Deck RefreshSAFE

An agent skill published by anthropics.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: deck-refresh
description: Updates a presentation with new numbers — quarterly refreshes, earnings updates, comp rolls, rebased market data. Use whenever the user asks to "update the deck with Q4 numbers", "refresh the comps", "roll this forward", "swap in the new earnings", "change all the $485M to $512M", or any request to swap figures across an existing deck without rebuilding it.
---

# Deck Refresh

Update numbers across the deck. The deck is the source of truth for formatting; you're only changing values.

## Environment check

This skill works in both the PowerPoint add-in and chat. Identify which you're in before starting — the edit mechanism differs, the intent doesn't:

- **Add-in** — the deck is open live; edit text runs, table cells, and chart data directly.
- **Chat** — the deck is an uploaded file; edit it by regenerating the affected slides with the new values and writing the result back.

Either way: smallest possible change, existing formatting stays intact.

This is a four-phase process and the third phase is an approval gate. Don't edit until the user has seen the plan.

## Phase 1 — Get the data

Use `ask_user_question` to find out how the new numbers are arriving:

- **Pasted mapping** — user types or pastes "revenue $485M → $512M, EBITDA $120M → $135M." The clearest case.
- **Uploaded Excel** — old/new columns, or a fresh output sheet the user wants pulled from. Read it, confirm which column is which before you trust it.
- **Just the new values** — "Q4 revenue was $512M, margins were 22%." You figure out what each one replaces. Workable, but confirm the mapping before you touch anything — a "$512M" that you map to revenue but the user meant for gross profit is a quiet disaster.

Also ask about **derived numbers**: if revenue moves, does the user want growth rates and share percentages recalculated, or left alone? Most decks have "+15% YoY" baked in somewhere that's now stale. Whether to touch those is a judgment call the user should make, not you.

## Phase 2 — Read everything, find everything

Read every slide. For each old value, find every instance — including the ones that don't look the same:

| Variant | Example |
|---|---|
| Scale | `$485M`, `$0.485B`, `$485,000,000` |
| Precision | `$485M`, `$485.0M`, `~$485M` |
| Unit style | `$485M`, `$485MM`, `$485 million`, `485M` |
| Embedded | "revenue grew to $485M", "a $485M business", axis labels |

A deck that says `$485M` on slide 3, `485` on slide 8's chart axis, and `$485.0 million` in a footnote on slide 15 has three instances of the same number. Find-replace misses two of them. You shouldn't.

**Where numbers hide:**
- Text boxes (obvious)
- Table cells
- Chart data labels and axis labels
- Chart source data — the numbers driving the bars, not just the labels on them
- Footnotes, source lines, small print
- Speaker notes, if the user cares about those

Build a list: for each old value, every location it appears, the exact text it appears as, and what it'll become. This list is the plan.

## Phase 3 — Present the plan, get approval

**This is a destructive operation on a deck someone spent time on.** Show the full change list before editing a single thing. Format it so it's scannable:

```
$485M → $512M (Revenue)
  Slide 3  — Title box: "Revenue grew to $485M"
  Slide 8  — Chart axis label: "485"
  Slide 15 — Footnote: "$485.0 million in FY24 revenue"

$120M → $135M (Adj. EBITDA)
  Slide 3  — Table cell
  Slide 11 — Body text: "$120M of Adj. EBITDA"

FLAGGED — possibly derived, not in your mapping:
  Slide 3  — "+15% YoY" (growth rate — stale if base year didn't change?)
  Slide 7  — "12% market share" (was this computed from $485M / market size?)
```

The flagged section matters. You're not just executing a find-replace — you're catching the second-order effects the user would've missed at 11pm. If the mapping says `$485M → $512M` and slide 3 also has `+15% YoY` right next to it, that growth rate is probably wrong now. Flag it; don't silently fix it, don't silently leave it.

Use `ask_user_question` for the approval: proceed as shown, proceed but skip the flagged items, or let them revise the mapping first.

## Phase 4 — Execute, preserve, report

For each change, make the smallest edit that accomplishes it. How that happens depends on your environment:

- **Add-in** — edit the specific run, cell, or chart series directly in the live deck.
- **Chat** — regenerate the affected slide with the new value in place, preserving every other element exactly as it was, and write it back to the file.

Either way, the standard is the same:

- **Text in a shape** — change the value, leave font/size/color/bold state exactly as they were. If `$485M` is 14pt navy bold inside a sentence, `$512M` is 14pt navy bold inside the same sentence.
- **Table cell** — change the cell, leave the table alone.
- **Chart data** — update the underlying series values so the bars/lines actually move. Editing just the label without the data leaves a chart that lies.

Don't reformat anything you didn't need to touch. The deck's existing style is correct by definition; you're a surgeon, not a renovator.

After the last edit, report what actually happened:

```
Updated 11 values across 8 slides.

Changed:
  [the list from Phase 3, now past-tense]

Still flagged — did NOT change:
  Slide 3 — "+15% YoY" (derived; confirm separately)
  Slide 7 — "12% market share"
```

Run standard visual verification checks on every edited slide. A number that got longer (`$485M` → `$1,205M`) might now overflow its text box or push a table column width. Catch it before the user does.

## What you're not doing

- **Not rebuilding slides** — if a slide's narrative no longer makes sense with the new numbers ("margins compressed" but margins went up), flag it, don't rewrite it.
- **Not recalculating unless asked** — derived numbers are the user's call. Your Phase 1 question covers this.
- **Not touching formatting** — if the deck uses `$MM` and the user's mappin

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.

[![ModelsAtlas audit: Deck Refresh](https://modelsatlas.com/skills/anthropics/deck-refresh/badge.svg)](https://modelsatlas.com/skills/anthropics/deck-refresh/)

Not your listing, or something out of date? Claim or update your listing.