Objective
What you're building
The Voice Profile: 3-5 real examples of your own writing that become a reusable few-shot fixture every one of the 6 generators includes in its prompt. No pipeline code yet — this is the one piece of the whole course that has to come from you specifically, not from a library.
Sketch the pipeline before writing any code
seed input (URL / YouTube link / raw text)
│
▼
preprocessor ──▶ clean text
│
▼
┌──────┴──────┬─────────┬──────────┬──────────┬──────────┐
blog twitter linkedin newsletter yt-script email
generator generator generator generator generator generator
└──────┬──────┴─────────┴──────────┴──────────┴──────────┘
│ (all 6 run in parallel — Lesson 4)
▼
orchestrator ──▶ aggregate results
│
▼
output (Notion / local files — Lesson 5)
Every box other than "Voice Profile" is either a preprocessing step (Lesson 2), a generator (Lesson 3), or plumbing (Lessons 4-5). The Voice Profile isn't a box in this diagram — it's an input threaded into all 6 generator boxes at once, which is exactly why getting it right once matters six times over.
Why voice, not generation quality, is the hard problem
Six technically-correct outputs can still all fail the same way
Ask Claude to "write a LinkedIn post about X" with no other context and you'll get something grammatically perfect, on-topic, and completely generic — the same voice every AI content tool produces by default. Do this across all 6 formats and you get 6 pieces of content that are individually fine and collectively worthless, because nobody can tell they came from you. This is the actual failure mode of most "content factory" tools on the market: they solve generation, not voice.
The voice profile: a few-shot fixture, not a personality description
The fix isn't asking the model to "write in a casual, direct tone" (adjectives don't transfer voice reliably) — it's showing it 3-5 real examples of how you actually write, which every generator prompt includes verbatim.
VOICE_PROFILE = """
Examples of my actual writing style:
1. "Shipped the auth refactor today. Took 3x longer than estimated because
I kept finding edge cases in the token refresh logic. Worth it though —
the old version was silently failing for maybe 2% of users."
2. "Unpopular opinion: most 'AI strategy' decks are just a slide with GPT-4
costs and a vague roadmap. The teams actually shipping AI features
don't have a strategy deck, they have a Tuesday they tried something."
3. [your 3rd real example]
""".strip()
Notice these examples aren't topically related to what the pipeline will eventually generate — that's the point. You're not giving the model example outputs to imitate line-by-line, you're giving it enough real samples of your sentence rhythm, word choice, and what you tend to editorialize about that it can extrapolate voice onto a new topic. Two examples is thin; five is usually enough; twenty is unnecessary and just adds token cost to every generation call.
You write a Voice Profile using 4 examples, all of which happen to be short, punchy one-liners. You then ask the pipeline to generate a long-form blog post about a technical topic. The output feels stilted and unnaturally short-sentence-heavy for a blog post. Why?
A Voice Profile transfers voice, not format-appropriate structure — those are different things. If all your examples are short social posts, the model has real signal about your word choice and tone but no signal about how you'd naturally write a longer piece. The fix (which you'll feel the need for by Lesson 3) is including at least one longer example if you want the blog-post generator to sound like you specifically at that length, not just "you, but compressed."
What you're building today
voice_profile.md: your own 3-5 real writing examples (pull from actual emails, posts, or notes you've written — not written fresh for this exercise, which defeats the purpose), plus the pipeline diagram above sketched in your own words before Lesson 2's code starts.