Oh My PPT Source Reading

Boundary

Use this skill only when the host prompt already provides source documents or source snippets.

Do not discover, infer, or guess source document paths. Use only the sourceDocumentPaths explicitly provided by the host prompt. If no source document path is provided, do not scan the project looking for one.

When to use

• sourceDocumentPaths are present.
• Retrieved source snippets are present.
• The task asks to preserve facts, metrics, terminology, conclusions, or page points from a reference document.
• A slide outline appears to come from a parsed reference document.

When not to use

• No source document path or source snippet is provided.
• The task is a pure visual/style edit that does not use source facts.
• The task is only about layout, animation, or chart mechanics with already supplied data.

Reading workflow

1. Treat retrieved snippets as an index into the source, not as final evidence.
2. Extract search terms from the current slide title, content points, user instruction, snippets, and any known entities/metrics.
3. Use the DeepAgents filesystem tool grep(pattern, path, glob) inside the provided sourceDocumentPaths to locate relevant headings, paragraphs, tables, dates, metrics, entities, and terminology. pattern is a literal string, not a shell regex; call grep multiple times for multiple search terms.
4. Use the DeepAgents filesystem tool glob(pattern, path) only if the provided path points to a directory or the host prompt explicitly says multiple matching source files may exist.
5. Use read_file only on targeted sections or line ranges found by snippets/grep. Do not read an entire long document into context at once.
6. If the initial grep/read did not yield enough source-grounded evidence, refine search terms and repeat grep -> targeted read. Try synonyms, broader or narrower scope, adjacent section headings, and known entities/metrics.
7. Base every exact fact, metric, date, name, and quoted conclusion on the inspected snippets and source passages — then build the slide into a full argument around them (analytical structure, reading conclusion), not just the bare evidence.
8. If an exact fact/metric/date/name is still missing after reasonable attempts, do not invent it — omit that specific claim or state it qualitatively. Do not leave the page sparse: expand it with analysis derived from the inspected material instead (see "When the inspected material is thin").

Example

Slide title: "Q3 Revenue Highlights" Content points: "YoY growth rate", "top product contribution", "regional breakdown"

1. Snippet mentions "revenue grew 15% YoY" -> search terms: revenue, 15%, YoY, Q3.
2. Call grep(pattern="revenue", path="/source.md"), then grep(pattern="15%", path="/source.md"), then grep(pattern="Q3", path="/source.md") -> matches near the Q3 finance section and regional table.
3. Call read_file(path="/source.md", offset=118, limit=60) only around those matches -> confirms the exact growth and regional values.
4. Build the slide with the confirmed facts. If "top product contribution" has no source backing, omit that exact claim (do not invent the product) — but keep the page full by expanding the angles you did confirm (see "When the inspected material is thin").

Evidence rules

• Preserve source terminology, product names, system names, roles, dates, metrics, conclusions, risks, decisions, and examples.
• Use only source facts relevant to the current slide. Do not move material for other slides into this slide.
• Do not invent exact facts, metrics, dates, system names, status claims, examples, risks, decisions, or conclusions.
• If a retrieved snippet conflicts with the source passage you inspected, trust the inspected source passage over the snippet.
• If the source conflicts with the page outline, follow the source facts. If the source conflicts with an explicit user instruction, follow the user instruction but avoid unsupported source claims.

When the inspected material is thin — expand into a full argument

A half-empty slide is a failure, not fidelity. When the inspected source for a slide is sparse (one number, one chart, a few facts), do not just render that bare evidence and leave the rest of the canvas blank. Expand the slide into a complete argument with analytical and presentational structure DERIVED FROM the inspected material:

• context or background the source implies
• comparison dimensions (vs baseline, vs other groups, vs a prior period)
• cause/effect or mechanism
• implications, consequences, or a "so what"
• evidence grouping, annotations, or a one-line reading conclusion

This is interpretation grounded in the source, NOT fabrication. The boundary:

• ✅ Allowed and expected: reasoning, structure, framing, and presentation derived from what you inspected — use it to make the slide feel complete and balanced (a complete argument, not more modules or a filled canvas).
• ❌ Forbidden: inventing exact facts, metrics, dates, names, quotes, or conclusions the source does not contain.

If an exact value is missing, state it qualitatively or omit that specific claim — but keep building the argument with the structure above so the slide feels complete and balanced.

Long-document discipline

• Read nearby sections progressively rather than large ranges. Prefer 50-80 lines around grep matches; avoid reading 200+ lines in a single call unless the section itself requires it.
• Keep notes mentally scoped to the current slide.
• Stop reading once the current slide has enough source-grounded evidence.
• Never use source snippets for one slide as evidence for another slide unless the source passage directly matches that other slide.

Output discipline

• Do not rewrite the reference document into a generic storyline, marketing narrative, consulting framework, or inspirational theme unless the user explicitly asks for that transformation. (This prevents drifting off the source's topic or genre — it does not forbid the analytical expansion above: comparison, implications, and a so-what that stay on-topic are expected.)
• Slide text should be concise and presentation-ready, but still traceable to source evidence.
• Compress long source passages; do not paste long verbatim excerpts.
• Preserve exact numbers and names only when they appear in the source passage you inspected.
• For charts/tables, use only source-backed labels and values. If exact values are unavailable, use qualitative wording instead of invented numbers.