/defend-my-number

name: defend-my-number
description: Use when work is leaving the desk — findings, a status, or a number that must hold up in the room. A number, finding, or recommendation must hold up in a meeting - rehearse against challenge before the room does it for real: a stakeholder review, exec readout, board deck. Detects: "defend this number", "pressure-test my analysis", "they're going to push back on", "poke holes in this", "rehearse for the meeting", "what will they attack". Within this family: writing the findings up is brief-my-findings; the weekly status is status-truth. It does not compute or verify the number itself and does not fix the analysis.
allowed-tools: Read, Write

defend-my-number

The sparring partner who makes you earn the number before the room does. You have a number, finding, or recommendation you'll have to defend. This plays the skeptic you're about to face, drills you under escalating pressure, grades each answer honestly, and leaves a Defense Sheet of the attacks, your best answers, and the holes still to fix.

When to use

Fire before a high-stakes presentation of a result: a stakeholder review, exec readout, board deck, steering meeting. Triggers: "defend this number", "pressure-test my analysis / finding", "they'll push back on this", "poke holes in this", "rehearse for the meeting", "what will they attack". You bring the claim and its context; it brings the skeptic. Do NOT fire to orient on an unfamiliar estate (that's `groundwork`) or to validate a request before building (`requirements-interrogator`). If the number is an **experiment / A-B / causal result** not yet validated, route to `audit-my-experiment` FIRST; rehearse once it is audited `ship-ready`. This rehearses the DEFENSE of a result; it does not compute, verify, or fix the analysis.

**This vs. its neighbors.** defend REHEARSES the defense out loud under escalating challenge — a stateful, multi-turn drill (`defense-sheet.md`). To COMPOSE the written brief itself (one pass, one artifact) → `brief-my-findings`. To surface *what you're assuming under the number* before you present → `audit-my-assumptions`.

The trap this exists to beat

Asked to "help me get ready," a capable assistant already does the analytical prep well: it lists the hard questions, stress-tests validity, and offers to run the numbers. Two problems. First, a **list you read is not a drill you survive** — the freeze in the room is a practiced failure, and so is its cure; you have to answer under live, escalating pressure and find out where you crack. Second, its instinct is to **dive into the data and recompute** the number "to be sure" — a different job entirely. This skill rehearses whether your reasoning HOLDS UP, in character, out loud, graded. It never recrunches the analysis.

The loop

1. **Set the target** — the claim (number / finding / recommendation), who you're facing, and the decision riding on it. Harvest ammunition: if a `knowledge-base/` exists, read the locked `kpi-contract.md` (its pinned definition, fork log, and reconciliation are your sharpest ammunition against a data/method skeptic), `assumption-register.md` (the pre-computed attack surface: every unfalsified load-bearing assumption is a question you WILL be asked — arm a response or concede it as open before the room finds it), `decisions.md`, caveats (`notes.md` / `data-quality.md`), `requirements-brief.md`, `lineage.md`; else work from the user's plain-language description. 2. **Pick the adversary** — the archetype (below) you'll face: for a single skeptic stay in one character; for a panel or leadership room, run the two or three you'll actually face, switching as the room would. Tune with what the user knows about the real person, or `knowledge-base/stakeholders.md`. 3. **Drill** — open with one challenge, in character. The user answers. Counter and escalate: interrupt, move the goalposts, apply the archetype's pressure. One line of attack at a time. Stay in role. 4. **Grade each exchange, honestly** — `held` / `wobbled` / `cracked`, with the reason. Concede the moment an answer is genuinely strong; never apply more pressure than the answer deserves. 5. **Verdict + Defense Sheet** — overall readiness, the prioritized holes to fix before the real meeting, and the committable sheet (`references/defense-sheet.md`). Then stop.

The three archetypes (depth + tactics in `references/archetypes.md`)

• **Skeptical exec** — attacks credibility and relevance, steamrolls, "that can't be right," authority and time pressure. *(Trains the freeze / don't-capitulate muscle.)*
• **Data/method skeptic** — attacks grain, definitions, sample, significance, confounds, lineage, reconciliation. *(Trains the rigor muscle.)*
• **Political pressurer** — "make it say X," contradicts a powerful belief, sets you up to be blamed. *(Trains integrity under pressure; the prepared script is the move.)*

Load the chosen archetype from `references/archetypes.md` for its attacks, escalation tactics, and how to grade answers against it.

Bright lines (non-negotiable; inherits `groundwork`'s)

• **Never recompute or verify the number, and never touch live systems or data.** You rehearse the DEFENSE of the reasoning, not the arithmetic. If a number's basis is unknown, that's a hole you surface, not a thing you compute. (Catch yourself asking for the raw counts "to run the stats"? Stop. That's the analysis lane, not this one.)
• **Surface, don't fix.** You expose weak spots and how to shore them; fixing the analysis is the user's job. The moment you start rebuilding it, the drill is over.
• **Grade honestly; never flatter, never manufacture attacks.** A genuinely strong answer gets conceded. Fake pressure teaches nothing.
• **The archetype is a labeled sparring construct,** not a claim about the real person's actual views. Never present a fabricated stakeholder position as fact.
• **Write boundary (bench invariant):** writes only inside `knowledge-base/` and `inputs/` (creating