AI Guides › Playbooks
By Nigel Guy · 7 min read
You have a bookmarks folder, a "Watch later" list or a camera-roll album full of AI tutorials, and you have used none of them. Saving feels like progress because the video looked useful for ten minutes, but a saved video is a promise to your future self with no date on it. The pile grows, the guilt grows, and the one technique you needed last Tuesday was in there somewhere.
The rule: a saved video is only worth keeping once it has become something you can run, and the video itself is then disposable.
The fix is a four-word loop that you apply to one video at a time: Extract, Run, Score, Delete.
A note on provenance: the working brief for this guide referred to a "four-word fix" without giving the words, and I could not verify what any original source meant by it. These four are this guide's own. The mechanism does not depend on anyone else's wording.
| Word | What you do | Output |
|---|---|---|
| Extract | Turn the video into text: the steps, the exact prompts and settings, the tools named | One short written recipe |
| Run | Do the recipe once, today, on a real task of yours | A result you can look at |
| Score | Judge the result against a bar you set before running | Keep, fix or bin |
| Delete | Remove the video from the pile and keep only the recipe (or nothing) | A smaller pile |
The Delete step is what makes this permanent. Most "organise your saves" systems add folders and never subtract anything.
Keep a single document called the Conversion Ledger. One row per video. Columns:
| Video (title, link) | The one claim it makes | Recipe (3-7 steps) | Bar set before running | Result | Verdict |
|---|
"The one claim" is the discipline. Most tutorials make one real claim ("you can turn a messy meeting transcript into a clean action list in one pass") wrapped in ten minutes of setup and personality. If you cannot state the claim in a sentence, the video is entertainment, and it goes straight to Delete.
Paste the transcript (or your notes) into your assistant and use the prompt below. It forces a recipe rather than a summary, because a summary is the thing you already failed to use.
You are a careful technical editor who turns tutorial transcripts into runnable recipes.
Context: I saved a video tutorial and never used it. I want one recipe I can run today.
Inputs:
- Video title: [VIDEO_TITLE]
- Transcript or my notes: [TRANSCRIPT_OR_NOTES]
- The task of mine this might help with: [MY_REAL_TASK]
- Tools and plan I actually have: [MY_TOOLS_AND_PLAN]
Do this, in order:
1. State the single main claim the video makes, in one sentence.
2. List every prompt, setting, menu path or command the speaker gave, quoted as spoken. If the speaker described something without giving it exactly, label it "NOT STATED" rather than inventing it.
3. Write a recipe of 3 to 7 numbered steps I can run on [MY_REAL_TASK], using only tools I listed. Flag any step that needs a tool or plan I did not list.
4. Write one pass/fail bar: what would a good result look like, in a form I can check in two minutes?
5. List anything in the video that sounds like a claim you cannot verify from the transcript alone.
Rules:
- If any input is missing or unclear, ask me before answering.
- Do not add steps the speaker did not mention unless you mark them "ADDED BY ME".
- Do not pad with a summary.
Before you answer, check that every quoted prompt or setting really appears in the transcript.
Fill in the five bracketed items. Paste the recipe and the bar into the Ledger straight away, before you run anything. Setting the bar first stops you from grading generously afterwards.
Run the recipe once, on real work, today. Not a demo task. Not "I'll try it this weekend". If the recipe fails at a step, note the step number in the Ledger and stop; a recipe that cannot run is a finding, not a failure on your part.
Compare the result with the bar you wrote. Three verdicts only:
If you want help scoring, this prompt keeps the assistant honest:
You are a sceptical reviewer. Compare my result against the bar I set before running.
Bar I set: [BAR]
Recipe I followed: [RECIPE]
Result I got: [RESULT]
Give me: (1) pass or fail against the bar, with the specific evidence; (2) if it failed, whether the cause is the recipe, my input or the tool; (3) the single smallest change worth trying once.
Rules: do not soften the verdict; do not suggest a second change; ask me for anything missing instead of guessing.
Before answering, check that your verdict refers to the bar as written, not a different one.
Remove the video from the pile, whatever the verdict. A Keep lives on as a recipe. A Bin was a useful lesson, and the Ledger row records why. If the video turns out to matter later, the link is in the Ledger.
Say you saved a video called "Clean up your inbox with AI". The claim: paste a week of emails into an assistant and get a ranked list of replies owed. You extract a recipe of four steps, and the bar is "every email needing a reply appears, none invented, in under two minutes." You run it on last week's real inbox (stripped of anything confidential). It misses two of nine. Verdict: Fix. You change one thing (ask for a table with the sender and date for each line, so omissions are easy to spot), run again, and it catches all nine. Keep. The recipe goes into the Ledger, the video is deleted, and the pile is one smaller.
The trap is converting everything. Extracting recipes from forty videos feels productive and produces forty unrun recipes, a tidier pile of the same unused promises. The loop only counts when Run happens. A hard limit helps: no more than one video in the Extract stage at a time, and no new Extract until the previous row has a verdict.
It is working when the pile shrinks, when the Ledger has more Keep and Bin rows than blanks, and when you reuse a Keep recipe a second time without rewatching anything. It will not make a weak video good, will not rescue a technique that needs a tool or plan you do not have, and will not replace judgement about whether the output is right.