Materials are sufficient for a concise daily brief: one FP&A tutorial separating the assumption layer from the formula layer, and one replicable reconciliation classification and escalation threshold framework. No padding.
This issue retains only materials that can be broken down into inputs, review steps, and output artifacts. Other sections are not padded.
Today’s Most Actionable Items (2)
1. Forecasting models: modify assumptions only, never formulas; use state files for incremental updates, avoid full recalculation each time
- Scenario: Rolling forecasts, operating assumption updates, monthly business reviews. The focus is “what changed on the business side → which model assumptions need updating”, not rebuilding an entire model table.
- Actions: The public tutorial decomposes the workflow into three Agents sharing the same directory structure: instruction file + read-only inputs + output directory + state file (recording the last processed point). This week, implement only the second Agent. Inputs are limited to: the existing model + one actual planning file (headcount plan / supplier pricing / sales forecast / expense proposal). The model modifies only assumption cells and lets the original formulas recalculate; modifying formulas or template structure is prohibited. Outputs must clearly document: which assumptions changed, source file reference, and impact on revenue / gross margin / headcount. The first Agent (daily scan of sales pipeline, procurement, headcount, and project tables, reporting only changes versus prior period) and the third Agent (monthly actuals → variance-driven) should not run in parallel this week. This is a tutorial demonstration, not a live financial model from any specific finance team.
- Review controls: FP&A owner pre-validates: no changes to formula areas, assumptions traceable to source files, no broken reconciliations. If mismatches exist or documentation is unclear, this version does not proceed to meetings. The model must not post entries, modify budgets, or issue final operating review materials.
- Outputs: Assumption change log (cell / old value / new value / file reference); recalculated P&L summary; risk alerts; FP&A revision log.
- Source: YouTube: Three Claude Finance Agents (Luke Finance) (tutorial / subtitled; source page shows approximately 3 weeks ago)
2. Reconciliation: classify into three layers first, then escalate by amount and aging; unexplained items must not close the period
- Scenario: Month-end bank reconciliations, GL control accounts versus sub-ledgers, intercompany matching. The focus is classification and escalation, not automatic posting of adjusting entries.
- Actions: The official reconciliation Skill divides differences into three layers. Layer 1 covers timing differences (outstanding items, in-transit, pending postings) — record expected clearance date, do not adjust. Layer 2 covers items requiring adjustment (unrecorded fees/interest, amount errors, duplicates, misclassified accounts) — produce draft entries only. Layer 3 covers items requiring investigation (unexplained causes, disputes, long-outstanding, recurring every period). Do not connect to production ledgers this week. Use only one cash account from the prior closed period: bank statement + GL cash balance. Reconcile both sides to the same cutoff date, produce the three-layer table, then assign owners by aging buckets (0–30 / 31–60 / 61–90 / 90+ days). Apply your own materiality threshold first; the example amounts in the source (single item >10k reviewed by supervisor, >50k by accounting) are for reference only — do not copy the thresholds.
- Review controls: Preparation and reconciliation must be performed by different individuals. Any unexplained difference, regardless of amount, must not close the account. Items aged 90+ days or increasing for three consecutive periods must be escalated; rolling to the next period is prohibited. The model must not post entries, record write-offs, or reach conclusions about individuals. The Skill explicitly states: assists reconciliation and does not constitute financial advice; sign-off remains the responsibility of qualified finance personnel.
- Outputs: Bank reconciliation statement (both sides brought to agreement); three-layer outstanding items schedule; aging report (amount / start date / category / owner); draft entries (if any); reviewer sign-off section.
- Source: GitHub: Anthropic Reconciliation Skill (official method template, not a live working paper from any company; source page shows 2026-03-13, with subsequent repository updates)
Accounting / Close / Controls
See Today’s Most Actionable Items item 2.
Inputs are the bank statement and GL as of the same cutoff date. The model only classifies items, assigns aging, and drafts reconciling entries. Humans handle only Layer 2, Layer 3, and items above threshold. Posting, write-offs, and period-close sign-off continue to follow existing authorization rules. The sole condition for expanding scope is running the same closed dataset twice with consistent classification and accounting’s ability to identify every flagged item.
FP&A / Planning / Reporting
See Today’s Most Actionable Items item 1.
Lock formulas and reporting definitions first, then allow the model to change assumptions only. Operating review materials must trace from the assumption change log back to source files. The approach can extend to an initial variance-driven draft from month-end actuals, but FP&A owner must still revise it; do not activate all three Agents in parallel this week.
Treasury / Cash / Risk
Data unavailable. No recent public cases from the past 365 days with both full text and reusable steps were identified for cash forecasting, bank transaction monitoring, or DSO/O2C. Item 2 covers only reconciliation and is insufficient for a standalone cash position report.
Tax / Compliance / Audit
Data unavailable. No new AI implementation cases or practical methods for tax research, SOX/internal controls, or audit evidence management from the past 365 days were identified.
CFO / Leader Team-Building Experience
Data unavailable. No new public finance leader shares from recent days that clearly describe division of responsibilities, review mechanisms, and performance metrics were identified. Compensation surveys or vendor product materials were not used to fill gaps.
LinkedIn data unavailable / verification not passed. No cross-verifiable startup finance headcount alternative cases were identified this period.
Open Source / AI Engineering References
See Today’s Most Actionable Items item 2.
The only directly replicable structures are: reconciliation type (GL–subledger / bank / intercompany) split by file or prompt; outstanding items must include category, aging, and owner; escalation thresholds written into instructions rather than decided ad hoc. This is a method template, not a production accounting system. No one may use it to post entries directly.
Items Requiring Verification
- One source claims reconciliation prompts must reference specific Drive exports (GL detail, AP aging, bank statements, sub-ledgers) and hard-code materiality; every finding must include entry number, account, user, and amount; exception scanning should return only items for review and must not reach conclusions about individuals. The prompt itself is shared only via DM; no public file was located. X: BojanRadojici10 (2026-08-25; single social media source / pending verification)
Small Experiments for This Week
- One cash account, three-layer reconciliation, no postings permitted: Follow item 2 using only the prior closed period’s bank statement and GL. Model produces matching / timing / adjustment-required / investigation-required classifications. Accounting samples 10 Layer 3 items to assess false-positive rate. If misclassifications exceed three items, adjust thresholds first; do not add a second account. No postings permitted by anyone.
- One planning file, assumptions only: Follow item 1. Select one actual headcount plan or supplier price update. Model lists only the assumption cells to change. FP&A opens the model and confirms formula areas untouched and reconciliations intact. Any formula changes or untraceable items result in rejection.
- One assumption change log, not for meeting distribution: Convert the output of experiment 2 into a single page showing old value / new value / source reference / gross margin impact. CFO or FP&A owner changes one assumption and observes whether the log updates correctly. If traceability is unclear, the log is not distributed for the operating review.