Cgpt Presentation Prep

CGPT Presentation Prep

Use this skill before generating slides from a manuscript, paper, report, or research workspace when facts, values, figures, or claims need to be verified. The goal is to make the deck build source-grounded before any PPTX design starts.

This skill prepares the evidence. Use cgpt-uw-pptx-deck or the Presentations plugin after the user confirms the extracted facts and visual plan.

Core Rule

Every number or factual claim intended for a slide must trace to a specific source location. Do not estimate, calculate, infer, or fill placeholders with guesses unless the user explicitly asks for analysis beyond source extraction.

Intake

1. Identify the source set:
   • user-named manuscript, report, deck, spreadsheet, script output, or folder
   • project MEMORY.md, README, SESSION-LOG, task files, or handoff notes when present and relevant
   • modular .qmd/.md sections included by a master manuscript
   • appendices, tables, figures, and replication outputs
2. Determine whether this is extraction mode or verification mode:
   • Extraction mode: build a new ground-truth table from sources.
   • Verification mode: check existing slide numbers, notes, or memory tables against sources.
3. Ask for missing source files only if they are necessary and cannot be found locally.

Ground-Truth Extraction

Read the relevant source material before proposing slides. Extract only facts useful for presentation planning:

• research question, setting, contribution claims, and headline findings
• sample sizes, units, time periods, exclusions, and data sources
• key tables and the few values a presenter would highlight
• key figures, captions, source paths, and whether the files exist
• core formulas, model terms, properties, proofs, or worked examples
• caveats, limitations, and audience-sensitive qualifications

Use this compact table when possible:

If a needed value is absent, write NOT FOUND IN SOURCE rather than filling it in.

Figure And Visual Asset Audit

For each existing figure or candidate visual:

• record file path, format, dimensions if available, and whether it exists
• judge whether it will survive 16:9 projection
• flag tiny labels, dense legends, bad aspect ratios, low resolution, or inaccessible colors
• recommend whether to reuse, crop, regenerate, redraw as a diagram, or replace with a placeholder

For new visuals, classify them before creating anything:

• Category A: can be generated from verified source values or from conceptual structure.
• Category B: requires running code, accessing external data, or calculating values not present in the source.

Ask the user to confirm the extraction and the Category A/B split before substantial generation or deck building.

Talk Planning Output

After extraction, propose a slide-ready evidence map, not a full deck unless requested:

• talk type and likely duration if known
• 3-5 core audience takeaways
• preliminary section arc
• candidate slides with each slide's source-backed claim
• visuals available now and visuals requiring more data/code
• risks: unverified numbers, missing files, dense figures, or values that need replication scripts

Boundaries

• Do not edit project memory, manuscript files, or source data unless the user explicitly asks.
• Do not start deck generation until the user confirms the extracted values or clearly waives confirmation.
• Do not claim a figure exists without checking the path when local files are available.
• Keep quotations short and only as much as needed to anchor the extracted fact.

Handoff

When the prep pass is complete, report:

• number of values or claims confirmed
• number of values not found or needing follow-up
• figure and asset readiness
• exact source files inspected
• whether the project is ready for cgpt-uw-pptx-deck or another deck-build workflow