idea-mining

---
name: idea-mining
description: Finds the ten short-video ideas your audience already searches for, ranked, each with a hook angle. Use before scripting.
---

# Find the ten video ideas your audience already searches for

Ten ranked ideas, each with its search phrase, a hook angle and why it should work.

## Claude does

1. Collect these, refusing to start without the first two: {NICHE}; {AUDIENCE} (who, and what they want done); {PILLARS} (3 to 5 themes, proposed if absent); {SEED_TERMS} (8 to 15 phrases they would type, drafted if absent).
2. Test every pillar the social-media-skills way: specific to that audience, tied to what the user sells, repeatable weekly, able to carry 10 or more ideas. Fewer than 10 is not a pillar.
3. Ask for 3 to 5 videos that already worked here. Isolate the one or two load-bearing drivers, then run the counterfactual (would it have worked without that?) and the replicability check (can this user shoot it?). Borrow the driver, never the format.
4. Hand permutation and scoring to Codex, then pick ten of its ranked candidates across pillars, no two making the same promise.
5. Write each as: the search phrase, the promise in ten words or fewer, one hook angle as a spoken line, and the driver it borrows. Leave unknowns blank: no volume or view count unless the user gave it or a tool returned it.

## Then Codex does

Codex takes the mechanical half: permuting every seed against a nine-part modifier grid, deduplicating, mapping each row to a pillar and scoring it on a fixed rubric. Hundreds of rows, no judgement.

Fill in {NICHE}, {AUDIENCE}, {PILLARS} and {SEED_TERMS} in CODEX.md first.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

## Claude checks

1. At least 60 candidates, each mapped to a named pillar; fewer means the grid did not run.
2. Strip any number with no source beside it. Uncited house numbers are claude-youtube's known weakness and read as data.
3. Top 20 rows must yield 12 or more distinct promises, or deduplication failed.
4. Cut any idea that cannot open in 1 to 3 seconds (Hoyos). Do not rescue it.
5. Reject rows with an em dash, or the words leverage, fundamentally, delve, harness, hey guys.

Rerun rather than patch by hand; patching hides which half broke.

## Rules

- Copy the driver behind a video, not its surface format.
- A pillar that cannot carry 10 ideas is dropped, not stretched.
- Specifics beat adjectives: "3 takes", not "a few takes".
- For Shorts there is no thumbnail; the first frame is it.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude builds and scores the grid itself, saying so first.

## Built from

- AgriciDaniel/claude-youtube, https://github.com/AgriciDaniel/claude-youtube, 350 stars (api.github.com, 5 Sep 2026): its ideate command.
- social-media-skills/skills, https://github.com/social-media-skills/skills, 66 stars: the pillar test, driver, counterfactual and replicability checks.
- AgriciDaniel/claude-shorts, https://github.com/AgriciDaniel/claude-shorts, 204 stars: the weighted rubric, adapted to score ideas.
- sergebulaev/tiktok-skills, https://github.com/sergebulaev/tiktok-skills, 22 stars: the Voice Rules ban list.
- Jenny Hoyos, https://www.marketingexamined.com/blog/jenny-hoyos-short-form-video-playbook, 10M views per Short, and YouTube's deep dive with Shorts lead Todd Sherman: the 3-second hook.

# Find the ten video ideas your audience already searches for - Codex task

## You are given

- {NICHE}: the subject area.
- {AUDIENCE}: who watches, and what they are trying to get done.
- {PILLARS}: 3 to 5 named themes. Every candidate must belong to one of these.
- {SEED_TERMS}: 8 to 15 phrases the audience would actually type.

## Produce

Two files in the working folder.

1. `candidates.csv`, columns exactly: `id,seed,modifier,phrase,pillar,hook,coherence,emotion,value_density,payoff,total,volume,volume_source,note`. At least 60 rows after deduplication.
2. `shortlist.md`: the top 25 rows by total, grouped under their pillar headings, phrase and total only.

Modifier grid, all nine applied to every seed:

1. question: "how do I X"
2. mistake: "X mistakes"
3. comparison: "X vs Y"
4. price: "what X costs"
5. beginner: "X for beginners"
6. speed: "X in 30 seconds"
7. cause: "why X fails"
8. constraint: "X without Y"
9. proof: "does X actually work"

Scoring: rate each of the five dimensions 0 to 100, then
`total = hook*0.30 + coherence*0.25 + emotion*0.20 + value_density*0.15 + payoff*0.10`.
Drop every row scoring under 60 before writing the files.

## Rules

- Leave `volume` and `volume_source` empty unless a keyword tool already present in this environment returned a figure. If one did, name the tool in `volume_source`. Never estimate a volume, a view count or a percentage.
- Deduplicate near-identical phrases across case, plurals, stopwords and word order. Keep the shortest wording.
- Every row maps to exactly one pillar from {PILLARS}. If none fits, set `pillar` to `unassigned` rather than forcing it.
- British English. No em dash characters anywhere.
- Banned words: leverage, fundamentally, delve, harness, hey guys.
- Specifics beat adjectives: "3 takes", not "a few takes".
- Do not write scripts, hooks, captions or thumbnails. That is a different task.
- Do not browse for private or paywalled data, do not create accounts, do not publish or send anything.

## Return

Print, in this order:

1. Row count before deduplication and after.
2. Count per pillar, including `unassigned`.
3. Count dropped for scoring under 60.
4. How many rows carry a volume figure, and the tool named in `volume_source`. Print `0 / none` if there is no tool.
5. The top 10 phrases with their totals.
6. The absolute paths of `candidates.csv` and `shortlist.md`.