Transcript to Article

Take a raw recording transcript and produce five things: a clean readable transcript, a finished article, a distribution package, social assets, and a scan report proving the output is safe to publish.

Before you write anything

Read these first. They are the source of truth and this skill does not invent around them.

1. brand-kit/brand.md: name, palette, fonts, CTAs
2. brand-kit/voice.md: house style and banned constructions
3. brand-kit/positioning.md: the frame, pillars, traps, stats, vocabulary
4. brand-kit/capabilities.md: read before making any product claim

If brand-kit/ is missing or still has its template placeholders, stop and tell the user to run brand-kit-setup. Do not proceed with invented positioning.

Also read docs/aeo-style-guide.md. It owns the answer-engine rules this skill depends on: the opening-answer rule, atomic answers under question headings, the key-takeaways block, comparison tables, and the FAQ section. This skill does not restate those rules and does not override them.

If house-style is installed, run it as the final voice pass over every asset before delivery. Do not hand-copy voice rules into this skill.

Speaker Roster

A transcript is worthless until you know who is talking and what each person is allowed to be an authority on. Establish the roster before Stage 1.

Get it from one of three places, in this order:

1. The user supplies it inline. Ask for it if the transcript labels are initials, "Speaker 1", or absent.
2. The brand kit. If brand-kit/positioning.md has entries under Internal Quotes, those names and titles are your roster for internal voices. Respect the Cleared for external use column: an uncleared speaker can inform the argument but does not get quoted by name.
3. The transcript itself, when speakers introduce each other on the recording.

Record the roster in this shape before you clean anything:

Rules the roster enforces:

• Attribute an insight only to a speaker whose Authority on column covers it. A finance lead does not get quoted on infrastructure because they mentioned it.
• Never invent a name, a title, or a company for an unlabeled speaker. Use "Speaker 2" and tell the user you need the identity.
• Guests from outside the company are external voices. Quote them verbatim and do not put your positioning language in their mouth.
• If the roster is empty and the user cannot supply one, write the article from the ideas only, with no attributed quotes, and say so in the delivery note.

Workflow

Stage 1: Clean the transcript

Produce a readable transcript that a human could publish as-is.

1. Cut fillers. Remove "um", "uh", "like", "you know", "actually", "basically", "right", "so" (as a sentence opener), "I mean", "kind of", "sort of", "obviously", "literally", and repeated false starts.
2. Standardize speaker labels. One bolded label per speaker, spelled the same way every time, matching the Speaker Roster exactly. Merge split turns from the same speaker into one block. Drop timestamps unless the user wants them kept for the video description in Stage 3, in which case park them in a separate list.
3. Repair sentences. Fix grammar and finish incomplete thoughts without changing what the speaker meant. When a sentence is unrecoverable, cut it rather than guessing.
4. Strip crosstalk and logistics. Scheduling, audio checks, and side conversations do not survive the clean.
5. Flag the raw material. Mark every quote, statistic, customer story, and sharp phrasing worth carrying into the article. List them at the end of the stage so Stage 2 has a shortlist rather than a re-read.
6. Verify every number. A statistic that appears in the transcript but not in Headline Stats in brand-kit/positioning.md is an unverified claim. Keep it attributed to the speaker who said it, or drop it. Never promote it to a house statistic.

Apply brand-kit/voice.md Grammar Constraints to the cleaned transcript too. It is a published asset, not scratch work.

Stage 2: Write the article

Default target is the article length in Length targets in docs/writing-templates.md, which is the only place asset lengths are set. For a pillar guide built from a long or multi-episode recording, take the pillar figure from the same table and follow the Pillar guide template in docs/aeo-style-guide.md.

Structure, in order:

1. Hook. A blunt claim, a real number, or a direct question. No setup, no throat-clearing, no "in this article we will".
2. Opening answer. Answer the article's core question inside the opening, at the length docs/aeo-style-guide.md prescribes. This is what an answer engine lifts.
3. Key takeaways block. Per docs/aeo-style-guide.md.
4. The stakes. Why the current approach fails, with the specific cost. Use numbers the transcript supplies. Where the failure has a name in Named Traps in brand-kit/positioning.md, use that name.
5. The mechanism. How the better approach actually works. Explain it well enough that a reader could argue for it in their own meeting.
6. The proof. Examples, scenarios, quotes, and numbers from the transcript, attributed per the Speaker Roster.
7. The close. The exact CTA from Calls to Action in brand-kit/brand.md for the blog and article context.

