The task
It’s Sunday night. You promised yourself you’d batch-film 4 videos this week and you have ideas for 1.5 of them. You also know, somewhere in your analytics, that 3 of your last 20 videos hit much harder than the rest, but you can’t quite articulate the pattern, and the videos you’ve been gravitating toward this week are closer to your flops than your hits. You want a list of 20 short-video ideas for the next 4 weeks that lean toward the patterns your audience has already voted for, mixed with a few uncomfortable bets in pillar areas you haven’t tested yet.
TL;DR
- Feed an AI your top 5 performers, bottom 3 flops, content pillars, and what you can actually film solo — get back 20 ideas with named hook angles, formats, and a 0-3 risk score, weighted toward your proven patterns.
- The pattern only matters because the platform rewards it: TikTok for Business research finds 63% of the highest click-through videos hook viewers inside the first 3 seconds, and TikTok’s 2026 distribution leans hard on completion rate, so the hook angle is the deliverable, not the topic.
- Which model: Claude Opus 4.7 or Sonnet 4.6 (1M-token context) handles a full analytics dump best; GPT-5.5 Thinking is the fastest generalist for the 5-idea sprint. Either works on free tiers for one prompt; paid plans start at $20/mo (as of June 2026).
- AI does the pattern extraction; you still do the weekly trend scan. Pull live demand from TikTok Creator Search Insights and feed the gaps back in.
- Re-prompt with fresh top/bottom data every 2-3 weeks — pattern data drifts, and old hits don’t predict next month.
Where AI helps — and where it does not
AI is strong at pattern-finding across your top performers: extracting the hook structures, formats, and topic angles that consistently work for your account, and applying them to new topics. It is also good at the “don’t repeat” filter; once you tell it why your bottom 3 flopped (your hypothesis), it will stop generating ideas that share those traits.
What AI cannot do: predict what’s about to trend. That is humans-on-the-platform work. Run a weekly scan with TikTok Creator Search Insights — its Content Gap and Searches by Followers filters surface topics your audience is searching for but nobody is serving — and watch what your peer accounts posted in the last 72 hours. Pair AI ideation (pattern-based, evergreen) with that weekly trend scan (timely, opportunistic), then feed the gap topics back into the prompt as extra pillars. AI also cannot judge which ideas are shootable with your current setup; a “dance with 4 people in a studio” idea is worthless if you film alone at home.
A specific failure mode: AI defaults to safe, slight-variation ideas that mirror your top performers too closely (same hook, same format, slightly different topic). After 8 weeks of this, audiences see the pattern and engagement drifts. Tell it explicitly: “at least 3 ideas must be uncomfortable for me, with a different format, different pillar, or different vulnerability level.”
Which AI model to use (June 2026)
Ideation is a long-context, pattern-matching job: you’re pasting a wall of analytics and asking the model to find the through-line. The differences below matter when you scale to a full month of data.
| Model | Context window | Best for this task | Plan to access |
|---|---|---|---|
| Claude Opus 4.7 | 1M tokens | Largest analytics dumps; strongest at naming a non-obvious pattern across 20 videos and holding your voice across all 20 ideas | Claude Pro $20/mo ($17 annual) |
| Claude Sonnet 4.6 | 1M tokens | Same context window, faster and cheaper; the default workhorse for the full 20-idea prompt | Claude Free (limited) or Pro $20/mo |
| GPT-5.5 (Thinking) | ~320 pages in-app (Plus); 1M only on $200 Pro | Fast 5-idea weekly sprints; strongest broad generalist | ChatGPT Free (tight limits) / Plus $20/mo |
| Gemini 3.1 Pro | 1M tokens | Good if you also want it to cross-check ideas against live search demand | Google AI Pro $19.99/mo |
Pricing and tiers as of June 2026, taken from each vendor’s pricing page (Anthropic, OpenAI, Google). For one month of ideation you almost never need the picker’s top reasoning tier — the workhorse model (Sonnet 4.6 or GPT-5.5 Thinking) is enough. Reach for Opus 4.7 only when you paste 8+ weeks of raw analytics and want the subtle pattern named precisely.
