YouTube Breakdown

Turn a YouTube transcript into one of four structured breakdowns. The four modes are genuinely different tools, not reshuffled versions of the same output — Study Notes wants depth and hierarchy, Quick Reference wants ruthless compression, Critique wants adversarial engagement, Actionable Insights wants a plan. Picking wrong wastes the run.

So: always ask which mode before generating. Never auto-pick.

Mode Selection

Present the four options:

Which mode?
1. Study Notes — deep, hierarchical, for videos worth retaining long-term
2. Quick Reference — ruthless distillation, for triage or recall-at-a-glance
3. Critique — steelman then dismantle, for skeptical engagement
4. Actionable Insights — actions + frameworks separated, for tactical content

When the trigger phrase implies a mode ("critique this video" → Mode 3, "what should I do with this video" → Mode 4, "study notes for this" → Mode 1), still confirm rather than assume: "Sounds like Critique mode — confirm or pick a different one?" One turn is cheap; a 900-word breakdown in the wrong register is not.

Auto-router (deliberately off). A pre-flight classifier that reads the first ~500 words and recommends a mode is specced under Future Extensions. Kyle's standing preference is explicit choice. Revisit only if mode-picking starts to feel like friction.

Input Handling

Don't hand-roll URL fetching. WebFetch against a third-party transcript site was the old approach and it fails silently — those sites rate-limit, restructure, and return a page of boilerplate that looks like a short transcript. youtube-transcript goes to the source.

If the input is ambiguous — a non-YouTube URL, or text too short to plausibly be a transcript — ask one clarifying question. Don't guess.

Long transcripts (>30k words): say so up front and offer to chunk-process — run the mode against sections, then stitch — rather than quietly producing a breakdown that thins out toward the end.

Output Handling

1. Display the breakdown inline. Full output, properly formatted. This always happens and works anywhere, including cloud sessions with no filesystem.
2. Append the mode's handoff line (each mode below names its own). It is an offer: never act on it without a yes. "Don't act on it" is not "end the turn" — keep going to step 3 in the same turn, and let Kyle answer both the handoff and the save question together.

3. Offer to save, and ask where. Kyle wants the choice every time, so propose a default and list the alternatives in one turn:

   Save this? Default is ~/Learning/youtube-notes/2026-08-19-<title>-study-notes.md. You also have teach workspaces at ~/Learning/party-line/, ~/Learning/gearhead-products/, ~/Learning/prime-agent/ — or name another path.

   Read ~/Learning/ before offering, so the workspaces you list are the ones that actually exist rather than the three named above, which are only an example of the shape. Filing a breakdown into a teach workspace puts it where /teach and /teach-research will find it, which is often the point.

4. If Kyle declines, leave it inline. Don't push, don't ask twice.

Filename and frontmatter

Slug: YYYY-MM-DD-{kebab-case-video-title}-{mode}.md — e.g. 2026-08-19-andrej-karpathy-on-llms-study-notes.md. If the title is unknown, use YYYY-MM-DD-youtube-breakdown-{mode}.md and offer to rename once the title is known.

---
title: "[video title]"
source: "[URL if available, else: pasted transcript]"
speaker: "[if identifiable, else: unknown]"
date_processed: [today, YYYY-MM-DD]
mode: study-notes | quick-reference | critique | actionable-insights
tags: [youtube, breakdown, {mode}]
---

The three free-text fields are quoted, and that is not cosmetic. Talk titles are full of colons — "Andrej Karpathy: Software Is Changing Again" — and a bare title: Andrej Karpathy: Software Is Changing Again is a YAML parse error, not a quirk. The whole frontmatter block then reads as nothing to every parser that touches it. Any " inside the value gets escaped as \".

The mode value is one of those four kebab-case strings exactly — not capitalized, not pluralized. It's what makes a folder of breakdowns sortable later, and a folder where half the files say study-notes and half say Study Notes doesn't sort.

Mode 1: Study Notes

For: videos worth retaining long-term — technical lectures, deep educational content, frameworks to reference later.

You are an expert study-notes generator. Internalize the transcript
deeply enough to teach it, then produce a hierarchical outline
optimized for learning and retention.

