MiralonMa/interpret-course-subtitles

Accurate Interpretation of Course Subtitles

The goal is to base subsequent answers on a reliable understanding of the course content. Applicable to courses in science, engineering, humanities, and social sciences. It does not come pre-loaded with a fixed replacement dictionary for any specific subject. By default, the skill directly fulfills the user's requested task—such as Q&A, summarization, or teaching—while identifying terminology as it reads the subtitles. It only generates a copy of the transcript when the user requests a corrected file. The original files and location information are preserved.

Step 1: Identify the Course and Existing Materials

1. Read the project conventions and context already provided by the user. Look for subtitles, course syllabi, textbooks, lecture notes, course slides, course glossaries, and any previously selected courses within the user-specified project. Prioritize locating files based on the file manifest and directory structure; do not scan unrelated disks. Support the case where the user only provides a subtitle fragment; do not require a full project or file.
2. Distinguish between a new lecture within the same course and a completely new course. Only verify the course identity if a different course has the same name but a different instructor, course level, or content structure that would affect interpretation. Do not re-initialize the course simply because the chapter number has changed.
3. For each new course, ask if the user wishes to create a dedicated glossary in the project directory. If the user has already made a choice for that course, or has explicitly provided an existing glossary for it, follow that choice and do not ask again. The installation of this general skill does not imply user consent to create glossaries for all courses. You may ask the user: "Would you like to create a dedicated glossary for '[Course Name]' in the project directory to reuse verified terms and record candidates for future verification?" Create the glossary only after receiving an affirmative answer. Its location and structure are described in the Glossary Convention. If the user refuses or has not yet responded, maintain a temporary mapping within the current analysis and continue with any feasible reading tasks; do not treat the wait as consent. If no project directory is provided, do not create a glossary in an unrelated location.
4. Check first, then decide whether to ask for reference materials. If you have already found relevant and usable textbooks, lecture notes, or slides for the course, use them first and avoid asking a generic question about adding references. Only ask once if no usable reference material is available: "No textbooks, lecture notes, or slides for this course were found in the project. Would you like to add reference materials, or shall I proceed using the subtitle context and publicly available information for now?" Combine this question with the glossary question to avoid excessive interruptions. If the user has already indicated that they do not want to add materials, do not ask again unless new circumstances arise. If materials exist but are inaccessible, corrupted, or entirely illegible, describe the specific situation and then ask for a usable version; do not claim that the project lacks material.
5. While waiting for a reply, feel free to identify the topic, compile points of doubt, and read the available material. If the user has requested a wait for specific materials, pause conclusions that rely on those materials. Ask further clarifying questions only on points that affect the outcome of the current task; do not ask for every typographical error.

Build a Re-checkable Context

• Preserve the filename, lecture number, original timestamps, or subtitle sequence numbers. If no timestamp exists, use line or paragraph numbers; do not fabricate timestamps. For SRT/VTT, keep the original cue identifiers. For TXT, do not arbitrarily add end times.
• First, grasp the topic and logical flow of the relevant chapter before interpreting ambiguous points. Read short texts in their entirety; for long lectures, read thematically and maintain overlap between sections. Keep a brief record of the course topics, candidate terms, their sources, and outstanding issues. When you have only read part of the content, clearly state the coverage; do not claim the entire text has been verified.
• Start from the complete argumentative paragraph that contains the ambiguous point. If necessary, expand to the same chapter, to earlier definitions, and to later review sections. Only expand further if the meaning remains ambiguous after examining the immediately surrounding sentences; do not mechanically perform a full-text search for every sentence.
• Search for repeated occurrences of the same concept, phonetically similar words, code-mixed segments (Chinese/English), abbreviations, and symbols. Repeated ASR errors are clues, not independent pieces of evidence. Do not treat a previous guess as a source for subsequent verification.
• Cross-sentence reconstruction is for understanding only; do not alter the original sequence. Distinguish between instructor statements, student responses, quotations, counterexamples, questions, and subsequent self-corrections. Colloquial repetitions, noise, and automatically inserted English words should not be automatically elevated to technical terms.

Propose Candidates for Ambiguous Points and Verify Them

For ambiguous points that affect understanding, retain the original text and propose a small number of discriminative candidates. Filter candidates based on the following evidence, rather than simply replacing them based on pronunciation similarity or search popularity:

• Whether the pronunciation or written form can explain the recognition error. Does the reading of a Chinese/English name or abbreviation correspond?
• Whether the candidate fits the topic, grammar, and anaphora of the paragraph. Does the course actually introduce this concept?
• Whether subsequent definitions, example problems, formulas, diagrams, or reviews support the candidate. Can it explain other occurrences of the term within the same course?
• Whether the candidate is valid within the current discipline, model, era, or theoretical framework. What evidence refutes it?

Clarify the purpose of the evidence:

Question to Answer	Priority Basis	What Cannot Be Claimed From This
What definition, notation, and conventions does the course use?	The user-specified order of course sources. Otherwise, refer to the official course materials and clear context.	An externally common definition is necessarily applicable to this course.
Does this term exist, and what does it usually mean?	Reliable sources: textbooks, university course materials, official standards, original papers.	The instructor said this exact word at this moment.
What is the most reasonable interpretation here?	Cross-referencing multiple contexts and course materials, having ruled out the main competing candidates.	The original audio has been checked and the exact words have been restored.
What did the instructor actually say?	The corresponding, actually audible recording; a clear, locatable source from the course; or user confirmation.	The original words can be confirmed by disciplinary common sense alone.

