TL;DR
Collect 15-25 of a competitor’s recent posts (hook + format + engagement numbers), paste them into Claude or ChatGPT with a teardown prompt, and ask for pattern-level findings: their hook formula, which format wins on which metric, the topic clusters they repeat, and the one gap you can own. AI does the analysis; it cannot scrape the numbers for you, so you bring the data. Below is the exact prompt, the model to use as of June 2026, and the four mistakes that produce useless output.
The task
You suspect a competitor in your niche is doing something structurally smarter than you — sharper hooks, a tighter format, a smarter cadence. You want a teardown that tells you what to reuse at the pattern level, not “post more videos.” Twenty posts is enough signal to spot a hook formula and a format-engagement split; five is not.
When this is the right job for AI
- You can hand the model 15-25 recent posts (titles, hooks, captions, engagement numbers).
- You want pattern-level insight, not “they get more likes.”
- You will adapt patterns to your own voice. AI is for analysis, not for cloning.
This is not the right job for AI when you want it to fetch the competitor’s posts. Neither ChatGPT nor Claude can reliably scrape current Instagram, TikTok, Xiaohongshu, or LinkedIn engagement counts — you paste the data; the model reasons over it.
Which model to use (as of June 2026)
Twenty posts plus hooks and captions is a few thousand tokens — well inside any current model’s window — so pick for reasoning quality, not raw context size.
| Model | In-app context | Files per message | Best for |
|---|---|---|---|
| Claude Opus 4.7 (Pro $20/mo) | 1M tokens (200K soft cap, then RAG) | 20 | Deepest pattern reasoning; persistent Projects knowledge base for one competitor over time |
| ChatGPT GPT-5.5 Thinking (Plus $20/mo) | ~196K-256K when “Thinking” is selected | 20 (raised from 10 on Feb 13, 2026) | Fast structured teardowns; pick Thinking, not Instant |
| Gemini 3.1 Pro (Google AI Pro $19.99/mo) | 1M tokens | — | Large multi-competitor batches in one pass |
Practical default: Claude Opus 4.7 for the analysis itself. If you track the same competitor every quarter, put their post log in a Claude Project so the model keeps the running history as persistent knowledge — Projects let you store far more than the 20-file chat cap, and Claude switches to retrieval (RAG) mode automatically once the log exceeds the 200K-token soft cap. Either way, paste engagement numbers manually; the model does not have them.
What to feed the AI
- Competitor handle + platform.
- 15-25 posts: title or first line, format (carousel / short video / long post), engagement (likes, comments, shares, saves).
- The three questions you want answered: hook formula? cadence? topic clusters?
- Your audience: who you are writing for.
Pull the engagement numbers from the platform’s own analytics or a scheduler export (HubSpot, Buffer, and most schedulers export likes/comments/shares per post to CSV). Include the post’s format every time — it is the single most useful column, because format usually splits engagement harder than topic does.
Copy-ready prompt
You are doing a content teardown for me.
Competitor: @creatorname (Xiaohongshu)
My audience: Chinese-speaking junior PMs (1-3 yrs experience), curious about AI workflows.
Recent posts (last 20, format: hook | format | engagement):
1. "I rewrote 3 years of PRD templates with AI" | carousel | 4.2k likes
2. "5 things GPT can do in a PM's day job" | text | 1.1k likes
3. "Cursor wrote my first user-interview guide" | short video | 8.7k likes
... [paste 20, in the original language used by the competitor]
Find:
1. The hook formula (their top 3 posts have hooks that share what structural property?).
2. The format-engagement pattern (does short video outperform carousel for them, and on which metric?).
3. The 3 topic clusters they keep returning to.
4. The thing they do NOT do that I could.
5. Two specific posts I should NOT copy directly because they are off-brand for my audience.
Cite at least one specific post number for every claim.
Output as a 5-bullet teardown. No "they are great at content"-style filler.Sample output structure
- Hook formula: all three top posts (#1, #3, #7) open with a number + a specific named tool (“I rewrote 3 years of PRD templates with AI”). Numbered + named-tool hooks out-perform generic “AI in a PM’s day job” hooks roughly 4:1 here.
- Format pattern: short video wins on reach (avg 8k likes vs carousel 4k), but carousels collect ~3x the comments. They optimize for reach; you could optimize for community signal — which matches the public 2026 data that static formats out-engage video per impression even when video wins discovery.
- Topic clusters: PRD workflows, user-research prompts, and Cursor for non-engineers. Two of the three overlap your audience.
- Gap: they never cover “AI for the junior-PM career” — résumé, interview prep, debriefs. That is your wedge.
- Skip: post #2, the “5 things GPT can do in a PM’s day job” listicle — generic, aimed at PM-curious non-PMs, not your audience.
The format-engagement split is real and worth checking against an outside baseline. Buffer’s 2026 analysis of 52M+ posts found Instagram carousels drove about +109% more engagement than Reels and beat them roughly 2x on engagement-per-impression, while video still won on reach — so “video for reach, carousel for comments” is usually a sound read, not the model hallucinating.
How to refine the output
- Output too high-level → add “every claim must reference at least one specific post number.”
- Wrong audience → add “do not analyze for a general PM audience; analyze for MY audience as defined above.”
- AI being sycophantic → add “be willing to say
they are wrong on Xif the data supports it.” - No clear gap → add “the deliverable is one wedge: a topic THEY are not covering that MY audience needs.”
Common mistakes
- Pulling 5 posts instead of 20. Too small a sample to separate a real formula from noise.
- Looking only at top performers. Their flops tell you what to skip; include a few low-engagement posts on purpose.
- Asking AI to fetch the engagement numbers. It can’t scrape — paste them.
- Copying the hook verbatim. Adapt the structure; bring your own specifics. Reusing a competitor’s exact wording reads as derivative to their audience and yours.
FAQ
- What if my competitor only posts video? Give the model the spoken hook (first 3 seconds) plus the thumbnail/cover text. It can analyze the hook structure without watching the clip.
- How often should I redo this? Quarterly per top competitor. Hook and format patterns shift slower than the algorithm panic suggests.
- Which AI is best for this in June 2026? Claude Opus 4.7 (Pro, $20/mo) for the reasoning; use a Claude Project if you track the competitor over time. ChatGPT GPT-5.5 in Thinking mode is a strong alternative — both let you attach up to 20 files per message.
- Can AI pull the competitor’s posts for me? No. ChatGPT and Claude cannot reliably scrape live engagement counts from Xiaohongshu, Instagram, or TikTok. Export from platform analytics or a scheduler, then paste.
- Can I run the same teardown on my own content? Yes — swap “competitor” for “my last 20 posts.” It is the most honest retrospective you can run, because the prompt forces post-level citations instead of vibes.
Related
- AI short-video ideation
- AI content calendar for creators
- AI personal brand voice design
- AI Xiaohongshu series content
- Comment-to-Content Prompts: Turn Replies Into Posts
External references: Claude file & context limits · Buffer 2026 content-format study
Tags: #AI writing #Creator #Content creation #Competitor #Social media