retest scheduled for tomorrow's batch. Derek's floor-wide "network is terrible" comment is a separate ticket, not scope creep into this record.

Thread 3 resolution — expected catches: root cause is quarterly access-recertification auto-expiry after manager non-response — not "permissions issue", not "IT changed something" (her claim), and definitely not an error. Resolution must include the approval step (policy-compliant re-add, not a favour), and prevention is the manager responding to recert emails. A draft that says "re-added user to group" with no approval mention documents a policy violation that did not happen.

Scoring: router catch = 2 · reinstall-as-ruled-out = 1 · attribution fixes (720, Netflix, user claims in threads 2/3) = 2 · silent-vanish mechanism in T2 = 2 · recert mechanism + approval step in T3 = 2 · diffs classified = 1.

Key Takeaway:

Drafts are free now; judgment is not. Ship the diff — what you corrected and why — and you have evidence of exactly the skill that survives the AI-drafted desk: turning a plausible draft into a true record.

© Anvi Innovate · BXUP — Be Exponentially Upskilled · BXUP is an independent learning platform, not affiliated with or endorsed by AWS, Microsoft, Google, NVIDIA, IBM, Salesforce, Databricks, Oracle or Anthropic. All trademarks belong to their respective owners. · bxup.in