build.allanninal.dev

Part 3 of 7 · Content repurposer series ~5 min read

How the strongest points get pulled

A long piece has a lot of words and only a few worth posting. The job here is to find those few. The drafter splits the source into passages, indexes them so it can match ideas back to exact text, and asks Bedrock to score and pick the handful that would actually make good short posts. The important rule runs through the whole step: every point that comes out the other side carries the exact passage it came from. If a point can’t be tied to the text, it doesn’t come out.

Key takeaways

  • The source is split into passages — small chunks of a few sentences each.
  • Each passage is turned into a number-fingerprint (an embedding) and stored so it can be matched and quoted.
  • Bedrock scores each candidate for how postable it is and picks the top few.
  • A grounding check confirms each chosen point is backed by a real passage; ungrounded ones are dropped.
  • Out comes a short list of points, each tied to the exact passage it came from.

The selection flow, per passage

Selection flow per passage when a piece is loaded A vertical flow diagram. At the top, an input box "Passage from source" with the chunk's text, its position in the piece, and the kind of source. Below that, a step "Embed the passage" — Titan Text Embeddings V2 turns the passage into a number-fingerprint stored in S3 Vectors. Below that, a check "Postable on its own?" — if no (it's filler, a transition, or boilerplate), route to "Dropped" (skip it). If yes, continue. The next step "Score for strength" — Bedrock Haiku reads the passage and rates how strong a short post it would make: a sharp claim, a surprising number, a clean story beat all score high. The next step "Pick the top few across the piece" — keeps only the highest-scoring passages so a long piece yields a handful of points, not a hundred. If a passage isn't in the top set, route to "Dropped." The next step "Grounding check" — confirms the point can be tied back to its exact passage; if it can't, route to "Dropped." If it can, the passage becomes a kept point, sorted into one of three buckets by score: Top point, Strong point, or Backup point. Each kept point keeps the exact passage it came from. A note at the bottom: every point carries its source passage — a point with no passage never leaves this step. Passage from source text · position · kind Step 1 Embed the passage Titan V2 → S3 Vectors Step 2 Postable on its own? filler → dropped Step 3 Score for strength Haiku rates how postable Step 4 In the top few? not top → dropped top score → top point Step 5 Tied to its passage? no passage → dropped Dropped skip it Top point highest score Strong point solid, worth a post Backup point held in reserve filler not top low end top mid score Every point carries its source passage — a point with no passage never leaves this step.
Fig 3. The selection flow, per passage. Five steps decide whether a passage becomes a point and how strong it is. The grounding check is the gate; a point that can’t be tied to its passage is dropped before it reaches the drafter.

Splitting into passages: small enough to quote

The first thing the drafter does with a loaded piece is cut it into passages — chunks of a few sentences each, broken at natural points like paragraph breaks or a change of subject. The reason for small chunks is the grounding rule: a short post should be able to point at the exact sentence it came from, and you can only do that if the chunks are small enough to be a real quote. A whole 1,800-word post is too big to be the “source” of a single line; a three-sentence passage is just right.

Each passage gets a position (where in the piece it is) and a note about the kind of source it came from. A passage from the opening of a post is treated a little differently from one buried in the middle — openings are often summaries, which make weaker standalone posts.

The fingerprint that lets the system match and quote

Each passage is turned into an embedding — a list of numbers that captures what the passage is about, using Amazon Titan Text Embeddings V2. Think of it as a fingerprint: two passages about the same idea have similar fingerprints, even if they use different words. The fingerprints are stored in Amazon S3 Vectors, a cheap place to keep them and search them.

This matters later, in the drafting step, for the grounding check. When a draft says something, the system can take that draft’s fingerprint and find the passage that best matches it — that’s the source it gets attached to. The fingerprints are what make “show me the exact passage this came from” a one-step lookup instead of a guess.

Scoring and picking the few worth posting

Now the model earns its place. Bedrock Haiku reads each candidate passage and scores it for one thing: how good a short post would it make on its own? A sharp claim scores high. A surprising number scores high. A clean before-and-after, a strong opinion, a concrete example — all high. A transition sentence, a throat-clearing intro, a “as we discussed earlier” — all low. The prompt is short and specific: “Rate this passage from 1 to 5 for how strong a standalone short post it would make. Return the score and one line of why. Do not rewrite it yet.”

Then the system keeps only the top few across the whole piece. A long post might have thirty passages; you don’t want thirty short posts, you want the best six or eight. The rules doc sets how many to keep. The highest scorers become top points, the next tier strong points, and a few more are held as backup points in case you skip several at review time and want more. Everything below the line is dropped — not deleted from the source, just not turned into a draft.

The grounding gate

Before any point moves on to drafting, one last check: can this point be tied back to its exact passage? For points pulled straight from the text this is automatic — the passage is the source. The check matters more in the next step, when the model rewrites a point into a post and might drift away from what the source actually said. Setting up the tie here, at selection time, is what makes that later check possible. A point with no passage behind it never becomes a draft.

This is the same instinct as a careful writer who won’t put a number in a post unless they can point to where it came from. The system makes that instinct a hard rule instead of a habit you’re trusting someone to remember.

Next post: how each pulled point becomes platform-sized drafts in your voice — the thread, the short posts, the quote-card caption — and how the grounding tie is checked one more time before a draft reaches you.

All posts