If your AI keeps giving you ideas that sound like everyone else’s, you need this. The Taste Filter makes AI challenge its own first answer before it ever reaches you.
AI is trained to predict the most likely answer, so it drifts toward the most average, expected version of everything. That’s great for research, where you want the consensus. It’s also exactly why your ads, scripts and captions come out sounding like everyone else’s.
So I built a skill that makes AI attack its own first draft. Before it hands me a hook, a script or an ad, it has to check whether this is the obvious answer every other AI would give, and whether a competitor could post the exact same thing. Then it rewrites, scores the rewrites and shows me what it killed. This is one of the ways I grew from zero to 300,000 followers in six months.
Turn on code execution. In Claude, go to Settings, then Capabilities, and turn on Code execution and file creation. Skills work on every plan, including Free.
Upload it. Go to Customize, then Skills. Click +, then Create skill, then Upload a skill, and pick the ZIP. Don’t unzip it first.
Use it. Ask for angles on anything, or paste a draft and say "run this through the Taste Filter."
Claude decides when to use a skill by reading its short description, so this one kicks in when you brainstorm or ask for the Taste Filter by name. (How skills work) Prefer a plain file? Download the .md and paste it into a Claude Project’s instructions instead.
What It Does
01
Generates with no format menu
Ten raw angles, with zero example formats. The second you suggest "a listicle or a before/after," you’ve capped the thinking at what already exists.
02
Attacks its own work
Every candidate runs a gauntlet. If any competitor could post it, if anyone would land on it in 15 seconds, or if it opens with "POV:", it gets tagged obvious. Most first drafts fail, and that’s the point.
03
Rewrites what failed
Each failure gets three reframes. It might flip the assumption or argue the unpopular side. It might borrow from a totally unrelated field, drill into one hyper-specific moment, or talk to one person instead of broadcasting.
04
Scores and shows the kill list
Taste, fit and specificity out of 10, ranked by the lowest score. You get the top 3, the emotion each one lands, and the list of what it rejected so you never get those ideas pitched back to you.
No Install? Use The Prompt
Paste this into any chat in Claude or ChatGPT when one thing came out sounding a little too safe.
Taste Filter, One-Off
Run this through my Taste Filter before you give it back to me.
WHAT IT IS: [hook / script / ad / caption / email / list of ideas]
WHO IT'S FOR: [the audience, as specifically as you can]
WHAT IT HAS TO DO: [stop the scroll / get the click / get a reply / sell X]
THE DRAFT: [paste it, or write "none, give me angles on: TOPIC"]
1. Ask yourself honestly: is this the obvious answer every other AI would give? Could any competitor in my category post this exact thing? If yes, say so.
2. Write 3 rewrites, each using a different move: flip the assumption, argue the unpopular side, or drill into one ultra-specific moment my audience would instantly recognize.
3. Score each rewrite out of 10 for taste (how unlike generic AI it is), fit (does it work for my audience) and specificity. Rank by the lowest score, not the average. Cut anything under 6.
4. For the winner, name the emotion it lands: recognition, curiosity, quiet validation, mild surprise, shared complaint or tension.
Give me: the winning version, one line on why it isn't generic, then a short "killed" list of what you rejected and why.
Rules: no "POV:", "Imagine if" or "Did you know" openers. Don't invent stats or results. If it needs a real number from me, write [NEEDS REAL NUMBER].
The Full Skill
Here’s the whole SKILL.md if you want to read it or tweak it before you install.
taste-filter / SKILL.md
---
name: taste-filter
description: Stops generic AI output. Attacks its own first draft of hooks, scripts, ads, posts and ideas, kills the obvious and rewrites sharper. Use for brainstorms or "run this through the Taste Filter".
---
# The Taste Filter
AI predicts the most likely answer, so its first draft drifts toward the most average, expected version of everything. That's useful for research, where you want the consensus. It's the reason creative work comes out sounding like everyone else's.
The Taste Filter has one move: **generate, attack your own output, then rewrite what fails.** Never hand over the first draft.
## Two modes
- **Ideas mode.** The user wants angles, hooks, concepts or campaign ideas on a topic. Run Steps 1 to 5.
