Lecture Slides Summarizer

Summarize Slides

Turn a long lecture PDF into a concise, exam-oriented LaTeX cheat sheet and compiled PDF with accurate page references, bilingual terminology, and only the diagrams that materially improve understanding.

Core principle: treat slide summarization as a controller workflow for long documents: read globally first, segment by lecture or topic, dispatch focused readers for each segment, merge into a single cheat-sheet draft, then run a fresh coverage verifier before building final artifacts.

Non-Negotiables

• REQUIRED COMPANION SKILL: use pdf for PDF intake and reading. This workflow is not optional.
• NO OTHER SKILLS: when this skill applies, use only summarize-slides plus pdf. Do not invoke dispatching-parallel-agents or any other additional skill. Any task-splitting, reader dispatch, verifier separation, or parallelization logic needed for this workflow is already defined inside this skill.
• Check that locally installed xelatex is available before promising the default required compiled PDF output.
• Default deliverables are .tex and compiled .pdf, with the PDF treated as the required final artifact unless the user explicitly asks not to produce it, and the default compiled PDF path requires xelatex.
• When this skill applies, a chat-only or inline summary is not a successful final deliverable unless the user explicitly asks for summary-only output or explicitly opts out of file creation.
• If the user gives parsed text, OCR text, extracted images, or page-by-page content originating from a lecture PDF instead of the raw PDF file, treat it as the same slide-summarization task and still produce the document artifacts by default.
• By default, place all generated artifacts in a single folder named Summary - <pdf-stem> inside the same directory that contains the source PDF. Do not scatter .tex, .pdf, helper assets, or intermediate files elsewhere.
• This containment rule is strict: create the output folder before any extraction, note-taking, OCR, or build step, and write every generated file into that folder from the start.
• Extracted text files are included in this rule. Files such as <pdf-stem>_full.txt, <pdf-stem>_p1-5.txt, OCR dumps, page-range extracts, verifier notes saved to disk, temporary LaTeX snippets, and any other intermediate .txt, .md, .json, .tex, image, or helper outputs must live inside Summary - <pdf-stem> only.
• Never place generated files next to the source PDF unless the user explicitly asked for that location. The source PDF's directory is not a scratch space.
• Here, same directory that contains the source PDF means the output folder sits alongside the PDF file itself. Example: if the source PDF is .../lecture/foo.pdf, the default output folder is .../lecture/Summary - foo/.
• The default writing style is Chinese-first.
• Important technical concepts, terms, and named distributions/theorems must include the original English in parentheses.
• Each summarized point should include the source slide page number or a tight page range by default. Omit citations only if an extraction failure is explicitly reported.
• The output style is cheat sheet in the sense of coverage, brevity, clarity, and fast lookup. It does not mean forced ultra-dense layout, automatic two-column formatting, or unreadable compression.
• Default layout is normal single-column reading unless the user explicitly asks for another layout.
• Do not produce a generic chapter summary when the user asked for exam points.
• Include formulas, important concepts, representative examples, and problem-solving techniques where they matter.
• Do not silently drop later sections of a long PDF because of context or time pressure.
• If extraction quality is too poor to preserve formulas, symbols, or page mapping, stop and ask rather than paraphrasing uncertain mathematics.
• If an important concept genuinely needs a figure to explain it clearly, include a LaTeX-drawn explanatory figure in the final document.
• Do not use screenshots or copied slide images as a substitute for a requested or necessary explanatory figure unless the user explicitly permits that fallback.

Skill Boundary

• Allowed skills for this workflow: summarize-slides and pdf only.
• Forbidden: all other skills, even if they appear relevant to planning, dispatch, verification, or writing.
• Task or subagent mechanisms may still be used when the platform supports them, but that is an execution detail and does not justify loading another skill.

Required Outcome

The default successful outcome for this skill is:

• a saved LaTeX source file named Summary - <pdf-stem>.tex
• a compiled PDF file named Summary - <pdf-stem>.pdf
• all generated files grouped in one output folder named Summary - <pdf-stem> inside the same directory as the source PDF unless the user explicitly requested another location
• no generated intermediate, extracted-text, OCR, notes, or helper files left outside that output folder
• a final response that reports the artifact paths and any blockers encountered

If no raw PDF path is available and the input is only source-derived text, OCR, extracted images, or page-by-page content:

• use the user-provided lecture name as the naming basis if one exists
• otherwise use the fallback stem lecture-slides
• default the output folder to Summary - <chosen-stem> in the current working directory unless the user specified another location
• name the artifacts Summary - <chosen-stem>.tex and Summary - <chosen-stem>.pdf

The following are not successful completions when this skill applies unless the user explicitly requested them:

