analytics-office

A discipline harness for AI-assisted analytics.

AI is genuinely good at analytics work. It fails in one specific way: not by being unable, but by being agreeable. Handed an inherited warehouse with no docs, it starts building instead of orienting. Handed "build me a dashboard with these KPIs," it builds exactly that — without ever asking what decision the dashboard serves. Asked what "active customer" means, it picks a sensible default instead of pinning the choice with the person who owns it. Handed a number, it writes the confident story. The fix is a model that refuses and redirects — that refusal-then-redirect is the whole product.

analytics-office is a bench of 19 read-only skills for Claude Code that switches the model out of answer-mode and into the discipline each moment of analytics work actually needs — orienting on an unfamiliar estate, interrogating requirements, proposing the questions worth asking when the stakeholder has none, locking metric definitions, designing models, auditing inherited premises, reviewing the code behind a number, diagnosing a break, exploring a dataset without dredging it, mapping the estate, scoping a change's blast radius, QA-ing the dashboard before the QBR, proving a migration ties out, briefing stakeholders, writing the weekly status, surviving the meeting. Each skill is engineered against a documented failure of the bare model. All of them write to one living knowledge base you resume from.

Read-only by construction · computes instead of eyeballing · no required sequence — any skill, any moment.

/plugin marketplace add debabsah/analytics-office
/plugin install analytics-office@analytics-office

---

The bench

Nineteen skills in five families — each family owns an ask-shape, and every member's description opens with its family's shared stanza (that structure is measured, not aesthetic; see Engineered, not vibed). The deep dive — every skill's job, trap, loop, and artifact, with diagrams — is docs/skills-deep-dive.md. Ordered here like a project — but there is no pipeline. Every skill fires independently, at any moment, with or without the others having run.

Shape — the work itself is still being shaped, before anything is built

You say	What fires	You walk away with
"I inherited this warehouse and the analyst left."	groundwork	a living knowledge-base/ — and a map of what you don't know yet
"Build me a dashboard with these KPIs."	requirements-interrogator	the decision behind the ask, the requested-vs-derived delta, a verdict
"We don't know what we need — what can our data tell us?"	worth-knowing	the question charter: candidate questions ranked by decision-weight, the unasked included, hypotheses never findings
"Lock down what 'active customer' actually means."	kpi-contract	a versioned contract — every definitional fork pinned by its owner or flagged [needs decision]
"How should I model this mart?"	model-contract	a logical star with the grain declared and gated on evidence — no DDL invented on a guess

Audit — a built thing is about to be trusted; the gate fires before the work leans on it

You say	What fires	You walk away with
"Turn this proc's output into the board number."	audit-my-assumptions	a graded register of every silent premise, falsified before you build on it
"Is this SQL right?"	review-my-query	findings graded Blocking / Latent / Advisory against the locked definition — a review, never a rewrite
"Is our knowledge base still true?"	kb-reconcile	a graded drift report — contradictions, stale claims, unsourced numbers
"QA my dashboard before the QBR."	review-my-dashboard	the assembly review: dashboards fail between correct parts — totals, defaults, titles, staleness

Validate — a measured result is about to drive a decision; the checks are computed first

You say	What fires	You walk away with
"Did our A/B test really win?"	audit-my-experiment	computed validity checks (SRM, peeking, multiplicity, power) gating the ship decision
"Can we plan against this forecast?"	audit-my-forecast	leakage, backtest, interval-honesty, and drift checks gating the plan
"The totals match — sign off the migration."	prove-my-parity	the stratified parity proof: offsetting errors caught, tolerance owned before results

Investigate — hands-in-the-data right now

You say	What fires	You walk away with
"Churn jumped to 11% overnight. Why?"	triage-my-number	a ranked differential across code / data / pipeline / definition / real change — plus a calibrated line for the exec who's asking
"Explore this data — find me insights."	explore-my-data	a harnessed exploration: every cut counted, found ≠ confirmed, the lucky cell never becomes the headline
"Draw the ER / lineage diagram of our mart."	map-my-estate	a cited map: every edge carries its evidence, guesses render dashed, islands stay islands
"What breaks if I rename this column?"	change-impact	the graded blast radius: breaks, silent meaning-drifts, and honest UNKNOWNs — before it ships

