Imported from Jakeintech/waybill (skills/report/SKILL.md). Install upstream with npx skills add Jakeintech/waybill --skill report. Copyright stays with the author.

Report

Turn ledger facts into a report a decision-maker can verify in two minutes. Brevity is part of the credibility: one page hard cap, every section has a line budget below, and when in doubt cut the sentence, not the number. The detail level scales the budgets (config.detail_default, or --detail / the user asking for a terse or full version): terse = the headline numbers in 1–3 lines per section (never dropping unattributed %, confidence, low_confidence labels, evidence tiers, or ranges — shorter never means less honest); standard = the budgets below; full = no collapse, every receipt and caveat rendered, still one claim per line. Read skills/ledger/references/methodology.md first and enforce its language rules throughout; read references/schema.md for field semantics.

The numbers come from the engine, never from estimation:

"${CLAUDE_PLUGIN_ROOT}/bin/waybill" mine --all                 # catch-up metering first, always
"${CLAUDE_PLUGIN_ROOT}/bin/waybill" query report --from <iso> --to <iso> [--audience <a>]
"${CLAUDE_PLUGIN_ROOT}/bin/waybill" query spend  --from <iso> --to <iso>    # spend-ledger detail
"${CLAUDE_PLUGIN_ROOT}/bin/waybill" query story <KEY>                       # "what did KEY cost"

You write the prose; the engine writes the numbers. Never recompute or adjust a number the query returned.

Resolve the request

• Window: default to the current sprint for "recap", the last full quarter for "quarterly", the review period the user names for a perf review, and the period since the last allocation (from config.json.allocations) for a token pitch. Confirm the window in the output header, not with a question, unless genuinely ambiguous.
• Audience preset (D12 redaction levels):
  • token-pitch, perf-review → --audience internal (keys and URLs stay; machine-local detail dropped).
  • anything the user will paste outside the org → --audience external (keys, titles, repos pseudonymized; numbers survive). Say in the footer that identifiers are pseudonymized and the internal version is available.
  • exploring their own data → self (the default).
• Presets: token-pitch (default when the goal is more tokens), perf-review, sprint-recap, quarterly, grant-report (maintainers: what the sponsorship shipped), incident (a windowed receipts pack over an incident timeframe). (Billing paperwork belongs to the invoice skill; per-item AI-involvement statements to disclose.)

Cold start: the bootstrap report

If the ledger has fewer than 3 shipped entries for the window — or the user asks for a "bootstrap report" or their "first report" — do not pad and do not refuse. Run "${CLAUDE_PLUGIN_ROOT}/bin/waybill" bootstrap for the zero-auth receipt; if MCP servers are connected, run the sync skill's flow for ~90 days first (confirm the window in one line), then produce a facts-only report. State plainly that no time-saved claims exist yet, and close with one line on how pre-registering upcoming tasks unlocks them. Never fabricate tier-3 claims to fill the gap.

Render (markdown, one page, this order)

1. Headline — one sentence: outcome + strongest range, from query report totals. Example: "Since the July grant: 34 points shipped across 2 epics, 11 PRs merged and deployed, an estimated 31–58 hours saved (pre-registered basis), at 0.9M tokens per point."
2. Shipped, with receipts — at most 5 lines, each KEY title (pts) — [PR](url) merged <date>, deployed <tag>, grouped under epic names, from report.shipped. Mark escrow-sealed items with "(pre-registered)". Everything else collapses to "+N more (ledger on request)".
3. Efficiency — tokens/point and tokens/PR from report.efficiency, with the previous window for trend if data exists (run a second query). When report.model_mix.by_model has 2+ rows, one line on tokens-per-point by dominant model ("opus-carried stories: 0.9M/pt; sonnet-carried: 0.4M/pt") — own history only, and rows built on a single story are anecdotes, said as such. When report.calibration.with_actuals > 0, one line on estimate calibration: coverage plus positions — and any above items (work that exceeded its no-Claude range) named, never dropped; that honesty is what makes the below items credible.
4. Spend ledger — from report.spend_ledger: total metered tokens, top accounts with confidence noted, open spend (tokens on not-yet-shipped stories), and the unattributed % — always shown, never hidden ("11% unattributed — shown, never hidden"). If the attribution inbox has open items, add one line offering to resolve them.
5. Costs & caveats — tokens spent vs. granted (utilization) from report.costs; USD using bundled Anthropic list rates by default (labeled "list-price equivalent"); rework from costs.reopened_count ("1 item reopened after shipping") and waste from costs.waste ("14 retried commands, 79 repeated reads across the window") — the unflattering numbers are what make the flattering ones believable; one line on what was excluded (judgment-tier claims, in-flight work).
6. The ask (token-pitch only) — from the forecast skill's output if upcoming work is known; otherwise state that the forecast can be generated next.
7. Methodology footnote — one line naming the evidence tier of each number class, plus "metering: deterministic, conservation-checked (waybill verify)".