• an inline chat summary only
• a summary draft that was never written to disk
• a claim that the work is complete before LaTeX was written and the PDF was compiled
• stopping at analysis, segmentation, or merged notes

When to Use

Use this skill when the user wants any of the following from a PDF slide deck, lecture PDF, lecture notes PDF, or 课件 PDF:

• a study sheet
• a cheat sheet
• a key-points summary
• a formula sheet
• a concise review PDF
• exam-point extraction from slides
• 总结这个课件
• 把这个课件整理成复习资料
• 提炼这份 slides 的考点
• 做一个公式和概念速查表
• 做成考试前看的 cheat sheet
• 总结 lecture PDF
• 总结 slides
• 提炼课程讲义里的重点

Common trigger words and phrases:

• 课件
• slides
• lecture PDF
• 复习资料
• 考点总结
• 总结一下这个课件
• 提炼考点
• 公式速查表
• 概念速查表
• cheat sheet
• 考前整理
• 考试重点

Requests like 总结一下这个课件 still default to the artifact-producing workflow above. Do not reinterpret them as permission to answer only in chat.

Do not use this skill for:

• homework solving
• essay writing
• article summarization where slide-page lookup is unimportant
• .pptx workflows that are not already exported to PDF

Portability

• Do not assume any custom skill other than pdf exists.
• This skill must be used together with pdf. Do not replace the pdf skill with an ad hoc fallback.
• Do not invoke any other skill to manage parallel work, planning, verification, or subagent dispatch. In particular, do not invoke dispatching-parallel-agents; follow the dispatch rules written here directly.
• If subagents are available and practical, prefer them for segment readers and a separate verifier. For very long decks, this is the preferred execution path when the platform supports it reliably.
• If subagents are not available, emulate the same roles sequentially: controller pass, per-segment read pass, merge pass, verifier pass.

Workflow Model

• the controller owns intake, global read, segmentation, merge, style normalization, build choice, and final reporting
• each segment-reader owns one lecture chunk or topic chunk only
• one fresh verifier checks coverage, page references, bilingual terminology, and exam usefulness after merge
• the controller never treats segment notes as already trusted final output

Dispatch Rules

This section replaces the need for any separate parallel-dispatch skill.

• Use one reader per independent lecture segment only when the segments can be understood without shared reasoning state.
• Good split examples: chapter blocks, explicit topic-title sections, or long independent page ranges.
• Do not split tightly coupled derivations, one multi-page proof chain, or one example whose steps depend heavily on each other across pages.
• When several segments are independent, dispatch them in parallel if the platform supports parallel subagents.
• When segments are not independent, process them sequentially even if parallel execution is available.
• The verifier must be a fresh pass with fresh context. Do not reuse one of the segment readers as the verifier.
• Each dispatched reader must receive only its own page range, the fixed return format, and the instruction not to write the final document.
• After all readers return, the controller must review for conflicts, merge them, and then hand the merged draft to a separate verifier pass.

Execution Flow

digraph summarize_slides_flow {
    "Intake" [shape=box];
    "Read PDF globally" [shape=box];
    "Segment by lecture/topic" [shape=box];
    "Dispatch segment readers" [shape=box];
    "Collect per-segment notes" [shape=box];
    "Merge into cheat-sheet draft" [shape=box];
    "Add needed LaTeX figures" [shape=box];
    "Dispatch fresh verifier" [shape=box];
    "Verifier approved?" [shape=diamond];
    "Write LaTeX" [shape=box];
    "Build PDF" [shape=box];
    "Final artifact check" [shape=box];
    "Report result" [shape=doublecircle];

    "Intake" -> "Read PDF globally";
    "Read PDF globally" -> "Segment by lecture/topic";
    "Segment by lecture/topic" -> "Dispatch segment readers";
    "Dispatch segment readers" -> "Collect per-segment notes";
    "Collect per-segment notes" -> "Merge into cheat-sheet draft";
    "Merge into cheat-sheet draft" -> "Add needed LaTeX figures";
    "Add needed LaTeX figures" -> "Dispatch fresh verifier";
    "Dispatch fresh verifier" -> "Verifier approved?";
    "Verifier approved?" -> "Merge into cheat-sheet draft" [label="no, revise and re-verify"];
    "Verifier approved?" -> "Write LaTeX" [label="yes"];
    "Write LaTeX" -> "Build PDF";
    "Build PDF" -> "Final artifact check";
    "Final artifact check" -> "Report result";
}

Step 1. Intake

• Confirm the source PDF path, or if no raw PDF is available, confirm the source-derived input and the naming basis for outputs.
• Confirm the output directory if the user already specified one.
• Determine expected outputs: .tex, .pdf, or both. Default is both, and the PDF is the required final artifact unless the user explicitly opts out of PDF output.
• Determine whether the user wants the whole deck or a subset. Default is the whole deck.
• Record any style instructions already provided by the user.
• If the user did not specify an output path, create and use a deterministic folder named Summary - <pdf-stem> inside the same directory as the source PDF and continue rather than waiting for unnecessary clarification.
• Treat creation of that folder as the first filesystem step of the workflow. Do not write extraction files to the source PDF directory and move them later; write them into the output folder immediately.

