cx-case-timeline
Use to reconstruct what actually happened to one customer across every conversation, channel and handoff, for an escalation, complaint, post-mortem or goodwill decision. Trigger for "build a timeline
SKILL.mdskills/compliance/cx-case-timeline/SKILL.md
GitHubCase timelines
Someone needs to decide something — a goodwill credit, a complaint response, whether an agent misinformed a customer — and the facts are spread across several tickets, two channels, a phone call, and a third-party thread. The deliverable is a timeline someone can act on and, if it comes to it, defend.
A case is not a ticket. Almost everything that goes wrong in this task comes from treating a ticket id as the boundary of the case.
Step 1: anchor on the customer, not the ticket
Start from the ticket you were given, then widen. The case includes:
- Earlier conversations from the same customer about the same issue, often with no link between them. The first contact is usually the one that matters most and is the one nobody finds.
- Channel switches. A customer who chatted, then called, then emailed has three records and one problem.
- Reopens and re-raises. A closed ticket followed by a new one two days later is one case with a gap in the middle, and that gap is frequently the finding.
- Merged and duplicate records. Follow duplicate pointers in both directions. A merged ticket can look abandoned while the work continued under another id.
- Internal-only threads — escalations to a back-office queue, a tracker issue, a third-party investigation. The customer-facing record often goes quiet precisely while these are running, and a timeline without them shows unexplained silence.
Search on every identifier the customer has: account id, email, phone, company, and the name as agents typed it. Report which identifiers you searched and which you could not.
Step 2: order by event time, not by record time
The single most common defect in a reconstructed timeline is wrong ordering, and it comes from sorting on the wrong timestamp.
- Event time is when the message was actually sent in the source system.
- Record time is when it landed in whatever you are querying.
Backfills, sync lag and re-imports make these diverge by hours or months. Sorting a timeline by record time produces a confident, plausible, wrong story — and it is wrong in the most damaging way, because it can reverse cause and effect.
Sort on event time, fall back to record time only where event time is absent, and mark every entry where you fell back. Where a timeline is used to establish who knew what when, an entry with an uncertain timestamp is an entry that cannot support that conclusion.
Normalise timezones and say which one the timeline is rendered in. Business-hours context matters too: a nineteen-hour gap over a weekend is not the same neglect as a nineteen-hour gap on a Tuesday.
Step 3: separate three different things
Keep these in separate columns. Collapsing them is how a timeline becomes an argument rather than a record.
- What the customer was told. Customer-visible messages only.
- What actually happened internally. Internal notes, status changes, escalations, third-party responses.
- What the record asserts. Evaluation summaries, dispositions, tags, and the case's own status history.
The divergence between these is usually the finding. A case where the customer was told "we're processing it" three times while the internal record shows no action between them is a specific, defensible finding that no single column reveals.
Step 4: attribute every turn correctly
- A bot reply is not "we responded". Automated acknowledgements, AI agent replies, and out-of-office autoresponders all look like agent messages in most data models. Label them, and never count them as a human response in a responsiveness finding.
- Name the human who acted, per turn, and note reassignments. A case handled by five people with no handoff notes is a routing finding.
- Check for the customer's follow-ups. A second unanswered message is much stronger evidence than a long first wait.
Do not trust a single author-type field in either direction; validate against the message content on a sample before building attribution on it.
Step 5: mark the gaps explicitly
A timeline with an unmarked gap is worse than no timeline, because the reader assumes completeness and decides on it.
State plainly:
- Conversations you know exist but could not retrieve, by id.
- Channels not covered — a phone call with no recording, a chat that was not synced, an in-person branch visit.
- Periods with no records at all, distinguishing "nothing happened" from "we have no data".
- Content unavailable: transcripts that failed to load, redacted messages, deleted records.
Say which of your conclusions would change if a gap turned out to contain something.
Template
Case timeline — <customer/account> — prepared <date>
Rendered in <timezone>. Ordered by event time; ~ marks an estimated timestamp.
Conversations in scope: #<id> (email), #<id> (chat), #<id> (voice), #<id> (internal)
Identifiers searched: <account id, email, phone, company>
Not retrievable: #<id> (<reason>)
<date time> <channel> <actor: customer | agent name | BOT | third party>
Customer-visible: <what was said, minimally quoted>
Internal: <note / status change / escalation>
...
Key findings
- <what the customer was told vs what happened>
- <unexplained gap of N business hours between X and Y>
- <first point at which the issue was known internally>
Gaps and their effect
- <gap> -> <which finding would change>
Guardrails
- Cite ids; quote minimally. Support transcripts carry names, addresses, card fragments, and financial and health disclosures. Quote only the specific sentence a finding rests on, redact the customer's identifying details, and say you redacted.
- Never reconstruct content you could not read. If a transcript failed to load, the entry is "content unavailable", not a paraphrase from an evaluation summary. This is the one failure that destroys the credibility of the whole document.
- Do not attribute intent. Record what was said and done. "The agent did not check the account" is a finding; "the agent did not care" is not.
- A timeline is not a decision. If the request is a goodwill credit or a complaint outcome, present the facts and the options; the decision belongs to the person with the authority for it.
- Expect this document to be read by someone outside the team. Complaints and escalations get shared with regulators, legal, and sometimes the customer. Write it so that is fine.
Present results to the user
- Scope — conversations included, identifiers searched, timezone, ordering basis.
- The timeline, in the template above.
- Findings, each tied to specific entries.
- Gaps, and which findings they could change.
- What this does not establish, said plainly. Ordering uncertainty, missing channels, unretrievable content.