# Claude never says no to a feature, so I built a skill that makes it argue for deleting one — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/4fd54bdd46fadf76

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-11-03 15:32 ingested 2026-11-04 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xmw242/claude_never_says_no_to_a_feature_so_i_built_a/)

## retained text



```
tl;dr: It's a skill that helps you safely delete features and reduce assumptions. Gist link here: [https://gist.github.com/amos501/919cc1941fb668dc1781bd4cc9a34fc4](https://gist.github.com/amos501/919cc1941fb668dc1781bd4cc9a34fc4)

I've been building a [personal finance app](https://taiga.finance/) with Claude Code, mostly by having an idea and asking for it. That works really well. It's also how I ended up with a codebase where the same calculation or could be built four different ways in different parts of the app.

**What the skill does**

/occam takes one GitHub issue and walks Claude through a fixed set of steps. The parts that matter:

1. **Reproduce first.**
2. **Write down the total number of assumptions and design the simplest model**
3. **Analyze the same five options, every time:** 
   1. keep it
   2. delete it
   3. keep it but label where the number comes from
   4. rebuild with fewest assumptions
   5. or something creative
4. **Crude scoring.** \+1 for each user story the option delivers, −1 for each assumption, magic constant, special case, caveat, or alert.
5. **A second agent red-teams the pick.**
6. **It stops before building**

See the example here

https://preview.redd.it/1b6ec0n00fzh1.png?width=1616&format=png&auto=webp&s=71e9f4d566a90a556eada6f143ec44d3ec2b72da

**Impact**

I ran this on 270 issues and removed 120k lines and added 113k, new tests included, so the codebase got smaller while gaining tests.

Most of the 270 issues came from a separate audit skill that walks user stories through one section of the app looking for duplicate logic, hidden assumptions, and one-off hacks, and files each finding as an issue. /occam then picks up from there.

**Why I think it works**

* Claude's instinct is to add. Making "delete the feature" a row it has to score every time puts removal on the table without me having to argue for it.
* Designing the simple model before seeing the options keeps it from anchoring on the current code.
* A fixed shape and a score turn "is this good code?" made it easy for me to review 20+ issues in a single sitting.

**Caveats**

* It's token heavy. I'd only run this on a subscription plan.

**Try it**

Point it at the part of your codebase you trust least. Or ask Claude to find places where the same thing is computed twice, have it file each one as an issue, and run /occam on them one at a time.
```



document fetched 2026-11-04 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xmw242/claude_never_says_no_to_a_feature_so_i_built_a
