by samuelfaj
Distill large CLI outputs into small answers for LLMs and save tokens!
# Add to your Claude Code skills
git clone https://github.com/samuelfaj/distillAgent command outputs are one of the biggest sources of token waste.
Logs, test results, stack traces… thousands of tokens sent to an LLM just to answer a simple question.
🔥 distill compresses command outputs into only what the LLM actually needs.
Save up to 99% of tokens without losing the signal.
npm i -g @samuelfaj/distill
You can also point distill at OpenAI-compatible providers such as LM Studio, Jan, LocalAI, vLLM, SGLang, llama.cpp-compatible servers, MLX-based servers, and Docker Model Runner.
Add in your global agent instructions file:
CRITICAL: Pipe every non-interactive shell command through `distill` unless raw output is explicitly required.
CRITICAL: Your prompt to `distill` must be fully explicit. State exactly what you want to know and exactly what the output must contain. If you want only filenames, say `Return only the filenames.` If you want JSON, say `Return valid JSON only.` Do not ask vague questions.
Bad:
- `distill "Which files are shown?"`
Good:
- `distill "Which files are shown? Return only the filenames."`
Examples:
- `bun test 2>&1 | distill "Did the tests pass? Return only: PASS or FAIL, followed by failing test names if any."`
- `git diff 2>&1 | distill "What changed? Return only the files changed and a one-line summary for each file."`
- `terraform plan 2>&1 | distill "Is this safe? Return only: SAFE, REVIEW, or UNSAFE, followed by the exact risky changes."`
- `npm audit 2>&1 | distill "Extract the vulnerabilities. Return valid JSON only."`
- `rg -n "TODO|FIXME" . 2>&1 | distill "List files containing TODO or FIXME. Return only file paths, one per line."`
- `ls -la 2>&1 | distill "Which files are shown? Return only the filenames."`
You may skip `distill` only in these cases:
- Exact uncompressed output is required.
- Using `distill` would break an interactive or TUI workflow.
CRITICAL: Wait for `distill` to finish before continuing.
No comments yet. Be the first to share your thoughts!
logs | distill "summarize errors"
git diff | distill "what changed?"
terraform plan 2>&1 | distill "is this safe?"
Examples with other providers:
distill config provider lmstudio
distill config model "your-loaded-model"
distill config provider jan
distill config api-key "secret-key-123"
distill --provider localai --host http://127.0.0.1:8080/v1 "summarize errors"
distill --provider docker-model-runner --model ai/llama3.2 "what failed?"
distill --provider openai-compatible --host http://127.0.0.1:9000/v1 "summarize warnings"
You can persist defaults locally:
distill config model "qwen3.5:2b"
distill config timeout-ms 90000
distill config thinking false
distill config provider lmstudio
distill config host http://127.0.0.1:1234/v1
Supported providers:
ollamaopenaiopenai-compatiblelmstudiojanlocalaivllmsglangllama.cppmlx-lmdocker-model-runnerFor pipeline exit mirroring, use pipefail in your shell:
set -o pipefail
Interactive prompts are passed through when distill detects simple prompt patterns like [y/N] or password:.
If you want Codex, Claude Code, or OpenCode to prefer distill whenever they run a command whose output will be sent to a paid LLM, add a global instruction telling the agent to pipe command output through distill.
~/.codex/AGENTS.md.~/.claude/settings.json, and its official mechanism for custom behavior is global instructions via CLAUDE.md.~/.config/opencode/opencode.json. Point its instructions field at a markdown file with the same rule.~/.copilot/copilot-instructions.md.rg -n "terminal|PERMISSION|permission|Permissions|Plan|full access|default" desktop --glob '!**/node_modules/**' | distill "find where terminal and permission UI are implemented in chat screen"
🔥 Saved ~98.7% tokens