A Claude skill that builds Instagram Reel openings that hold attention. Spoken hook, on-screen text, open loop, rehook, visual anchor, payoff.
Works for Reels, TikToks, and Shorts. Installs into Claude Code, Codex, Cursor, and 60+ other agents.
npx skills add arhamhi/hooksmith --skill '*' -a claude-codeFor other agents, swap the target (-a codex, -a cursor) or drop the flag to pick interactively.
Then teach it your context once:
/hooksmith init
or in any agent:
Use hooksmith init for my creator account.
Init writes a HOOKS.md file with your niche, viewer, proof, voice, formats, and forbidden claims. Every later command reads it, so the hooks come out sounding like you instead of like a swipe file.
Most hook prompts give you something technically fine and completely scrollable.
Before:
Want to start an AI agency? Here are 5 mistakes to avoid so you can get more clients.
There is nothing really wrong with it.
That is the problem. It sounds like every other post.
Hooksmith makes the viewer feel called out before it explains anything:
After:
If you've been building your little AI agency for a few months now...
watched all the videos, joined the communities, switched your offer like four times...
and you still haven't closed a single client...
or even if you closed one or two, and then it just... stopped.
and now you're sitting there stuck...
and you keep telling yourself it's the next tool. the next niche. the next funnel.
but if you're honest, you've been busy for months and you haven't actually moved.
That is the whole point.
Do not ask your agent for "10 viral hooks." It will give you 10 tiny variations of the same flat line. Use Hooksmith when you want the full opening system.
A good Reel opening is not one clever sentence. It is a retention path. Every generate option comes back like this:
Spoken hook:
If you fell off the gym for 2 months, do not restart with a full routine. Start with the 3-day version.
On-screen text:
The 3-day gym restart
Open loop:
What are the 3 days and why is day one not a full workout?
Rehook:
Day one is not even there to train you.
Visual anchor:
Show a crossed-out 6-day split, then a simple 3-day card.
Payoff:
Reveal the 3-day restart and explain why it removes the pressure that makes people quit again.
Notice the spoken hook and the on-screen text do different jobs. That split is one of the core frameworks inside.
| Command | Use it for |
|---|---|
init |
Create HOOKS.md for your niche, voice, formats, proof, and CTAs. |
generate |
Create 3-5 meaningfully different hook systems. |
audit |
Score a hook or first 15-30 seconds out of 50. |
rewrite |
Rebuild a weak opening into a stronger hook stack. |
identity |
Create viewer-first identity-stack openings. |
rehook |
Add attention resets inside the first 3-10 seconds. |
caption |
Create caption hooks and comment prompts. |
Example prompts:
Use hooksmith generate for this Reel:
Topic: how busy people can start training again without a perfect routine.
Payoff: a 3-day restart plan.
Proof: I used this after falling off for 2 months.
Format: talking head with gym B-roll.
Use hooksmith audit:
[paste the first 15-30 seconds or your hook stack]
Use hooksmith rewrite:
[paste weak opening and intended payoff]
The skill lazy-loads references per command, so it stays cheap until you use it:
- the Three Rules every hook must pass
- two-pronged hooks: visual and spoken tracks doing different jobs
- the YouTube-title method for on-screen text, with keyword tables and ten fill-in templates
- six hook families with formulas, plus ten named patterns
- the identity stack, with a fully annotated reference example
- the two-sentence amplifier and its filler test
- a 50-point audit rubric
- an anti-pattern list it checks before every final output
A private version of this workflow helped shape a Reel that reached 110K views from an account with around 50 followers.
That does not mean Hooksmith guarantees views. It means the underlying workflow was used on a real breakout Reel before being cleaned up into a public skill.
Read the public-safe case study in examples/case-study-110k-reel.md.
Hooksmith should reject:
- fake proof
- fake follower counts
- fake client stories
- fake timelines
- fake revenue claims
- hooks that promise virality
- hooks that blame the viewer
- hooks with no real payoff
If the input is too vague, Hooksmith should ask for the missing payoff instead of returning generic lines.
Fork it, remix it, and add niche packs if you have real examples.
Good contributions include:
- stronger hook audits
- public-safe before/after examples
- niche-specific keyword banks
- better fake-proof refusal tests
- new command references that stay focused on short-form retention
If this helps your agent write better Reel openings, star the repo and send it to someone still asking ChatGPT for "10 viral hooks."