Course materials take precedence in determining the course context, but this does not imply the materials are free of errors. When conflicts arise, report the discrepancies and the basis for your interpretation; do not silently align the course with other textbook systems. When scanning PDFs, slide images, formulas, or circuit topologies, examine the original, legible image. OCR provides only clues; do not fabricate relationships from blurry images.

Conduct Public Searches When Necessary

When terms are obscure, candidates remain conflicting, factual accuracy is uncertain, or the user requests verification, search for the candidate terms, their English names, their discipline, and key features within their definitions. Propose the candidates before searching for evidence to avoid confirmation bias by searching for only one candidate.

• Prioritize official course websites, university course pages, legitimate public content from designated textbooks, standards bodies, and original papers. Select sources based on the specific discipline. For fields like medicine or law, also distinguish by era and scope of applicability.
• Open and read the body text that supports a judgment. Record the title, link, and relevant section or page numbers. Summaries, search snippets, multiple repostings, and AI-generated paraphrases cannot serve as independent, definitive conclusions.
• Only search for the necessary concepts; do not send complete, non-public subtitles, course login details, or personal information. Do not automatically upload entire documents to online OCR, ASR, or other conversion services for verification.
• If reliable sources cannot be found or the internet is unavailable, lower the strength of your conclusions and clearly state the gaps in evidence. If one round of extended context and targeted source verification still cannot distinguish between candidates, preserve the ambiguity. Only loop back if new, discriminative evidence is found.

Independently Check High-Impact Content

Negation words, comparison direction, conditions, numerical values, decimal points, units, positive/negative signs, superscripts, subscripts, upper/lower case, formulas, proper nouns, and dates should not be corrected simply because "it should be so by common sense." Distinguish between instructor slips of the tongue, classroom counterexamples, student mistakes, and ASR errors. If the original text contradicts the theory, clearly indicate the conflict and explain “as stated in the subtitle” vs. “the common theory” separately. Do not rewrite the instructor's point of view.

Terms can be expressed in the following states; these are levels of evidence, not measured probabilities:

• Original wording / course usage verified: Clearly record the listening location, the corresponding clear material, or user confirmation. Do not claim that a term is verified just because it appears in the textbook.
• High-confidence interpretation: The context and course/discipline materials corroborate each other, and the main competing candidates have been ruled out. This interpretation can be used to continue answering, but should not be presented as a word-for-word quotation.
• Pending verification: There are still conflicting candidates that affect the meaning. Provide the candidates, the reasons for distinguishing them, and the missing evidence. Frame the conclusion in a conditional manner.
• Unintelligible: Insufficient information. Do not fabricate a sentence. Point out the original location.

Local confidence in a term does not imply the whole sentence or formula is correct. Use new evidence to re-verify earlier interpretations and their affected conclusions. Downgrade, retract, or overturn previous judgments when faced with counter-evidence. Repeated reading without new evidence does not upgrade confidence. Do not promise a "95% recognition accuracy."

Default Output and Delivery on Demand

By default, output the answer requested by the user. Corrections with high confidence and low risk can be naturally applied to explanations. For an obviously incorrect word that appears for the first time and affects meaning, briefly explain: "Here it is understood as..." and include a timestamp or line number. Only elaborate on the crucial corrections, sources, and pending points that are relevant to the final conclusion. Do not force output a line-by-line correction log or a complete new subtitle.

For content that is still pending verification, present the most discriminative candidates and their impact. Do not include an uncertain premise in a definitive conclusion. Provide direct links for public sources next to the relevant judgment. For local materials, specify the file and locatable position. The instructor's original words, inferred interpretations, external knowledge, and supplementary explanations by the agent must be distinguishable.

If the user agreed to maintain a glossary, only record course evidence that can be re-traced and valuable points of doubt, updating according to Glossary Convention. Do not write data from a single course into the skill's own operational code, a global dictionary, or global memory.

Only export a corrected copy, revision comparison, or correction report upon an explicit request. Retain the original subtitles with their timestamps/sequence numbers and generate a new file in the original format. Do not overwrite the original text. Do not perform a global phonetic replacement. In a "verbatim correction," do not delete colloquial speech, reorder arguments, or insert textbook content. When a "smooth lecture notes" version is requested, clearly label it as a reorganized text.

Additional Means and Completion Check

The most valuable supplementary evidence is usually a clear original slide, a term index from the designated textbook, a subsequent explanation from the same course, and a short re-listening of key audio passages. When authorized and the tools are available, you may partially re-listen or use course-specific terminology to prompt a secondary ASR run. The second recognition result is still only a clue. Do not initiate batch transcription, paid services, software downloads, or full-course uploads by default, and do not force the user to provide audio/video.

Pre-delivery checks: Are counter-evidence, negations, and conditions missing from the candidates? Have model inferences been turned into direct quotations? Are the locations of ambiguous points preserved? Has a full-text consistency check led to unconditional replacements? Has an unapproved glossary been inadvertently created? Examples of complex ambiguities can be found in Decision Examples; only read them when needed.