// the find
lidangzzz/AI-Coding-Style-Guides
A set of coding style guidelines for Vibe Coding or SWE-Agents that maximize efficiency and improve human readability.
A TOML file of prompts (not a tool) that instruct an LLM agent to write minified-looking code — stripped whitespace, one-letter variable names, merged lines — to save tokens in agentic coding workflows. Aimed at people running vibe-coding or SWE-agent loops who feel context-window and API-cost pressure from verbose source.
The 8-level compression table is a genuinely useful framing: a dial from readable to maximally dense instead of an all-or-nothing choice. The worked KMP and JSON-parser examples show real before/after character counts at each level, so you can see concretely what each step costs in readability. It's a single TOML file with no install or build step, so testing the idea costs five minutes, not a dependency.
Every number in the README is a character count, not a token count — tokenizers already compress repeated whitespace and short identifiers cheaply, so the dramatic 23-50% figures almost certainly overstate real token savings, especially at levels 5-8. There's no benchmark anywhere on whether agents write more correct code, debug faster, or cost less end-to-end with this style; only output size is measured, never the cost of running the compression pass itself or the downstream cost of an agent reasoning about variables named `l`, `j`, `t`. "Ask the LLM to decompress it" hand-waves away that stack traces, diffs, and debuggers all show the compressed version — the actual friction of hitting a bug in this code is never addressed. It's a philosophy doc plus prompts from a single contributor, with no linter, formatter, or CI check enforcing any rule, so there's nothing stopping a codebase from drifting out of whatever compression level it claims to target.