What to feed the AI
- Your niche and 3-5 content pillars (the topics you actually want to be known for)
- Your top 5 performing videos in the last 8 weeks, each with: hook (verbatim), topic, format, retention curve summary. Pull the curve from your native dashboard — TikTok’s Search analytics, YouTube Studio’s audience-retention graph, or Instagram’s Reels insights — and note the exact second viewers drop off (strong Shorts hold 70%+ through the first 3 seconds)
- Your bottom 3 performers in the same window, each with: hook, topic, format, and your hypothesis about why it flopped
- The format mix you can actually shoot (talking head, voiceover-over-b-roll, on-screen text only, duet/stitch); solo creators have a different mix than team-supported ones
- The hook patterns you’ve already tested too many times (so the model doesn’t repeat them)
- One pillar you’ve been avoiding because it feels uncomfortable but the audience keeps asking for
- Your posting cadence (3/week, daily, 4/week); affects whether 20 ideas is one month or two
- The “voice you” (formal, dry, sarcastic, earnest), and one peer creator whose tone yours most resembles
Copy-ready prompt
Paste this into Claude (Sonnet 4.6 or Opus 4.7) or ChatGPT (GPT-5.5 Thinking). Replace every [bracketed] placeholder with your own data first.
You're helping me ideate 4 weeks of short-form video at [cadence].
Niche: [paste]
Content pillars: [3-5]
Top 5 performers (hook verbatim + topic + format + retention summary): [paste]
Bottom 3 (hook + topic + format + my flop hypothesis): [paste]
Format mix I can shoot solo: [paste]
Hook patterns I've used too often: [paste]
Pillar I've avoided but audience asks for: [paste]
Voice: [dry / earnest / sarcastic / playful] — like [peer creator]
Generate 20 short-video ideas. For each, return:
1) One-line topic — specific enough to film tomorrow.
2) Hook angle — the actual first 3 seconds, not just "the topic." Reference which pattern family (counter-intuitive / stake / number / admission / open question / if-you / before-after / mid-action).
3) Format — storytime / tutorial / reaction / list / voiceover-b-roll / on-screen-text-only.
4) Which pillar it ladders to.
5) Why this lines up with my top performers (one phrase reason — what pattern from my hits it inherits).
6) A 0-3 "risk" rating: 0 = safe variation of past hit; 1 = adjacent angle; 2 = new pillar or new format for me; 3 = uncomfortable bet.
Rules:
- At least 3 ideas must be risk-2 or risk-3. The other 17 should lean toward what worked.
- Vary the format across ideas; do not produce 12 talking-head storytimes in a row.
- For each idea, the hook must be a different first-3-seconds than my last 5 hits.
- Do not repeat hook patterns I marked as overused.
- If an idea couldn't actually be filmed with my solo setup, flag it with [NOT SHOOTABLE] and drop.Shorter variant — fast 5-idea sprint
Niche: [paste]. Top 3 hits this month: [hooks]. Give me 5 shootable ideas for this week only. Each: one-line topic, hook (specific first 3 seconds), format. Lean toward my hit patterns. No "thought leadership" generic ideas.Sample output
A useful idea row: “Topic: I tracked my AI tool spend for 90 days and the ‘cheap’ tools cost me more than the premium ones. Hook angle (admission pattern): ‘I switched to the cheap AI plan and it cost me 4x more — here’s the spreadsheet.’ Format: voiceover over screen recording of the spreadsheet. Pillar: AI tooling decisions. Inherits from: your Notion vs. Linear breakdown that hit (admission + spreadsheet evidence). Risk: 1.”
A useful uncomfortable idea: “Topic: The AI tool I built that nobody used. Hook angle (admission): ‘I built an AI feature my users explicitly asked for. Then 7% of them used it.’ Format: talking head + screen capture of the feature. Pillar: product judgment. Inherits from: your ‘why we sunset feature X’ video (admission + specific number). Risk: 3 — first time you talk publicly about a feature failure.”
A useful “what to drop” reflection: “Idea dropped: tutorial on prompt engineering basics. Reason: your tutorial-format videos have flopped twice in the last 8 weeks; your audience prefers admission + specific-number patterns over how-tos. Save this topic for a written post.”
How to refine
- Force a risk distribution: “Re-read the 20 ideas. Confirm at least 3 are risk-2 or risk-3, and at least 12 are risk-0 or risk-1 (safe variations or adjacent angles). If the mix is off, swap ideas until it balances.”
