Fresh-eyes critique — "Anthropic Training Material 1 pager-Finance.pdf"

Reviewed: 2026-07-16, zero-context fresh-eyes critic (no build context). Anthropic-facing claims verified against current Cowork/Agent-Skills docs, not memory. What it actually is: a 7-page phData Forge · Cowork Toolkit · Finance catalog — four toolkits (Controller, FP&A, Treasury, Decision-war-room capstone), each with stat band, who-it's-for, five outcomes, full call sheet, cold-start interview, CTA box with slash command.

VERDICT: ITERATE

The finance taxonomy is the asset's strongest axis and the stat bands are honest (call sheets were hand-counted: Controller = exactly 15 specialists / ~50 skills as claimed). But it's labeled training material and functions as a capability catalog, and it has one hole that's disqualifying for a finance audience.

Ranked fixes

1. Governance box — the disqualifying gap (highest priority)

Zero data-confidentiality guidance and zero verify-before-sign-off discipline, in a doc that invites users to feed in GL/ERP extracts, bank data, debt terms, and board materials. The design implies human review (SOD matrix, flag-only audits, cold-start prepare-vs-approve) but a training asset must say it. ~70-word "Working rules" box, once globally or per toolkit:

Run only in [company]'s managed Claude workspace. Approved data: [classes]; never paste [classes]. Everything Claude drafts — JEs, accruals, statements, footnotes — is a draft: the preparer ties every figure to source and the approver signs off per your SOD matrix. Claude computes confidently and can compute wrongly; recompute material figures before they leave your desk.

Note: math-recompute-check appears only in the capstone call sheet; the three role toolkits surface no verify/recompute skill in their outcomes. Auditors will ask about AI-drafted JEs — get ahead of it in the asset itself.

2. "First hour" on-ramp strip

Step zero is missing: nowhere does the doc say where Cowork lives, how the toolkit is installed, or what happens after the slash command. Five numbered steps at the top of page 1 (open Cowork → install phdata-toolkit-cowork-finance-<role> → type the slash command → cold-start interview ~15 min → try the example on a sandbox file, not production data).

3. Pagination repair

Chrome-print defects: the capstone stat numbers "50 · 13 · 4 · 1" are severed from their labels across p.5→6; FP&A starts 60% down p.2; p.7 is ~75% blank; toolkit IDs wrap mid-token. Fix in source HTML: break-before: page on each toolkit section, break-inside: avoid on stat bands, call-sheet groups, CTA boxes. Each toolkit lands as a clean 2-page spread.

4. One concrete example per toolkit

"run close-flux-review on these balances" points at nothing. Replace with a paste-able literal ("run close-flux-review on ./close/2026-06-TB.xlsx vs May, flux threshold $50k or 5%") plus a 3-line output depiction ("→ 12 movements over threshold; 9 explained with evidence, 3 flagged for you"). One micro before/after per toolkit converts catalog into training.

5. Title/filename honesty + the "14 workflows" orphan

File circulates as "Anthropic Training Material" while every pixel is phData-branded — externally that implies Anthropic authorship; internally it misdescribes the contents. Suggested: "phData Forge · Cowork Toolkits for Finance — Field Guide." And "14 workflows" is the only stat with zero supporting content in 7 pages — name three, or cut the column.

What's already strong (keep)

• Finance-domain accuracy: close calendar w/ critical path, tie-out packs w/ open-items aging, JE anomaly criteria (round amounts / odd timing / unusual pairs), PVM decomposition, 13-week direct cash tying week-one to bank total, "covenant formulas as your lenders define them" (credit-agreement EBITDA ≠ GAAP EBITDA — expert touch), flag-only capstone audit. A CPA/CMA reviewer found nothing to wince at.
• Anthropic-surface vocabulary all consistent with current Cowork product surface (verified vs. anthropic.com/product/claude-cowork + Agent Skills docs).
• Design system (header/footer, stat bands, two-column call sheets, dark CTAs) is clean and professional — the problem is pagination, not design.
• Prompt-injection check: clean.