Preset adjustments: perf-review leads with epic-level outcomes across all work_types (including review, incident, and docs work), frames the headline for the user's manager, and omits token-efficiency metrics unless asked; sprint-recap drops the ask section entirely. grant-report ("what did the sponsorship fund") windows on the grant period, leads with shipped items and closed issues (GitHub-issues keys render naturally), includes the AI-cost line as spend transparency, drops the ask unless the user is renewing, and defaults to --audience internal since maintainer work is public anyway — offer export --pack as the funder's verification. incident windows tightly on the incident timeframe the user names, filters shipped rows to work_type: "incident" (plus anything the user tags in; in-flight incident work under a labeled "In flight" line only if asked), leads with the timeline facts (sessions, turns, tokens over the window — when work started, how long it ran), and skips efficiency metrics entirely — nobody prices a fire by the token.

Hard rules

• Every metric traceable to event ids; keep the mapping, produce on request. The engine's numbers are the numbers.
• Ranges stay ranges. No midpoints, no "approximately 45 hours".
• Tokens are the native unit (D6 — bundled Anthropic rates ship as defaults via pricing import; override with pricing set); dollars appear only when a pricing basis is configured, always labeled with the pricing_version — and when spend_ledger.pricing_coverage.priced_pct is below 100, with what the dollars cover ("$412 covers 92% of tokens; the rest unpriced"). A silently partial total is a broken receipt.
• No peer comparisons, ever. Own-baseline and published team/industry benchmarks only.
• Delivered value includes only merged/deployed work. In-flight items appear only under a clearly labeled "In flight" line, and only if asked.
• No adjective without an artifact link.
• If the ledger is too thin for a section, say so in plain empty-state language: "Nothing shipped in this window. The ledger doesn't pad."

Output handling

Print the report in the conversation. Offer, without doing it unprompted, to save it to a file — a saved report is the export pack: Markdown, written by this skill. If the user wants a formatted document (HTML, DOCX, PDF), hand off to the host's document tooling (e.g. a docx/pdf skill when one is available) — document rendering is deliberately not an engine feature (see docs/skills.md, "the export-pack boundary"): the engine supplies the numbers, never the presentation. Confluence publishing is strictly opt-in: only when the user explicitly asks, publish via the Atlassian MCP tools, confirm the destination space/page first, and remind them in one line which audience level the report was rendered at before it leaves their machine.

Arm the recipient: the verification pack

When the report's audience is a decision-maker (token pitch, review packet), offer — in one line, without doing it unprompted — to attach the verification pack:

"${CLAUDE_PLUGIN_ROOT}/bin/waybill" export --pack --from <iso> --to <iso> --out <dir>

It writes a directory holding the verbatim event lines behind the report's numbers, the engine itself, and a README whose one command — node waybill.mjs verify --home . — lets the recipient re-run id determinism, escrow seals, and conservation offline. Two things to say when offering it: the engine refuses to build a pack from a ledger that doesn't verify green, and pack contents are unredacted by design (redaction would break the id checks), so it travels only where the internal report would.