Monthly briefing skills: two source files, no seventh card

The shortlist and decision

The small search also surfaced Equipt ad-performance-report (paid ads only), Two Minute Reports Instagram Insights (live MCP/Instagram connection; manual data forbidden), and a Google Ads manager repository (Google Ads plus local scripts). Those were scope exclusions from accessible primary pages, not full package or security reviews. A prompt, dashboard, connector, or service was not counted as a portable-skill candidate.

Anthropic: manual data in the selected file, connected options around it

The selected performance-report/SKILL.md was observed as 392 lines on main. It asks for report type, period, source, optional comparison and audience. With a connected analytics tool, it instructs discovery and automatic pulls; otherwise it asks the user to paste or share metrics, including CSV, spreadsheet or key numbers. It offers a monthly template, channel summaries, a comparison table, trends, hypotheses, recommendations and some rate definitions. It notes attribution gaps and forecast uncertainty. These are written instructions, not observed output.

The selected file's pasted-data path could avoid a service only with separate constraints. The root README and marketing README describe installing a broader Claude Cowork or Claude Code plugin. The connector map and plugin MCP configuration describe external analytics endpoints, including Supermetrics and Amplitude; the surrounding README describes Slack and other integrations. None was installed, connected, permission-checked or tested. The selected file's post-report menu offers to draft a stakeholder email; it does not order an actual send or spend change.

The gap is specific: no mandatory export identity, timezone, extraction date, row IDs, units/currencies, attribution definition, comparable-window check, same-source rate constraint, zero-denominator halt, deduplication approval, claim-to-source appendix, discrepancy stop or human sign-off before forwarding. A formula table does not itself recalculate supplied numbers. Its “what worked” framing can overclaim if owner context is missing, although it labels some reasons as hypotheses. Thus the finding is incomplete guardrails, not a claim that every output is causal or wrong.

Provenance limit: the selected-directory listing showed only SKILL.md but was cached two weeks earlier, not a verified 29 September inventory. Path history showed March 2026 changes, and an 18 March change visibly added a Supermetrics connector option to this path. This is historical path evidence, not a current-file hash, pin or tag attribution. The root license is Apache-2.0.

Skill Me: a sent report is the stated endpoint

The original agency-monthly-report/SKILL.md was observed as 120 lines. It asks for the budget holder's metric and target, current/prior platform numbers, work shipped and earlier bad-news notices; it says modeled or estimated numbers should be labeled. The results-first template has limited figures, misses, corrections, a next-month plan and one client ask. But Step 5 tells the user to send on a fixed day through the contracted channel; the named deliverable is a sent monthly report. Merely reading the file sends nothing. Its instructions nonetheless cross the Run 8 draft-only boundary without an authorized review of source rights, recipients, version, arithmetic and claims. An example attributes CPA improvement to creative without an evidentiary test.

The root README describes site, npx skills add and manual-copy routes as project claims, not verified compatibility. The MIT license names Alexander Ouellet (2026). Root history showed July sync commits; commit a28c4ce includes the selected path. That is a historical change, not a pin of today's file. Selected-directory and path-history requests failed with cache miss, so file inventory and tag attribution remain unverified.

The file also refers to onboarding and client communications neighbors. They mention signed SOW, payment, contacts, asset/platform access and contracted delivery channels. Those are surrounding workflow context, not a connector or automatic account access in this selected report file. No script or MCP requirement was observed in the file; the unavailable directory prevents a whole-directory text-only claim.

Seven gates the titles do not settle

“Absent” means absent in the inspected selected text and immediately relevant surface, not impossible for a human to add. No first-use walkthrough was run. Source inspection cannot establish calculation accuracy, privacy, compatibility, safety, speed or marketing outcomes.