- **Filter mode.** The user hands you a draft (a hook, script, caption, ad, email) or says "run this through the Taste Filter." Skip Step 1. Treat their draft as the only candidate and run Steps 2 to 5 on it, then return the rewritten version.
## Context first
Use whatever real context you have: the brand, the audience, the product, the data, what is actually true. If the chat, the Project or attached files describe the brand or audience, read them before you start. If there's no context, work from the topic and ask at most one question, only if the answer would change the angle completely.
## The rule that makes this work
**Never show the user a menu of formats or example angles to pick from.** Examples in a prompt become the ceiling. The moment you offer "you could do a listicle, a before/after or a testimonial," you've capped the thinking at what already exists.
## Step 1: Generate with full latitude
Generate 10 candidate angles. No example formats, no genre lists, no "here are some popular approaches."
## Step 2: Attack every candidate
Tag a candidate `obvious` if any of these are true:
- **Could this exact angle belong to any competitor in the category?** (The sharpest test. Use it hardest.)
- Would anyone who thought about the topic for 15 seconds land on it?
- Is it a talking-head testimonial, a before/after or a listicle ("3 things I wish I knew")?
- Does it open with "POV:", "Imagine if", "Ever wondered", "Picture this" or "Did you know"?
- Does it lean on a worn trope: morning routine, what's in my bag, get ready with me, day in the life?
- Does the brand or product show up before the audience gets anything of value?
- Is it aspirational lifestyle that quietly signals "you'll never have this"?
- Does it claim a result with no specifics?
- Would it make someone stop scrolling? If you can't say why in one sentence, it won't.
Be ruthless. Most first drafts are obvious. That's expected.
## Step 3: Rewrite everything you tagged
For each tagged candidate, write 3 reframes using these transforms. They are rules to apply, not examples to copy.
- **Inversion.** Flip the assumption. If the obvious version says "here's what to do," the inverted one says "here's why what you're doing works, and when it doesn't."
- **Contrarian.** Argue the unpopular position. If everyone agrees on X, find the real case where X is wrong.
- **Cross-domain.** Connect the topic to an unrelated field: architecture, weather, cooking, music, sports. The collision is where the idea lives.
- **Specificity drilling.** Drill into one ultra-specific case the audience will recognize instantly. The narrower it gets, the more people feel seen.
- **Direct eye contact.** Stop broadcasting. Write to one specific person about a thing only they would notice.
## Step 4: Score honestly and cut
Score every candidate and reframe out of 10 on:
- **Taste:** how unlike generic AI output is this, really?
- **Fit:** does it work for this brand, audience and product?
- **Specificity:** is there a real, particular observation in it, or a shape with nothing inside?
All three must be 6 or higher. Rank by the **lowest** of the three scores, not the average. A great idea that fails one axis is a bad idea. Return the top 3 (ideas mode) or the single best rewrite plus two alternates (filter mode).
## Step 5: Name the emotion
Every survivor must land one of these, and you must name which: **recognition** ("that's me"), **curiosity** (an open loop), **quiet validation** ("I wasn't crazy"), **mild surprise** (a small reversal), **shared complaint** ("finally someone said it"), **tension** (something at stake). If you can't name it, it isn't finished.
## Output format
```markdown
# Taste Filter: {topic}
## Top 3
### 1. {short name}
**The angle (or rewrite):** {one or two sentences, or the full rewritten draft in filter mode}
**Why it isn't generic:** {one specific sentence}
**Emotion:** {one of the six}
**Scores:** taste {n}/10 · fit {n}/10 · specificity {n}/10
### 2. ...
### 3. ...
## Killed
- {candidate}: {why it was obvious}
```
Always show the kill list. It proves the work happened and stops the user from re-suggesting ideas you already rejected.
## Hard rules
1. No format hints during generation.
2. Never return anything that hasn't been through Step 2.
3. All three scores 6 or higher. Rank by the minimum.
4. Name the emotion or it isn't done.
5. Show the kill list.
6. Never invent facts, stats or results to make an angle sharper. If an angle needs a number the user hasn't given, mark it [NEEDS REAL NUMBER].
The Real Win
AI doesn’t come with taste. You have to build it in. This skill bakes your standard, "is this actually good, or just likely," into everything the AI hands you.