# I made Claude Code write the plan before the code, with no permission to touch the code. Here is where its specs kept going wrong — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/ccef88306f693732

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-11-16 21:17 ingested 2026-11-17 02:40

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xy1jug/i_made_claude_code_write_the_plan_before_the_code/)

## retained text



```
>I'm the author of the app described here, so take this as a project write-up.  
>  
>I use Claude Code daily, and the code is rarely the problem. The problem is what I never told it: what is out of scope, which decisions are already made, what "done" means when something fails. Claude fills those gaps with guesses and I find out in review.  
>  
>So I built Kitsune, a desktop app for the step before coding. I fill in a short brief: why the change is needed, what is in and out of scope, decisions already made, and acceptance criteria including error cases. Then the app starts my own Claude Code CLI on that one change. For that run Claude can read the repo, while shell, web access, and built-in file edits are disabled via `--disallowedTools`. It can only write planning files (specs, design, tasks, or ADRs — or [`spec.md`](http://spec.md) / [`plan.md`](http://plan.md) in a `docs/` workflow) through a scoped MCP tool. The app also adds an MCP server to my user-level config, not the repo, so Claude can ask for the brief, related ADRs, and the API catalogue instead of searching for them.  
>  
>The part I learned the most from was reading what Claude wrote. The same problems kept coming back:  
>  
>  
>  
>These became automatic checks in the app. During and after a run it flags them, critical warnings can be sent back to Claude in a fix loop, and I review a diff of every file afterward with the option to discard the whole run (nothing is committed).  
>  
>Two other things I took from this. Denying code edits during planning changed the output: with nothing to implement, Claude spends the run on the spec instead of rushing to code. And a five-minute brief did more for the result than any amount of prompt wording.  
>  
>Limits: public beta. macOS is signed and notarized, Windows and Linux builds are not signed yet. No model is included, it runs your own Claude Code. Documents stay on your machine; the server holds account and document metadata only. It works with OpenSpec or a structured `docs/` folder (it can scaffold it for you). There is a free tier for briefs, reading, and running the agent to write planning files (up to 3 personal projects); the automatic checks, fix loops, and MCP context tools are part of Solo Pro (with a 14-day free trial).  
>  
>[https://getkitsune.dev](https://getkitsune.dev)  
>  
>What does Claude Code get wrong most often in your plans or specs? I'm deciding which checks to add next.
```



document fetched 2026-11-17 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1xy1jug/i_made_claude_code_write_the_plan_before_the_code