# PHASE 1: COMPREHENSION (internal, do not output)
1. Identify the video's domain, the speaker's expertise, and the
   core thesis.
2. Map the logical structure: what claims are made, what evidence
   supports each, and how ideas connect.
3. Flag areas where the speaker glosses over nuance, makes
   unsupported claims, or omits obvious counterpoints.

# PHASE 2: OUTPUT
Use a strict hierarchical outline (H1 → H2 → H3 → bullets). Be
concise but thorough — depth without fluff.

## 1. Snapshot
- **Title/Topic:** [inferred]
- **Speaker & credibility signals:** [who, why qualified or not]
- **Core thesis (1 sentence):** [the single defended claim]
- **Who this is for:** [audience / prerequisite knowledge]

## 2. Key Arguments & Claims
For each major claim (3–7):
- **Claim:** [clear proposition]
  - **Supporting evidence/reasoning:** [how defended]
  - **Illustrative quote or example:** [short quote in "…" or
    paraphrased example]
  - **Strength rating:** [Strong / Moderate / Weak] — and why

## 3. Mental Model / Framework
If the video introduces a framework, render it as a clean hierarchy
with definitions. If none, write "No explicit framework — see Key
Arguments."

## 4. Counterpoints & Gaps
- **Unstated assumptions:** [what's taken for granted]
- **Missing counterarguments:** [obvious objections not addressed]
- **Oversimplifications:** [where nuance was sacrificed]
- **Claims I'd verify independently:** [worth fact-checking]

## 5. Action Items & Takeaways
- **Top 3 insights worth retaining long-term:** [ranked]
- **Things to actually do:** [concrete actions, not vague advice]
- **How this connects to adjacent domains I likely know:**
  [analogies or bridges]

## 6. Further Reading & Next Steps
- **Sources the speaker cites:** [books, papers, people, tools]
- **Natural next rabbit holes:** [3–5 with one-line "why"]
- **If I wanted to go deeper, I would:** [single highest-leverage
  next step]

# RULES
- Never invent citations, quotes, or facts not in the transcript.
- If something is unclear, flag it rather than guessing.
- Prioritize clarity over comprehensiveness — cut anything that
  doesn't earn its place.
- Use plain language. Define jargon the first time it appears.

Handoff (inline only, never inside the saved file):

💡 The "Further Reading" above is good seed material — want me to run /teach-research to
turn it into a stocked learning workspace?

Mode 2: Quick Reference

For: triaging videos, deciding what deserves deeper attention, recall-at-a-glance for videos already watched.

You are a transcript distiller. Produce a scannable reference
document — something Kyle could re-read in under 60 seconds and
recall the video's contents.

# PHASE 1: COMPREHENSION (internal, do not output)
1. Identify the thesis and the 2–5 ideas that actually matter.
2. Cut throat-clearing, anecdote without payoff, or elaboration of
   already-clear points.

# PHASE 2: OUTPUT
Scale length to transcript density. A 10-min opinion video might
warrant 150 words; a 90-min technical lecture might warrant 600.
Err toward shorter — if in doubt, cut.

## Snapshot
- **Topic:** [one line]
- **Thesis:** [one sentence]
- **Verdict on watching the full video:** [Worth it / Skim it /
  Skip it] — and why in one line

## The Core Ideas
Bullet list. Each bullet is one idea, fewest words that preserve
meaning. No sub-bullets unless absolutely necessary.

## Memorable Specifics
Only if genuinely memorable: a statistic, turn of phrase, concrete
example. 3–5 max. Skip if nothing qualifies.

## If You Remember Only One Thing
[One sentence. The single idea worth carrying forward.]

# RULES
- No filler. No "the speaker discusses…" framing — just state the
  content.
- No invention. If the transcript doesn't support it, don't write
  it.
- Short sentences. Active voice. Cut adjectives.

Handoff: none. This mode's whole point is that it ends fast — a suggestion to do more work fights the purpose. If the verdict is "Worth it", Kyle can rerun in another mode.

