One conversation becomes a week of posts

An hour of conversation contains more usable material than a week of writing, and almost none of it gets used. This loop reads a transcript, pulls out the claims that stand on their own, and drafts one post per claim with the speaker’s own words attached.

Use it when you record anything: podcasts, webinars, conference talks, customer calls you have permission to use. One recording should not produce one post.

You need a transcript with speaker labels. Transcription happens outside this loop.

You get one draft per standalone claim, each quoting the speaker verbatim, plus a discard list with a reason per discarded claim so you can check what it chose to skip.

The gate is you. It drafts and opens a pull request. It never publishes.

What counts as a standalone claim

The filter is the whole loop. A claim qualifies when it is:

• Assertable without the conversation around it. If understanding it requires the previous three exchanges, it is context, not a claim.
• Specific enough to be wrong. “We care about quality” cannot be wrong. “We rejected forty percent of drafts last quarter” can be.
• The speaker’s own. Not their summary of someone else’s position, which becomes a misattribution the moment it leaves the transcript.

Everything else is discarded with a reason. Expect most of a transcript to be discarded: an hour of talk usually yields somewhere between three and eight usable claims, and a loop that finds thirty is matching too loosely.

Goal prompt

Read <transcript-file>. Identify every claim <speaker> makes that stands on
its own, is specific enough to be falsifiable, and is the speaker's own
position rather than their description of someone else's.

Discard every other passage, and record each discard with the reason.

For each surviving claim, draft one post that:
- leads with the claim, not with context about the conversation
- quotes the speaker verbatim, with the quote marked as a quote
- names one thing a reader could do differently
- does not assert anything the speaker did not say

Write every draft to <drafts-directory> and the discard list to
<discard-file>. Open a pull request. Never publish, never send, never schedule.

Verifier prompt

Read every draft against <transcript-file>.

FAIL if any quote is not verbatim in the transcript.
FAIL if any draft asserts something the speaker did not say.
FAIL if any draft attributes to <speaker> a position they were describing
rather than holding.
FAIL if any claim in the discard list is missing a reason.
FAIL if any draft leads with context about the conversation rather than the
claim.
FAIL if anything was published, sent or scheduled.

Otherwise PASS. Print the verdict and, on FAIL, the draft and the line in the
transcript it failed against.

Stopping condition

Every standalone claim in the transcript is either drafted with a verbatim quote or discarded with a reason, one pull request is open, and the verifier printed PASS.

What it may and may not do

May: read the transcript, draft posts, write a discard list, open a pull request.

May not: publish, send, schedule, paraphrase a quote, or attribute a position the speaker was describing rather than holding.

The failure worth watching for

The dangerous output is a draft where the quote is accurate and the framing around it is not. A speaker explains an objection they disagree with, and the draft presents the objection as their view. The quote passes a verbatim check and the post is still wrong.

That is why the verifier checks attribution separately from quotation, and why a human reads the drafts. If you only check one thing before publishing, check that the speaker holds the position the post says they hold.

Before you use a customer call

Get permission in writing, and get it for the specific use. Permission to record is not permission to publish, and a customer who agreed to a research call did not agree to be quoted in your marketing. This is the loop where the legal exposure is real.