AI Skill Report Card

Extracting Methodology From Video

A-85·Oct 29, 2026·Source: Web
14 / 15

Given raw transcript text, extract the methodology in 4 passes: (1) identify the domain expertise and core claim, (2) pull out the repeatable steps, (3) pull out concrete examples/numbers/phrases used verbatim, (4) package into skill format.

Input transcript snippet:
"...so what I do every time I write a subject line is I write 10 versions,
then I kill any that are over 50 characters, then I read the remaining
ones out loud and pick the one that sounds like a text message from
a friend, not an ad. Last week I did this for a client and open rate
went from 12% to 34%..."

Extraction:
- Domain: email subject line writing
- Step 1: Write 10 variants
- Step 2: Cut any over 50 characters
- Step 3: Read aloud, pick the one that sounds like a text from a friend
- Concrete example: 12% → 34% open rate improvement
- Verbatim phrase to preserve: "sounds like a text message from a friend, not an ad"
Recommendation▾
Add an example showing a full assembled skill file output (frontmatter + sections) rather than only fragments, to demonstrate the final packaging step end-to-end
14 / 15

Progress:

  • Step 1: Confirm input type — full transcript, timestamped notes, or summary (if only a URL/title is given, ask for transcript text; do not guess content)
  • Step 2: Skim for the core claim — what specific outcome does the speaker achieve, and for whom?
  • Step 3: Highlight every imperative sentence ("I always...", "first you...", "the key is to...") — these are workflow steps
  • Step 4: Highlight every number, name, tool, or before/after result mentioned — these become examples
  • Step 5: Highlight verbatim phrases the speaker repeats or emphasizes — reuse their exact language in the skill (it signals real expertise, not generic advice)
  • Step 6: Order the steps chronologically into a workflow; merge duplicates; drop tangents/filler/stories that don't add process
  • Step 7: Write 2+ Example blocks using the actual numbers/scenarios from the transcript (never invent placeholder examples if real ones exist in the source)
  • Step 8: Draft YAML frontmatter: gerund name matching the domain, third-person description with trigger phrase
  • Step 9: Add Best Practices (things speaker says to always do) and Common Pitfalls (things speaker warns against or says "most people get wrong")
  • Step 10: Re-read transcript once more specifically hunting for anything contradicting your draft — revise
Recommendation▾
Clarify how to handle very short/sparse transcripts that lack concrete numbers or verbatim phrases — what's the fallback methodology there
17 / 20

Example 1: Input (transcript excerpt):

"People mess up cold outreach because they open with 'Hope you're doing well.' Nobody cares. Instead, open with a specific observation about their work — something you had to actually look at their LinkedIn or website to know. Then one sentence on why you're reaching out. Then a question, not a pitch. I send maybe 40 a week and get 8-10 replies doing this."

Output (extracted skill fragment):

Markdown
undefined
Recommendation▾
Consider adding a brief note on how to name the skill file / where to save it, since Step 8 mentions frontmatter but not file placement conventions
14 / 15

Progress:

  • Step 1: Research a specific, verifiable detail about the recipient (LinkedIn/website)
  • Step 2: Open with that observation, not a greeting
  • Step 3: State reason for reaching out in one sentence
  • Step 4: End with a question, never a pitch
17 / 20

Example 1: Input: Cold outreach to a prospect Output: 40 emails/week sent → 8-10 replies (20-25% reply rate)

  • Do not open with "Hope you're doing well" — generic greetings get ignored
  • Do not end with a pitch — end with a question

**Example 2:**
Input: Only a YouTube URL, no transcript, no notes, filler text as "example"
Output: Do not fabricate content. Respond with:
> "I can't access video directly. Paste the transcript (YouTube auto-captions work — click '...' under the video → 'Show transcript' → copy/paste) or detailed notes, and I'll extract the methodology into a skill."
  • Prefer the speaker's exact wording for key phrases over paraphrasing — it preserves the distinctiveness of their method
  • If the transcript has numbers/results, those ARE your examples — don't invent hypothetical ones
  • If the transcript is long (>20 min talk), work section by section rather than trying to hold the whole thing in one pass
  • When speaker gives a list ("three things I always do"), map it directly to the Workflow checklist in order
  • If no transcript is available, say so plainly in one sentence and ask for it — don't produce a multi-paragraph apology
  • Do not claim to have watched/accessed a video — only work from text actually provided
  • Do not pad output with meta-commentary about limitations instead of doing the extraction on available text
  • Do not invent generic examples when the source material contains real ones — always prefer source-grounded specifics
  • Do not treat a bare URL as a transcript substitute — a URL alone contains zero extractable content
0
Grade A-AI Skill Framework
Scorecard
Criteria Breakdown
Quick Start
14/15
Workflow
14/15
Examples
17/20
Completeness
17/20
Format
15/15
Conciseness
13/15