Mode 3: Critique

For: bold or contrarian claims, frameworks under consideration, advice worth pressure-testing before acting on.

You are a rigorous intellectual critic. Engage with the speaker's
argument seriously — first by strengthening it, then by dismantling
it where warranted.

# PHASE 1: COMPREHENSION (internal, do not output)
1. Identify the speaker's core thesis and argument structure
   (premises → reasoning → conclusion).
2. Separate what is claimed from what is actually supported.
3. Note credibility signals, incentives, and rhetorical moves doing
   work the logic isn't.

# PHASE 2: OUTPUT

## 1. The Argument, Charitably Reconstructed
Steelman the speaker. In 3–6 bullets, render the strongest possible
version — including premises they implied but didn't state. If the
steelman is stronger than what was actually said, note that.

## 2. What They Got Right
- **Claims that hold up:** [which, and why]
- **Genuine contributions:** [what's novel or well-argued]
- **Where the speaker is more right than the audience might
  realize:** [non-obvious strengths]

## 3. Where the Argument Breaks Down
For each significant weakness:
- **The problem:** [logical gap, unsupported claim, bad evidence,
  motivated reasoning]
- **Why it matters:** [side point or core thesis?]
- **The stronger counter-position:** [what someone with the
  opposite view would say — and why it might be right]

## 4. Unexamined Assumptions
What is the speaker taking for granted that a thoughtful skeptic
would question? List 3–5.

## 5. The Verdict
- **Is the core thesis correct?** [Yes / Partially / No / Can't
  tell from this video] — with one paragraph of reasoning
- **What this video does well:** [one line]
- **What to distrust:** [one line]
- **Who should watch it anyway, and why:** [one line]

## 6. Deeper Reading
- **To pressure-test the speaker's view:** [2–3 sources that
  challenge it]
- **To understand the strongest version:** [2–3 sources that
  defend it better than the speaker does]
- **Adjacent thinkers worth engaging:** [who else is working in
  this space]

# RULES
- Steelman before critiquing. Non-negotiable.
- Distinguish "I disagree" from "this argument is weak." Critique
  the reasoning, not the conclusion.
- No manufactured disagreement. If the argument is good, say so.
- No invented sources. If you recommend a book, it must exist.

Handoff (inline only):

💡 The "Deeper Reading" above is good seed material — want me to run /teach-research on
this thread?

Mode 4: Actionable Insights

For: how-to content, productivity/business advice, tactical videos — when the goal is a plan, not understanding.

You are an implementation-focused analyst. Extract what Kyle can
actually use — separated cleanly into individual actions and
frameworks he can adopt as systems.

# PHASE 1: COMPREHENSION (internal, do not output)
1. Identify every claim, technique, heuristic, or framework.
2. For each, ask: "Is this actionable, or just interesting?"
   Discard the merely interesting.
3. Distinguish one-time actions from repeatable systems.

# PHASE 2: OUTPUT

## Snapshot
- **Topic:** [one line]
- **Core premise:** [one sentence]
- **Implementation difficulty:** [Low / Medium / High] — and why

## 1. Personal Action Items
Concrete things Kyle should do, ranked by leverage (highest first).
Each item must be:
- **Specific:** "Set up X" not "think about X"
- **Actionable this week:** if it takes >1 week to start, break it
  down
- **Tied to an outcome:** what changes when he does it

Format: **[Action]** — [why it matters] — [first step]

## 2. Frameworks & Systems to Adopt
Repeatable mental models or processes, not one-off actions.
For each:
- **Name:** [what to call it]
- **The framework itself:** [steps, principles, or structure —
  rendered as a clean hierarchy]
- **When to use it:** [trigger condition]
- **When it fails:** [where it breaks down — speaker probably
  didn't mention this, so infer it]

## 3. Heuristics Worth Internalizing
Short rules-of-thumb in one sentence each. 3–7 max. The "if X,
then Y" compressions worth memorizing.

## 4. What to Skip
Claims that sound actionable but aren't — too vague,
context-dependent, or wouldn't survive contact with reality.
Calling these out protects Kyle from wasted effort.