Deliver — work is leaving the desk

You say	What fires	You walk away with
"Write up my findings for the VP."	brief-my-findings	a brief where every claim carries its provenance and open questions stay open
"The CFO will grill me on this number."	defend-my-number	a live sparring drill, graded honestly, and a defense sheet of what held and what cracked
"Write my weekly status update for steering."	status-truth	a provenance-graded status where every green earns its color and slips carry their delta

The rooms go deeper: eight modes

A mode adds a job to a moment a skill already owns — a branch inside the host, zero new routing surface (that's the growth rule: modes before skills). The daily ones:

• Micro-brief (brief-my-findings) — three sentences for the exec with the claim discipline compressed, never dropped; an expired verdict is never quoted as standing.
• Delta brief (brief-my-findings) — "what changed since the last readout," composed as a diff of the record.
• Meeting armament (defend-my-number) — the 30-minute card: state, holding lines, likely attacks, and the explicit do-not-say list.
• Morning brief (groundwork) — open loops on one screen: pending paste-backs, expired verdicts, aging questions, the next move.
• Decision archaeology (groundwork) — "why did we decide X?" answered only from the record, with citations; "the record is silent" is a legitimate answer, confabulated history is the named enemy.
• Handoff package (groundwork) — the curated KB tour that normally walks out the door with a departing analyst.
• Meeting capture (groundwork) — raw notes to candidate record entries, every attribution confirmed-or-[unconfirmed], owner-pinned before write.
• Test-design-from-contract (kpi-contract / model-contract) and the change-request gate (requirements-interrogator) — a locked contract projected into acceptance-grade test specs; a mid-flight "can you also add ___" met with a delta ledger and an owner-pinned accept/defer/reject.

---

What "discipline harness" means

LLMs in analytics fail through answer-mode: the pull to be immediately useful. Answer-mode inherits a stale filter as fact because "that's what the proc does." It eyeballs a check it could compute. It resolves a contested definition with a "sensible default, confirm later." It smooths an open question into a clean narrative because the deck reads better that way. None of these are knowledge failures — they're discipline failures, and they produce confident, well-formatted, plausible, wrong output.

A harness is the countermeasure, built into every skill:

• A trap, named. Each skill documents the exact thing a capable model does by default — then refuses it. The skills know their own failure modes before you hit them.

• Bright lines. Non-negotiables with teeth: never touch a live system, never compute the user's deliverable, never resolve an owner's decision silently, never grade a guess as a finding.

• Anti-evasion tables. The mid-task rationalizations, pre-rebutted. Two real rows:

  The thought	The reality
  "The QBR's in an hour, I'll just write the SQL so they're unblocked."	Surface, don't build. The contract is the deliverable; the runnable query is downstream of the pinned definition.
  "It's obviously X."	Obvious = untested. Hold the differential until a check confirms.

• A graded artifact, every time. No skill ends in vibes. Each emits a committable file where every line carries a status. The signature example — kpi-contract's fork log:

  Fork            Options               Pinned             Why it matters
  Revenue basis   bookings/recognized   recognized         biggest gap vs Finance
  Refunds         gross/net             net                gross overstates by refund rate
  Attribution     first/last/multi      [needs decision]   changes who gets credit
  

• Verdicts that carry. A "not ship-ready" from an audit cannot be upgraded into a win by the write-up downstream. The brief inherits the verdict; it does not soften it.

• Engineering constraints, enforced. Every skill body is capped at 200 lines by a structural validator (depth lives in references/, loaded on demand), and every skill declares least-privilege tool access — the validator rejects a wildcard grant. The whole bench is about 1,300 lines of always-loaded skill text — roughly 4,000 counting the on-demand references and kits.

---

One living knowledge base

Every skill reads from and writes to the same knowledge-base/ directory in your project — current truth in STATE files, history in an append-only timeline, and one graded artifact per job done.

flowchart TD
    KB[("knowledge-base/<br/>STATE + timeline + decisions<br/>catches.md + house-rules.md + inputs/")]
    GW[groundwork] -->|stands up + seeds| KB
    KC[kpi-contract] -->|locked fork log| KB
    AA[audit-my-assumptions] -->|assumption register| KB
    ME[map-my-estate] -->|the cited map| KB
    KB -->|the contract anchors the review| RQ[review-my-query]
    RQ -->|graded findings| KB
    KB -->|known defects become first suspects| TN[triage-my-number]
    KB -->|the map becomes the blast radius| CI[change-impact]
    KB -->|verdicts carried with their age — expired ones can't be quoted| DEL[brief-my-findings<br/>status-truth<br/>defend-my-number]