| name | talk-takeaways |
| description | Read a talk transcript (ideally plus the SRT for accurate timestamps) and produce a key_takeaways.md — a flat list of timestamp-anchored emoji-prefixed bullets for jumping back into the video. Use when the user wants a navigation index, not a read-through summary. |
| version | 0.1.1 |
| last_reviewed | 2026-05-26T00:00:00.000Z |
Talk Takeaways — Sub Skill
Read transcript.txt (and captions.SRT for precise timestamps) and
write key_takeaways.md. This is the re-navigation entry point — the
doc someone opens when they've watched the talk and want to find a
specific moment, or when they want to skim the structure before deciding
which sections to watch.
(This format / file name was first used in
yegor-pm-skills for a
Yegor Bugayenko talk; the structure is what's inherited, not anything
attributable to Yegor.)
When to invoke
talk-distill runs this after talk-transcribe (often in parallel
with talk-summarize).
- User asks for "key takeaways" / "a timestamp index" / "jump points" /
"the highlights with timestamps" / etc.
Required structure
# Key Takeaways — <Talk title> (<Speaker>)
**Video:** <one-line context>
**Link:** https://www.youtube.com/watch?v=<id>
**Duration:** <h:mm:ss>
## Key takeaways for quick navigation
> Timestamps are approximate (deduced from the SRT). Append `&t=XXmYYs`
> to the YouTube link to jump.
- **00:00** <emoji> Preamble — <whatever the meta-setup is>. Skip to ~01:00 for content.
- **01:00** <emoji> <One-line gist of the next topic>
- **05:00** <emoji> <Next topic>
[…20-50 bullets total, depending on talk length…]
Bullet style
Each bullet has three parts:
-
Timestamp in **HH:MM** or **MM:SS** format, depending on talk
length. Use bold so it stands out.
-
Emoji glyph that hints at the topic category (no strict mapping —
pick what feels right). Common patterns:
- 🎯 for "key insight" / "thesis statement"
- 🧠 for definitions / mental models
- 🛠️ for tools / mechanisms / how-it-works
- 🏗️ for architecture / design decisions
- 💬 for Q&A / audience interaction
- ❓ for open questions / future work
- ⚠️ for caveats / warnings / common mistakes
- 📚 for historical context / references to other work
- 🔌 for "this connects to / is built on…"
- 📐 for formal/spec content
- 🐛 for war stories / things that went wrong
- 🎬 for "this is where the demo / live coding happens"
- 🏁 for closing / wrap-up
-
One-line gist. Specific. Should make sense without surrounding
context. Compare:
Weak: "Discussion of state machines."
Strong: "Hierarchical states + parallel regions: how Tony's
traffic-light example has 4 parallel branches alive at once."
Picking bullets
Aim for one bullet per major topic shift, not per paragraph. The
transcript will have hundreds of paragraphs; a good takeaways doc has
20-50 bullets.
Topic shifts to bullet-mark:
- Speaker introduces a new section ("Okay, so now I want to talk about X").
- A demo / code-walkthrough begins.
- A war story is told.
- A Q&A exchange surfaces a non-obvious point.
- The speaker flags a "this is the BIG idea" moment.
- The talk closes / Q&A begins.
Skip:
- Filler ("uh, yeah, where was I").
- Pure tech-check chatter.
- Restatements of previously-covered material.
How to assign timestamps
Two ways, in order of preference:
-
From the SRT directly. For each bullet, find a distinctive
sentence from the corresponding transcript paragraph and grep -n it
in captions.SRT. The timecode line just above the matching content
line gives you the precise timestamp.
grep -n "as most of you already kind of know" captions.SRT
-
Approximate from paragraph index. If only the transcript is
available: (paragraph_index / total_paragraphs) * total_duration
gives a rough timestamp. Round to a sensible 5-minute granularity
to make it clear it's approximate.
The current srt_to_transcript.py script does NOT preserve inline
timestamps in transcript.txt. A future version of talk-transcribe
might emit a transcript_timestamped.txt to make this easier. Until
then, the SRT-grep approach is the precision route.
Output
${SUBDIR}/key_takeaways.md
Examples
Two reference takeaways docs that this skill should produce something
comparable to — written by hand with the same structure before the
skill existed:
The sister repo where this file shape originated (different emoji
choices, same structure):
Caveats
- Timestamps are jumpable, not exact. Even from the SRT, the block
boundary may be a few seconds off from where the topic actually starts.
Good enough for navigation; not for legal citation.
- Emoji is decoration. Don't agonize. If unsure, pick one that's
roughly topical and move on. The information is in the text, not
the glyph.
- The bullet count is a quality indicator. Fewer than ~15 bullets
for a 1-hour talk usually means under-segmentation (you're collapsing
topics). More than ~80 means over-segmentation (you're per-paragraph
bulleting). Re-balance.