Monthly Client Report: how to use (collection guide, type: inst, public)

This collection is not a set of independent tools: it is one production line for the AI Connect monthly client report, where each skill's output is the next one's input. The frame is AI Connect's and never changes; the content is the client's and changes entirely.

The map

• absence-check: consolidates the month's absences: confirmed periods, candidates, pending identifications, operational pause dates.
• monthly-report: the chassis: JSON in, A4 report out, on the fixed AI Connect template. Client specifics come from the client profile.
• report-nine-check: correctness gate: every number recomputed from source. Nothing ships without it.
• report-quality-gauge: quality gate: 0-100 score on communication, three improvements maximum.

The four-layer access model (read before running for a client)

1. Code (this collection, public): generic skills, zero client data.
2. Configuration (closed collection per client, e.g. "Client Ops: AZOS", unlisted): the client profile snippet + pointers. Ask the client owner for the link: that is the "opening" of the closed collection.
3. Data (never in SkillSafe): rosters, absence boards, exports live in the client folder (Drive ACL). Whoever runs the report exports their own agentjobs JSON with their own token: fresh data beats copied data.
4. Secrets (individual): access tokens and keys are personal. Skills say "ask the admin for a token", never carry one.

The monthly path

export JSON (your token) → absence-check → monthly-report
  → report-nine-check (loop until PASS) → report-quality-gauge
  → export PDF (A4, template dialog) → send with the month's e-mail

First month for a new client is dominated by the seven survey decisions (see monthly-report's references/template-multicliente.md); from the second month on, it is just the JSON.

When NOT to use

• Fortnightly touchpoint decks (16:9): separate skill, different document.
• Investigation or ad-hoc analyses: this line produces one artefact only.
• A month whose data you cannot export yourself: do not run over someone else's stale copy; ask for access instead.

Anti-patterns

• Skipping gates. A report that skipped nine-check is not "probably fine"; it is unverified. One wrong number spends the credibility of all the others.
• Editing the chassis for one client. Client needs go in the client profile or the data folder. The day the frame bends for one client, every client's report stops being recognisable.
• Publishing inference. Absence candidates, guessed names from handles, provisional classifications: all of it is question material, never report material, until confirmed.
• Comparing months computed under different rules. Recompute the previous month from its export first, or print the "no comparison" note.

Privacy rule (the one that overrides everything)

Skills carry rules, never records. If you find yourself typing a person's name into anything in this collection, stop: names belong in the client data folder, behind its ACL.