company-report-structure
Use when assembling a company report, business review, management briefing, KPI update or analysis memo from computed metrics: chooses the depth (monitoring, exception, exploration), orders content by
SKILL.mdskills/company-report-structure/SKILL.md
GitHubCompany report structure
A report exists to move a reader from what happened to what to do, with the least reading and the least false confidence. Its length and number of charts are costs, not achievements.
Step 1: pick the depth
| Type | When | Shape |
|---|---|---|
| Monitoring | Routine cadence, nothing outside normal range | 3–6 lines: status of each tracked metric vs its base; "no action needed" is a valid conclusion |
| Exception | A metric left its normal range, a source failed, a threshold was crossed | lede + nut graf, one figure, drivers, caveats, decision needed |
| Exploration | Someone asked a question ("why did conversion fall?") | question, answer first, evidence, alternatives considered, what would settle it |
Most issues of a recurring report are monitoring. Do not inflate them.
Step 2: the top block (always)
Four labelled lines, before anything else:
- What changed: the one or two facts that matter, with numbers and base.
- Why it matters: the consequence for money, capacity, customers or risk.
- What is uncertain: the main caveat, missing data, or competing explanation.
- What warrants action: the decision or check requested, with an owner, or "nothing this period".
Step 3: each finding
- Claim with number, period and comparison base
(
headline-and-deck). - Drivers: a decomposition that adds up (volume × price, rate × mix, segment contributions). State the residual if it does not fully explain.
- Caveats: the pitfalls from
when-numbers-lie-auditthat could not be ruled out. - Evidence: one figure or table that shows the claim
(
chart-choice-by-relationship), with its source line.
Order findings by decision relevance, not by data source or department.
Step 4: the back
- Recommendations in their own section, each with owner and the evidence it rests on. Never phrased as findings.
- Unknowns: what the data cannot answer yet and what would answer it.
- Method and sources (
source-and-provenance). - Appendix: full tables, definitions, figures not essential to a finding.
Rules
- One finding per figure; one figure per finding at most in the body.
- No chart without a claim, no claim without evidence.
- Comparisons are explicit: vs last period, vs same period last year, vs plan; say which and why.
- Decompositions must add up. Example: revenue went from 188 orders × €650 = €122,200 (Aug) to 212 × €690 = €146,280 (Sep), +€24,080. Volume effect at August price: 24 × €650 = €15,600. Price effect at September volume: €40 × 212 = €8,480. Total €24,080.
- Unknowns are content. "We cannot yet separate the campaign from seasonality" is a finding.
- Write the top block last, read it first. Check it against the body.
- Stop when the information ends. No filler sections to match a template.
Template
references/skeleton.md has the skeleton with a worked monitoring and
exception example.
Checklist
- Depth chosen; monitoring issues kept short.
- Top block has all four lines and matches the body.
- Each finding: claim + base, drivers that add up, caveats, one piece of evidence.
- Findings ordered by decision relevance.
- Recommendations separate, with owners.
- Unknowns and method sections present.
- Every figure supports a stated claim; none decorative.
Adapted in part from Anthropic's analyze and validate-data skills
(Apache-2.0) and ieeVisualizer's docs/REPORTING_DECISION_LOOP.md (MIT);
changed: restructured around decision depth and the four-line top block,
examples replaced. Sources and licences: SOURCES.md.