compress-code-responses

Reduce AI output token cost by having the model return format-minified code (whitespace and line breaks only), then restore readable formatting locally with Prettier.

Works with:

Tool	How it loads instructions
Cursor	.cursor/skills/compress-code-responses/SKILL.md
OpenCode	.opencode/skills/, .claude/skills/, or .agents/skills/
GitHub Copilot CLI	.github/copilot-instructions.md + .github/instructions/*.instructions.md
PHPStorm + Copilot	Same .github/ files as Copilot CLI (details)

How it works

1. Model output: code is sent with minimal indentation, blank lines, and spacing. Variable names, strings, comments, and logic stay the same.
2. Local restore: after files are written, run Prettier with this repo's config to format them back.

You pay fewer completion tokens; your repo still gets normally formatted code.

Quick start

Install the default skill:

git clone https://github.com/rgarciar1931/compress-code-responses.git
cd compress-code-responses
npm install
npm run install-skill
npm run ensure-deps

On Windows PowerShell, if script execution policy blocks npm, use npm.cmd:

npm.cmd install
npm.cmd run install-skill
npm.cmd run ensure-deps

Use a specific release when needed:

# Stable release
git checkout v1.0.0

# Adobe Commerce / Magento 2 release
git checkout v1.0.0-adobe

Adobe Commerce users can instead install from the adobe-commerce branch for Magento-specific formatting support:

git checkout adobe-commerce
npm install
npm run install-skill

npm run install-skill writes user-level copies under paths such as ~/.cursor/skills/, ~/.claude/skills/, ~/.agents/skills/, ~/.config/opencode/skills/, and ~/.copilot/copilot-instructions.d/.

To vendor only the skill in one project, copy skills/compress-code-responses/ to that project's .cursor/skills/, .claude/skills/, .agents/skills/, or .opencode/skills/ directory.

Formatting after AI edits

From the directory that contains this package:

npm run ensure-deps
npm run format -- src/foo.js src/bar.php

Check without writing:

npm run format:check -- src/foo.js

Token comparison

The benchmark simulates "without skill" as normally formatted code and "with skill" as whitespace-minified model output, then verifies Prettier can restore a layout-equivalent file locally.

Latest local run against richer JS, PHP, and JSON fixtures:

Measure	No skill	Skill raw	Reduction
Characters	5,231	4,059	22.4%
OpenAI cl100k_base	1,415	1,048	25.9%
OpenAI o200k_base	1,424	1,091	23.4%
OpenAI p50k_base	1,770	1,359	23.2%

Per sample:

Sample	Chars no skill	Chars skill	Char reduction	Restored
JavaScript	1,869	1,505	19.5%	yes
PHP	2,221	1,688	24.0%	yes
JSON	1,141	866	24.1%	yes

Run it yourself:

npm test
npm run test:token-comparison

The script writes detailed artifacts to test-output/token-comparison/:

• *-1-no-skill-formatted.txt: formatted baseline
• *-2-skill-raw-minified.txt: simulated model output
• *-3-skill-plus-prettier.txt: restored with local Prettier
• summary.json and summary.md: aggregate counts

OpenAI token counts are calculated locally with tiktoken. Character counts are included as a provider-neutral proxy for providers that do not ship a stable offline tokenizer in this repo.

Platform setup

Cursor

1. Run npm run install-skill.
2. The skill installs to .cursor/skills/compress-code-responses/SKILL.md.
3. Ensure the agent runs npm run format -- <files> after edits.

OpenCode

Skill paths, any one of which works:

• .opencode/skills/compress-code-responses/SKILL.md
• .claude/skills/compress-code-responses/SKILL.md
• .agents/skills/compress-code-responses/SKILL.md

Global paths include ~/.config/opencode/skills/ and ~/.claude/skills/.

GitHub Copilot CLI / VS Code / PHPStorm

npm run install-skill writes:

• .github/copilot-instructions.md
• .github/instructions/compress-code-responses.instructions.md

Global Copilot instructions are written to ~/.copilot/copilot-instructions.d/compress-code-responses.md.

Supported formats

Prettier, through this repo's config, supports JS/TS, JSON, YAML, HTML, XML, CSS, Markdown, PHP, PHTML, and related extensions.

Tests

npm test
npm run test:token-comparison

npm test runs the unit tests with explicit file paths for consistent behavior on Windows, macOS, and Linux. npm run test:token-comparison regenerates the benchmark artifacts described above.

Project layout

compress-code-responses/
|-- skills/compress-code-responses/SKILL.md
|-- scripts/
|   |-- ensure-deps.js
|   |-- format.js
|   `-- install-skill.js
|-- platform/
|-- lib/
|-- tests/
|   |-- fixtures/
|   |-- equivalent.test.js
|   |-- minify.test.js
|   `-- token-comparison.js
|-- prettier.config.js
`-- README.md

Limitations

• Prose in chat is not restored by Prettier; savings are largest on code-heavy responses.
• Copilot chat does not run shell commands; you or your agent must run npm run format after accepting edits.
• Minification is instruction-based; models may occasionally send readable code, and the local formatter still applies afterward.

License

MIT