Do not use the contrast formula. "Most companies do X. Winners do Y." and "It's not about X, it's about Y." are listed under AI Tells to Avoid in brand-kit/voice.md and lose points in the quality panel. When you need to show a gap between approaches, use a comparison table per docs/aeo-style-guide.md, or state the better approach directly and let the contrast be implicit.

Headings. Question-based H2s that match how a buyer would actually phrase the search, each opening with a standalone atomic answer per docs/aeo-style-guide.md.

Positioning discipline.

• The qualities in The Pillars in brand-kit/positioning.md are invisible. The reader absorbs them. Never write the framework's name, "our pillars", or any internal label into the article.
• Every product claim traces to What We Do in brand-kit/capabilities.md. Anything in What We Do Not Do is off limits regardless of what a speaker said on the recording.
• Use Canonical Vocabulary in brand-kit/positioning.md for terminology and check Superseded Names in brand-kit/voice.md before writing any product name.
• Sell the idea, not the product. The company should read as the natural conclusion of a sound argument, not the subject of it.
• Link internally using Link Map in brand-kit/capabilities.md. Make a descriptive phrase the anchor text. Never write a bare "Learn more: URL".

Stage 3: Distribution package

Produce all four, exactly to spec.

Three title variations. Label which is which.

1. SEO. Keyword-led, matches real search intent.
2. Provocative. Challenges the reader's current assumption.
3. Direct. States the core argument plainly.

Meta description. 155 characters maximum. Count them and state the count. Action-oriented. Answers "why should I read this".

Video description. Roughly 200 words. Include timestamps for the key topics (use the timestamps parked in Stage 1). Include the CTA link from Calls to Action in brand-kit/brand.md. End with 3 to 5 tags.

Show notes. A 3 to 4 sentence summary, then 3 to 5 "What you'll learn" bullets. Each bullet is a specific takeaway, not a topic label.

Stage 4: Social and visual assets

If social-generator is installed, hand it the finished article and the chosen title and let it produce this stage. It owns the per-platform specs, the text overlay, and the image prompts. Do not duplicate its rules here.

If it is not installed, produce the minimum set:

• LinkedIn post. Lead with a real number or a real claim from the article. Beats and length come from LinkedIn post in docs/writing-templates.md. Article link at the end. Hashtags come from Hashtags in brand-kit/brand.md, three to four of them, never invented here. If that section is empty, ship the post without hashtags and say so.
• Short-form post. Under 280 characters. One sharp idea. Link.
• Thread. Five posts: the problem, the cost, the mechanism, the proof, the CTA. Each under 280 characters.
• Image prompt. Built on Illustration Style in brand-kit/brand.md, visualizing the article's central tension. Colors come from Palette in the same file. If Illustration Style is missing, say so and skip the prompt rather than inventing a look.

Stage 5: Pre-delivery scan

Run this as a dedicated pass over every asset, including the cleaned transcript. Report the result as a short checklist with a pass or fail per line. Fix and re-scan before delivering. Do not deliver with a known failure.

Output Format

Deliver in this order. It matches references/article-template.md exactly, so the two never disagree.

## Speaker Roster
[the table, confirmed with the user]

## Stage 1: Cleaned Transcript
[cleaned transcript, then the flagged quotes and stats]

## Stage 2: Article
[article in markdown, ready for the CMS]

## Stage 3: Distribution Package
### Titles
### Meta Description
### Video Description
### Show Notes

## Stage 4: Social and Visual Assets
### LinkedIn Post
### Short-Form Post
### Thread
### Image Prompt

## Stage 5: Pre-Delivery Scan
[the checklist, pass or fail per line]

Reference Files

references/article-template.md is the blank skeleton for the whole delivery. Its section order is the Output Format order above. If you change one, change both.

Notes

• A real number from the recording beats a generic claim every time. The flagged list from Stage 1 is the most valuable output of that stage.
• The atomic answer under each H2 is what an answer engine quotes. Write it first, then write the expansion around it.
• Long recordings are pillar guides, not long articles. Switch templates rather than padding.
• When the transcript contradicts the brand kit, the brand kit wins and you tell the user which claim you dropped and why.