ehteshambuildagents bl-mine
One good long asset holds four to six posts. Most people extract one, post the summary, and throw the rest away.
Input
A transcript, an article, a newsletter, a script, a call recap. If the user gives a URL and a transcript tool is available in the session, use it. Otherwise ask them to paste the text. Read the whole thing before extracting anything - the best post is rarely in the first third, because that is where the setup lives.
Extract, do not summarise
A summary of a video is not a post. Nobody wants the summary. Go through the asset and pull out the fragments that survive on their own:
| pull out | because |
|---|---|
| the claim that made you pause | it will make a reader pause |
| any number said out loud | numbers are the scarcest thing in a feed |
| the aside that was not the point | asides are where the honest material is |
| a disagreement with received wisdom | the highest-comment shape there is |
| a story with a turn in it | the highest-dwell shape there is |
| the question the interviewer asked twice | it was the interesting one |
| the thing the speaker almost did not say | usually the best post in the asset |
List what you found before writing anything, in one line each. The user knows their own material and will tell you which ones are actually worth it.
Then build each one as a standalone post
Every extracted post must work for a reader who has never seen the source. That means:
- No "as I said in my video". The post is the thing, not an advert for the thing.
- Each carries its own opener from
../bl-draft/openers.json. - Each uses a different opener pattern. Five posts from one asset that all open the same way read as batch-produced, because they were.
- No two posts make the same argument. If two extracts collapse into one point, that is one post, not two.
Spread them out
Five posts from one asset go out over two to three weeks, not five days. The same audience sees all of them, and a source mined visibly in one week reads as a content run rather than a week of thinking.
Hand the user a sequence with the gap between each, and note which one to lead with. The strongest post goes first, not the introduction.
Link back once, at most
If the asset is public, one post can point to it, and it should be the last one, not the first. Posts with external links are shown to fewer people, so spend that cost once, deliberately, on the post best placed to convert.
Before delivering
Run all of them through ../bl-scrub/scrub.py in one pass and score each. A
batch is where flat cadence hides, because you stop hearing it by the fourth
draft.
Every number in every post comes from the asset itself or from ledger.md.
Transcripts are full of approximations said out loud - "we maybe doubled it" is
not a number, and it becomes {{your number}}.