- Make hooks specific, not topics: “For each idea, the ‘hook angle’ must be the actual sentence I’d say in the first 3 seconds, not a topic restatement. ‘About AI failure modes’ is a topic; ‘I built an AI feature my users asked for and 7% used it’ is a hook.”
- Tie each idea to a past hit: “Every idea must include a one-phrase reason explaining what pattern from my hits it inherits. If you can’t connect it to a hit, the idea is a guess, not a pattern bet.”
- Drop the unshootable: “If an idea requires a team, a location, or equipment I don’t have, mark [NOT SHOOTABLE] and replace with a similar idea I could film tomorrow. Solo creators can’t ideate from team-creator examples.”
- Vary format: “No more than 6 of the 20 ideas should share the same format. If 12 are talking-head, force 6 of them into voiceover or on-screen-text-only formats and reshape the hook accordingly.”
Common mistakes
- Ignoring your bottom performers: they teach more than top, because they tell you what your audience explicitly rejected; always feed the model both
- Same format repeatedly: your top hit was talking-head storytime, so AI gives you 18 talking-head storytimes; audiences notice within 4 weeks and engagement softens
- No hook angle stated: AI gives you topics, you write hooks tired at midnight; the hook is the idea, not the addendum
- Treating AI ideas as scripts: these are ideas with hook angles; scripting comes after, with the short-video-script-with-ai workflow
- Not refreshing pattern data: your hits 3 months ago don’t predict what hits this month; re-prompt every 2-3 weeks with fresh top/bottom data
- Picking only risk-0 ideas: safe variations on past hits eventually stop being hits; you need a few uncomfortable bets per month for the channel to grow
- Generating ideas without pillars. Without pillars, AI produces a grab-bag and your audience can’t tell what your channel is “about”
- Letting AI generate 20 and shooting all 20. Quality over volume; pick the 8-12 you actually believe in and let the rest go
FAQ
- How often should I re-prompt with fresh data?: Every 2-3 weeks if your cadence is daily or 4/week, every 4 weeks if you post 2/week. Pattern data drifts; what worked 6 weeks ago may not work now, especially on TikTok where algorithm preferences shift.
- Should I make all 20 ideas?: No. Pick 8-12, lean toward the ones with high “inherits from your hits” signal plus the 2-3 uncomfortable bets. The remaining ideas are scaffolding; they’re proof you’ve considered alternatives, not a shooting list.
- What if my top performers are flukes?: Run the prompt with an 8-week window instead of 4 weeks. Flukes wash out at 8 weeks, and the real pattern emerges. If after 8 weeks the top 5 are still flukes, your account is too early in pattern formation; keep posting and re-run in 4 weeks.
- Can I use this for YouTube long-form ideation too?: Partially. The pattern-extraction works; the format mix needs to be replaced with long-form formats (deep-dive, breakdown, story-arc). Hook patterns translate; the timing does not (long-form hooks are 30-60 seconds, not 3).
- The model keeps generating safe ideas. What changes?: Add: “Push me. At least 3 ideas should make me uncomfortable to film: a pillar I’ve avoided, a vulnerability level I don’t usually show, or a format I’m bad at. Without these, the channel doesn’t grow.”
- Which AI should I use — ChatGPT or Claude?: For the full 20-idea prompt with a month of pasted analytics, Claude Sonnet 4.6 or Opus 4.7 handles the long context best (1M tokens at standard pricing, as of June 2026) and holds your voice across all 20 ideas. For the quick 5-idea weekly sprint, GPT-5.5 Thinking is faster. Both have free tiers that can run one prompt; paid plans (Claude Pro, ChatGPT Plus) are $20/mo.
- Does the risk-0 vs risk-3 framing actually map to the algorithm?: The risk score is about your channel’s pattern, not the algorithm directly — but they connect through the hook. TikTok for Business reports 63% of the highest-CTR videos hook viewers in the first 3 seconds, and 2026 distribution leans on completion rate (strong Shorts hold 70%+ through three seconds). A risk-0 idea reuses a hook that already cleared that bar for you; a risk-3 idea is an untested hook that might clear a higher one. You need both.
Related
- Reel hook prompts
- Content calendar prompts
- Content calendar creator AI
- Short video script with AI
- TikTok hook AI
- Short-form video hook AI
- Cross-platform repurpose
Tags: #AI writing #Content creation #Creator #Short-form video