lecture-condenser

name	lecture-condenser
description	Condense multiple lecture-notetaker .md files from one series into a flat pre-exam priming document of recognisable key points, grouped by lecture and topic. Recognition tool for fast scan before exams — not full learning notes or flashcards.
disable-model-invocation	true

Lecture Condenser

Invocation

Manual only. Apply this skill only when the user invokes /lecture-condenser or explicitly selects it. Do not load or run it based on conversation context alone.

Produce a flat priming document of key points from a set of lecture notes in one series. Recognition tool, not a learning artefact. The user scans it before an exam to dispel the "I don't know what I know" feeling — each line is something they should recognise, and not recognising one is the signal to revisit the source .md.

Before starting

1. Read /mnt/skills/public/pdf-reading/SKILL.md for slide handling.
2. Read references/condenser-template.md — full output structure with a worked example. Follow it closely.
3. Read references/key-point-criteria.md — the mechanical filter for what counts as a key point.

Inputs required

Input	Format	Required?
Lecture notes	2+ .md files from lecture-notetaker	Yes
Slides	.pdf per lecture	Yes — used for extraction, not in output
Series name	String	Yes — ask if not provided

If fewer than 2 .md files are provided, confirm with the user — the condenser is designed for series-level review, not single-lecture review.

If slides are missing for any lecture, proceed with .md only for that lecture and flag in the output frontmatter that the slide signal was unavailable for that lecture.

Workflow

1. Confirm series scope

Ask: which series, which lectures, and confirm all .md files are from the same series. If they span topics, recommend running the skill once per series — the priming function works best when the user is loading one coherent topic into memory at a time.

2. Parse inputs

For each lecture:

• Parse the .md: extract learning objectives, key claims under each, definitions, mechanisms, callout-flagged content, ⚠️ markers, and any explicit exam-relevance flags.
• Extract slide text via /mnt/skills/public/pdf-reading.
• Note slide titles, bolded text, and bullet structure — these inform the key-point filter.

3. Apply key-point filter

A claim qualifies as a key point if it satisfies any two of:

• Appears on a slide as title, bullet, or bolded term
• Referenced in the LO list of the source .md
• Has an associated mechanism, definition, or numerical value (recallable as a discrete unit)
• Appears in more than one lecture in the input set

See references/key-point-criteria.md for examples and edge cases. When in doubt, drop the claim — the priming function is destroyed by bloat.

4. Group and order

• Spine: lecture series → individual lecture → topic cluster within lecture
• Topic clusters come from how the slides organised the content, not from LOs (LOs are a learning-time construct; for review, slide structure matches the user's mental map of the lecture)
• No cross-lecture deduplication — if a concept appears in three lectures, it appears three times. That repetition is itself information.

5. Compress to one-line recognition units

Each point is a single line. Mechanism, definition, classification, threshold, named entity. No prose. No explanation. Recognition, not exposition.

Source claim → recognition unit examples in references/condenser-template.md.

6. Generate the markdown

Follow references/condenser-template.md exactly. Flat structure — no foldable callouts, no PDF++ embeds, no questions, no answers, no spaced review schedule.

7. Deliver

Save the .md and present.

What this skill is NOT

• Not a summary you read to learn (that's lecture-notetaker full mode)
• Not flashcards or active recall (different cognitive function — recognition vs recall)
• Not a place to add explanation or context (that's lecture-ftg)
• Not exhaustive — claims that don't pass the two-signal filter are out, even if true

Why no questions, no callouts, no embeds

The priming function depends on fast scan. Anything requiring interaction (folding a callout, clicking through to a slide, answering a question before continuing) breaks the scan. The user reads top-to-bottom in one sitting, recognises everything, and closes the doc. Different artefact from lecture-notetaker's full-mode notes, which are designed for active engagement during learning.

Edge cases

Situation	Action
One lecture only	Confirm with user; condenser is series-level by design — single-lecture review is what lecture-notetaker already produces
Lectures from different series	Flag and recommend running once per series
Slides missing for a lecture	Drop the slide signal for that lecture; require both other signals to qualify a claim
Source .md is sparse	Flag at top of output, proceed with what's available
100+ key points	Generate as-is. The series is what it is — don't truncate
Same claim appears twice in one lecture	Include once (intra-lecture dedup); cross-lecture is preserved

References

File	When to read
references/condenser-template.md	Always — before generating
references/key-point-criteria.md	Always — for the filter logic