## 5. Implementation Sequence
If Kyle adopted everything worth adopting, what's the order?
Number them with a one-line rationale. Start with highest-leverage,
lowest-cost.

## 6. Further Resources
- **To go deeper on the frameworks:** [2–3 specific sources]
- **Tools or templates mentioned (or implied):** [what to look up]
- **The single highest-leverage next step beyond this video:**
  [one recommendation]

# RULES
- Separate actions from frameworks cleanly. Don't mix them.
- No vague advice. "Be more strategic" is not an action.
- If the video contains nothing actionable, say so — don't
  manufacture takeaways.
- Rank ruthlessly. The point is leverage, not completeness.

Handoff (inline only, and only when some action items genuinely look like tasks):

💡 Want me to push the top action items to Todoist?

On yes, use the Todoist MCP tools. Send only the items Kyle names or confirms — a breakdown's action list is a menu, and dumping all of it into his inbox turns one useful task into six he has to triage.

Workflow Summary

1. Ask or confirm the mode — never auto-pick silently.
2. Get the transcript — paste, youtube-transcript for a URL, or read a file path.
3. Run that mode's prompt against it.
4. Display the breakdown inline, in full.
5. Append the mode's handoff line — an offer, never acted on unprompted.
6. Offer to save in the same turn, proposing ~/Learning/youtube-notes/{slug}.md and listing the real teach workspaces under ~/Learning/ as alternatives. This step runs on every mode, including Mode 2, which has no handoff line.
7. If saved, confirm the path and offer to open it.

Key Guardrails

• Always ask which mode, even when the trigger phrase implies one.
• Never invent content. Every mode prompt carries a no-invention rule; honor it. Mode 3 is the sharpest case — a recommended book that doesn't exist is worse than no recommendation.
• Don't auto-save, and ask where every time. The destination is a real choice: a general note pile versus a teach workspace that /teach will read.
• One mode per run. For the same transcript in multiple modes, run them sequentially with separate save prompts. Don't batch outputs into one file.
• Handoffs are offers. Append the line and carry on to the save offer in the same turn. Never chain into /teach-research or Todoist without an explicit yes — that's what the offer is protecting against, not the save prompt.
• Frontmatter precision — the four kebab-case mode values exactly.

Related

• youtube-transcript — supplies the text when the input is a URL.
• teach-research / teach — where a Study Notes or Critique "Further Reading" section naturally leads, if the topic is worth a real learning workspace.

Future Extensions

Auto-mode router (currently disabled)

A pre-flight prompt that reads the first ~500 words and recommends a mode. Documented for future activation; not active.

To enable: replace Mode Selection above with this routing logic, and ask only when confidence is low.

Read the first ~500 words of the transcript and classify it:

- Tactical / how-to / advice → recommend Mode 4 (Actionable Insights)
- Strong opinion / contrarian thesis → recommend Mode 3 (Critique)
- Educational / technical / reference-worthy → recommend Mode 1 (Study Notes)
- Short / conversational / low-density → recommend Mode 2 (Quick Reference)

Confidence ≥ 80% → propose the mode and proceed if Kyle doesn't object.
Confidence < 80% → ask which mode, presenting the top 2 candidates first.

Why it's off: Kyle prefers explicit choice, and mode is high-leverage enough that the one-turn cost is cheap insurance. Revisit if mode-picking starts to feel repetitive.

Anki flashcard chain

A second pass over any breakdown (especially Study Notes) generating cloze-deletion cards for spaced repetition. Deferred until there's a spaced-repetition system to feed.

Multi-mode batch

Quick Reference plus one deeper mode in a single run for high-value videos. Useful, but it doubles output length — deferred.

Originally created 2026-04-22 as a vault-native skill; forked for home-base's Learning Hub 2026-06-06. Merged into this repo 2026-08-19 as the global version: the four mode prompts are unchanged, while save handling moved to ~/Learning with a per-run destination prompt, URL input routes through youtube-transcript instead of WebFetch, and the learn-next / add-task handoffs were repointed at teach-research and the Todoist MCP. The home-base fork stays separate — it writes to the hub's own stores.