If there is no raw PDF path, choose the output stem using the rule from Required Outcome and create the deterministic output folder from that stem instead of waiting for clarification.

Default style unless the user says otherwise:

• language: Chinese
• layout: single-column normal reading
• purpose: exam review
• tone: concise, clear, non-verbose
• output: Summary - <pdf-stem>.tex and Summary - <pdf-stem>.pdf inside the folder Summary - <pdf-stem> that sits in the same directory as the source PDF

Step 2. Check Build Environment

• Verify xelatex exists before promising the default required compiled PDF.
• The default Chinese-first compiled PDF workflow requires xelatex for Chinese text and math.
• If xelatex is missing, stop and report that the environment prerequisite for the default required PDF output is missing unless the user explicitly changes the output expectation away from compiled PDF. Do not stop at .tex only when PDF output is still required.

Step 3. Read the PDF Globally First

Do not jump straight into chunk summaries.

• If any extracted text, OCR output, or saved notes are written to disk during this step, they must be written inside the output folder, never beside the source PDF.
• Read enough of the whole PDF to understand its macro-structure.
• Identify title pages, agenda pages, lecture separators, topic transitions, and summary slides.
• Build a segmentation map before deep reading.

For very long decks, segmentation plus per-segment reading is mandatory. Do not summarize a few-hundred-page deck in one shot.

Treat a deck as very long whenever one-pass reading would plausibly cause dropped coverage, unstable page mapping, or weak recall of later sections. In practice, any deck with a few hundred pages should be treated this way.

For long decks, the controller must first produce:

• overall topic list
• segment boundaries by page range
• expected number of lecture/topic segments
• suspected formula-heavy or concept-heavy regions
• a segmentation map covering the full page range

The segmentation map must cover all pages exactly once unless a page is explicitly marked as intentionally excluded, such as a duplicate title slide or blank separator page.

Step 4. Segment the Deck

Use the coarsest segmentation that preserves topic integrity.

Prefer, in order:

1. lecture boundaries already present in the slides
2. explicit topic-title boundaries
3. major concept transitions

Do not split a tightly coupled derivation across multiple segment readers unless necessary.

If a segment boundary is not obvious from lecture titles or topic-title slides, record a short justification for that split.

Step 5. Segment Reader Phase

If subagents are available and practical, dispatch one fresh segment-reader per independent segment.

If subagents are not available, the controller must still create separate per-segment notes before any merged summary is written. Do not read and summarize the full deck in one fused pass.

If the proposed segments are not independent enough to read safely in parallel, keep the same per-segment structure but execute sequentially.

Each segment-reader should return exactly these sections:

• segment id
• page range covered
• main exam points
• core formulas
• key concepts with English terms
• important examples or recurring problem patterns
• problem-solving techniques or pitfalls
• figure candidates if any
• coverage concerns or uncertainty

Segment readers are not allowed to:

• write the final document
• decide the final layout
• skip page references casually
• collapse the work into a generic prose summary

Step 6. Merge into Cheat-Sheet Draft

The controller merges all segment notes into one coherent exam-review structure.

The target output is not a page-by-page narrative. It should be organized for fast review, typically by lecture or topic block, with concise bullets under each block.

Every major block should try to include:

• what to know
• formula(s)
• why it matters or when it appears in problems
• one representative example pattern if useful
• one practical solving hint if useful

Style Rules for the Draft

• Chinese main text.
• Important terms in English in parentheses right after the Chinese term.
• Short bullets, not large explanatory paragraphs.
• Prefer exam language like 定义, 公式, 判据, 技巧, 典型例题, 易错点.
• Avoid filler transitions and lecture-storytelling.
• Avoid copying slide sentences verbatim unless a definition must remain exact.
• If a point is too detailed for a cheat sheet, compress it into the minimum needed to recall the method.
• If a point becomes so compressed that it loses meaning, expand slightly instead of forcing density.

Page Reference Rules

• Attach a source page citation to each summarized point by default.
• Prefer exact pages.
• Use short page ranges only when one idea truly spans several nearby slides.
• Do not downgrade to very broad section-level ranges just to save effort.
• If page mapping is uncertain because extraction is ambiguous, say so explicitly and resolve it before finalizing. Missing citations are acceptable only when extraction failure is explicitly reported.