Method Β· Milestone Companion: Structure Beats Eloquence
What this milestone asks
Record a 10-minute video titled "How an agent works," publish it publicly, and collect five pieces of feedback. The stage behind it is about craft: structure beats eloquence; live demos follow iron rules; analogies are your best friend; every talk leaves exactly one takeaway. Ten minutes is deliberate β long enough to require real structure, short enough that structural failure is cheap to fix.
Why this milestone
Writing forgives; video does not. A reader who hits a confusing paragraph can reread it. A viewer who hits a confusing thirty seconds clicks away, and the retention curve tells you exactly where. The video milestone forces the two skills that separate watched evangelists from unwatched ones: narrative structure and delivery under constraints. Five pieces of genuine feedback β not likes, not "great video!" β are worth more than five hundred silent views, because they show where your mental model of the audience diverged from the actual audience.
The one-takeaway rule exists because audiences retain one idea per talk whether or not you planned it. If you do not choose the idea, the audience chooses for you β usually the demo, not the principle. For "How an agent works," a defensible single takeaway is: an agent is a model in a loop with tools, and the loop is where the design lives. Everything in the video either serves that sentence or gets cut.
How to use the resources
Study narrative structure before writing a script. Watch Andrej Karpathy's Intro to LLMs and Deep Dive into LLMs with a notebook. These are reputationally the benchmark for LLM technical storytelling; your job is not to enjoy them but to dissect them. Timestamp every transition: when does he state the problem, dismiss the naive approach, introduce the new idea, show rather than tell? Underneath you will find the golden structure: problem β wrong approaches β new idea β demo β call to action.
Study visual explanation. Watch 3Blue1Brown's attention animation the same way β once for content, once for the analogies, once purely for how visual pacing substitutes for verbal explanation. You will not have that animation budget, but you can steal the discipline: one visual idea on screen at a time, each earning its presence.
Read the two craft books. Chris Anderson's TED Talks: The Official TED Guide to Public Speaking gives you the through-line doctrine β the single idea everything hangs on. Steven Pinker's The Sense of Style cleans your script of jargon and abstraction. Read Anderson before outlining; read Pinker between the first and second script drafts.
Mine your Stage 1 essay for analogies. The analogies that survived your layperson retelling are proven assets. The canonical set β agent as new hire, context as working memory, RAG as an open-book exam β works because each maps a familiar constraint onto an unfamiliar one. Use at most three analogies in ten minutes; more is a zoo, fewer forces depth.
Script, rehearse, record β in that order. Write the full script word for word, then record three complete rehearsals. The live-demo iron rules apply even to recorded video: always have a recording of any demo as backup, and never show on camera a run that has not succeeded three times before. Publish on your chosen platform β Bilibili or Zhihu for a Chinese-language audience, YouTube otherwise β then actively solicit the five feedbacks; they rarely arrive uninvited.
Study stage control even though you are on camera. David Malan's CS50 lectures are the benchmark for pacing and audience control in technical teaching β watch his use of silence, repetition, and movement. These skills are dormant in video but detonate productively at Stage 4.
Common failure modes
- Feature tour instead of a through-line. Planning, memory, tools, multi-agent, and evaluation in ten minutes means the viewer retains nothing. Cut until one idea remains.
- Demo-first structure. Opening with the impressive demo before the viewer has a problem. A demo without a problem is a magic trick; with a problem, it is an argument.
- Analogy sprawl. Five half-explained analogies in ten minutes. The viewer remembers the metaphors, not the mechanism.
- Improvising the demo. A live run that has not succeeded three times off-camera first. A recorded failure is not charming authenticity; it is a structural error you chose to keep.
- Counting likes as feedback. "Great video!" contains zero information. The five feedbacks must be specific: what confused, what surprised, what they would retell.
Acceptance checklist
- A 10-minute (Β±2 minutes) video titled "How an agent works" (or close variant) exists and is publicly accessible.
- The video follows a discernible problem β wrong approaches β new idea β demo β call-to-action structure.
- A viewer could state the single takeaway in one sentence; you have verified this by asking.
- Every demo shown had succeeded at least three times before recording, or a pre-recorded backup was used.
- No more than three analogies are used, and each is fully explained.
- Five specific pieces of feedback are collected in writing, with names or handles attached.
- At least one revision decision (made or consciously deferred) is recorded per feedback item.
Going further
The five feedbacks are your first FAQ entries β in Stage 3 they seed your topic pool and weekly queue. If the video performed decently, resist making another one immediately; Stage 3 is about cadence, and a video made on schedule teaches more than one made in excitement. If it performed poorly, the gap between your script and the audience's retention curve is the most honest measurement of your current craft. In Stage 4, this video becomes your audition tape for conference talks.