cx-narrative-reporting
Use to write CX commentary that makes a number decision-ready — what moved, why (decomposed), so what, do what — separating noise from signal and stating confidence honestly. Trigger for "write the co
SKILL.mdskills/cx-operations/cx-narrative-reporting/SKILL.md
GitHubNarrative reporting
The chart is done. The number moved. The commentary says "QA fell 4pts driven by Team B's lower scores."
That sentence is wrong more often than it is right — and it is still the template most reports use, because naming a team is easier than proving the move is real, splitting mix from performance, and saying what anyone should do about it.
This skill writes commentary that survives the reply-all thread: structured, honest about uncertainty, and tied to a decision or a deliberate non-decision.
The four-part structure
Every metric narrative answers four questions in order. Skip or reorder them and readers fill the gaps with assumptions.
| Part | Question | Length |
|---|---|---|
| What moved | Direction, magnitude, period, comparison, denominator | One sentence |
| Why | Decomposed cause — rate, mix, coverage, composition, definition | One short paragraph |
| So what | Customer, cost, or risk implication if the move is real | One or two sentences |
| Do what | Named action, owner, or explicit "no action yet" with trigger | One sentence |
Example skeleton (fill with verified content only):
What: QA mean score fell 3.1 points (78.2 → 75.1) week-on-week on n=412 completed evaluations, ISO week 30 vs 29, same scope and definition v4.
Why: The move exceeds sampling noise (95% interval ±1.8pts). Decomposition shows ~70% mix (higher share of chat, structurally lower scores) and ~30% rate (chat itself scored lower). Coverage was stable; no definition change.
So what: If chat share stays elevated, headline QA will run below target even if chat quality stabilises — the aggregate is a mix story as much as a quality story.
Do what: Team lead for chat to confirm whether the chat rate drop is real in isolation by Friday; no headcount change until then.
If what moved fails the noise check, stop after stating that. Do not write why, so what, or do what for noise — the correct do what is "no action; recheck next period."
Step 1: noise before narrative
Before any "why", establish whether the movement exceeds what random sampling would produce. If you cannot compute an interval, say "direction only; n too small for confidence" — do not borrow confidence from tone.
Rules:
- Inside noise — report the interval, state "no action recommended", do not decompose. Decomposing noise invents drivers.
- Borderline — say borderline; prefer no action unless the cost of waiting is high and the proposed action is reversible.
- Outside noise — proceed to decomposition.
Weekly reports that flag "significant" moves at 95% without multiplicity correction will cry wolf — either widen thresholds for segment flags or label them "for review, not action."
Step 2: decompose before blaming
Aggregate rates move for four channels: rate change, mix change, coverage change, and composition change (entrants/exits). Check coverage before attributing a rate story — especially for QA and CSAT.
In prose:
- Lead with the headline split (rate vs mix vs entrants) in the metric's own units, not percentages of percentages.
- Rank segments by contribution to the move, not by worst rate.
- Name definition or instrument change before team performance — rubric changes look like collapses.
- State what you ruled out in one clause when it matters.
If decomposition is not available, say "cause not decomposed" rather than guessing. "We could not split mix from rate with current data" is acceptable; a invented team name is not.
Step 3: so what without drama
So what connects the number to something the reader cares about — customer experience, cost, risk, forecast credibility — without catastrophising or cheerleading.
- Tie to threshold or target only if the target is agreed and the definition matches.
- Distinguish structural (mix, definition) from operational (rate, process).
- If implications depend on an unverified mechanism, say "if X holds, then Y" — not Y as fact.
Avoid: "customers are suffering" without evidence; "great progress" on a move that is mostly mix or noise.
Step 4: do what with accountability
Every narrative ends with one of:
- Named action and owner — specific, proportionate to evidence.
- Deliberate wait — what you are waiting to see, and when you will reconsider.
- No action — inside noise or immaterial to targets.
"Monitor closely" without a trigger is not an action. "Share with leadership" is not an action unless someone is decision-ready.
Actions must match confidence:
| Evidence | Appropriate action |
|---|---|
| Noise | None |
| Real but unexplained | Investigate by date X; no irreversible moves |
| Real, decomposed, mechanism plausible | Scoped operational change |
| Definition break | Fix registry and relabel series; do not coach agents |
Voice and honesty
- British spelling where natural; plain sentences over jargon.
- Denominator always adjacent to rates in the narrative body.
- No invented confidence — "likely", "suggests", and "we ruled out X" beat false precision.
- Separate changed from notable — most periods have nothing worth a paragraph; saying "no material movement" is valuable.
- Data notes in prose — late arrivals, suppressed cells, restatements, definition version — one sentence at the end, not omitted.
Traps
- Explaining noise — the most common way narratives lose trust.
- Team name as cause — correlation from lowest rate, not contribution.
- Relative vs absolute change — "5% drop" when you mean five points on a rate.
- CSAT or QA without response/coverage context — when composition shifted.
- Confidence by adjective — "significant" without the interval shown.
Present results to the user
- Noise verdict — interval or explicit "n too small", and whether to proceed.
- Four-part narrative — what moved, why, so what, do what.
- Decomposition summary — rate vs mix vs composition, top contributing segments with denominators.
- Confidence statement — what is established, what is hypothesised, what was ruled out.
- Alternative readings — at least one plausible counter-explanation if evidence is partial.
- Data notes — late data, suppression, definition version, comparability limits.