How do you make an AI assistant follow your project's coding conventions?

Question

Accepted Answer

An AI defaults to a generic average of its training data, so to get **your** style you have to **show and enforce** it — through examples, an explicit rules file, and a linter loop that catches what slips through. Don't rely on it to guess your standards (WordPress Coding Standards, PSR-12, etc.).

## Techniques (in order of leverage)

- **Provide the linter config and examples** — paste your `.eslintrc`, `phpcs.xml`, or `.editorconfig`. These are the rules in machine-checkable form, and the AI can read them.
- **Keep a rules/instructions file** — a `CLAUDE.md` or system prompt that states the standard explicitly: "This project follows **PSR-12**. Use 4-space indent, braces on the next line for functions, snake_case for hooks." Standing instructions beat repeating yourself every prompt.
- **Show representative existing code** — one or two clean files teach naming, structure, and idioms faster than any description. The AI pattern-matches what it sees.
- **Run the linter on AI output and feed errors back** — the most reliable step. The linter is the ground truth; paste its errors and ask the AI to fix them.

## The linter loop

```bash
# 1. AI generates the change, then you run the project linter on it:
phpcs --standard=phpcs.xml src/NewClass.php
# => src/NewClass.php  12 | ERROR | Expected 1 space after FUNCTION keyword; 0 found
#    src/NewClass.php  14 | ERROR | Line indented incorrectly; expected 4 spaces, found 2

# 2. Paste those exact errors back to the AI: "Fix these phpcs errors."
# 3. Re-run until clean. The linter — not the AI's claim — is the source of truth.
```

This loop converges fast because the linter gives precise, line-level feedback the AI can act on directly, instead of you describing the style in prose.

## Why it matters

Code that ignores project conventions creates **review friction and inconsistency** even when it's functionally correct — reviewers waste time on formatting, and the codebase drifts. An AI has no awareness of your house style unless you supply it, so the work is to make the rules **explicit and checkable**: a rules file for intent, representative code for idioms, and the linter as the enforcer in the loop. The same discipline (config + examples + automated check) is how you onboard a human teammate to your conventions, too.