zachdaniel/plz

A (wildly unsafe to use outside of a containerized environment) magical LLM based shell command - zachdaniel/plz

13 stars · 0 forks · 2 watchers · Shell

README

Plz A wildly dangerous piece of shell magic that runs claude in yolo mode to do whatever you want. (yes, this is a shitpost) See the examples below. Anyone who lives out of shell will get why this is cool. Magic Shell Command that integrates well with nushell echo "zachdaniel,josevalim" | plz "get their github usernames and locations" | jq . | plz "capitalize the keys" | where Username == "zachdaniel"

╭───┬────────────┬────────────────╮ │ # │ Username │ Location │ ├───┼────────────┼────────────────┤ │ 0 │ zachdaniel │ Summit, NJ │ ╰───┴────────────┴────────────────╯ Installation Put the single plz script on your PATH. Then wire up your shell so plz becomes pipeline-aware — it can see what you pipe from and to, which is what lets it pick the right output format (e.g. JSON when you pipe into jq). plz init <shell> prints the glue; there are no other files to install. The glue uses your shell's own parser ((z) in zsh, ast in nushell) to locate the running plz within the pipeline — no fragile string munging. If a pipeline contains two identical plz "…" calls with different neighbors (so the correct output format is genuinely ambiguous), plz errors out loudly instead of guessing. plz shells out to the claude CLI, which must be installed and authenticated. Get the script With Homebrew (no separate tap repo — tap this repo by URL and install --HEAD): brew tap zachdaniel/plz https://github.com/zachdaniel/plz brew install --HEAD plz Or just drop the single plz file anywhere on your PATH. Then wire up your shell: Zsh

~/.zshrc

then, in config.nu:

Tip: PLZ_DEBUG=1 <your pipeline> prints the resolved piped_from / piped_to context and exits without calling Claude — handy for confirming pipeline detection.

Examples Tell it to do stuff ls -la | plz "extract the filenames" Output (any shell) README.md plz

Pipe into it It knows what data format is coming in echo '{"users": [{"name": "Alice", "age": 30}, {"name": "Bob", "age": 25}]}' | plz "extract just the names" Output (any shell) Alice Bob

Pipe out of it into anything It also knows what data format it needs to go to (i.e json, because of jq) curl -s https://api.github.com/users/zachdaniel | plz "get the username and location" | jq . Output (any shell) { "username": "zachdaniel", "location": "Summit, NJ" }

It understands nushell's structured data When you pipe a nushell value (a table, record, or list) into plz, the wrapper hands the engine clean JSON plus the value's shape (its columns and types) — instead of nushell's rendered ASCII table — so the model works with named fields. Plain text and byte streams pass straight through untouched (and keep streaming). It also prefers for its final return to be structured data for display in nushell ls -la | plz "extract the filenames" sys | plz "summarize memory and cpu as one line" It can fall back to an LLM for messy data When a regex or parser would be too brittle, the generated command can pipe through plz transform "<instructions>" — an LLM filter that reads stdin and writes only the transformed result. plz reaches for this on its own when it helps (e.g. turning ad-hoc log lines into JSON for jq), but you can use it directly too: tail -n 50 /var/log/app.log | plz transform "turn each line into a JSON object" (Deterministic tools are still preferred when they'll do the job — they're faster and they stream; plz transform reads its whole input.) Put it all together echo "zachdaniel,josevalim" | plz "get their github usernames and locations" | jq . | plz "capitalize the keys" Output (nushell) ╭───┬────────────┬────────────────╮ │ # │ Login │ Location │ ├───┼────────────┼────────────────┤ │ 0 │ zachdaniel │ Summit, NJ │ │ 1 │ josevalim │ Kraków, Poland │ ╰───┴────────────┴────────────────╯

Output (zsh) [ { "Username": "zachdaniel", "Location": "Summit, NJ" }, { "Username": "josevalim", "Location": "Kraków, Poland" } ]

source