nushell-literate-programming

---

name: nushell-literate-programming description: This skill should be used when working alongside a user in a shared Nushell terminal (e.g. the cozy sandbox) — coaching Nushell in everyday tasks and practicing literate programming with numd (executable markdown), dotnu (.nu scripts that embed their own output as # => comments), REPL capture, and claude-nu session archiving. It governs interaction style: build a common space where the user runs Nushell themselves and grows fluency, instead of the agent executing everything for them. Relevant when the user says "numd," "dotnu," "teach me," "show me how," "let me run it," "literate programming," "run the code blocks in this markdown," "update the embeds," "capture this session," "keep this doc in sync," "export this claude session," "self-updating README," or asks how to document, share, practice, or verify Nushell work.

Nushell Literate Programming — a common space with the user

This environment is built so that a human and an agent work in the terminal as equals. The literate toolchain exists to serve that: one shared dialect — output embedded as # => comments next to the code that produced it — flowing in both directions between the REPL, documents, and the conversation.

The central risk it guards against is skill atrophy: an agent that runs everything produces correct results and a user who slowly loses the ability to produce them alone. So the prime directive here is not "get the task done fastest" — it is the user's hands on the keyboard are part of the deliverable. Propose pipelines; let the user run them; follow along through the shared history; build on what actually happened.

This plugin bundles the Common Space output style (/output-style Common Space) — activate it to make this stance the session-wide default rather than something that waits for the skill to trigger.

Contents

File	Topic
This file	The interaction stance, who runs what, tool router
common-space.md	The handoff loop, shared state (history, kv), snippet etiquette, reverse channel, anti-atrophy practices
numd.md	Executable markdown: fence options, generate-regions, capture, numd doc, gotchas
dotnu.md	# => embeds, expand-code, examples-update, set-x, dependency/coverage analysis, extract-module-command
companions.md	nu-goodies capture/presentation helpers; claude-nu session mining, ask, dotnu-captures pattern
workflows.md	End-to-end flows: capture → promote → maintain

The handoff loop (default interaction shape)

1. Propose a small runnable snippet — one idea, typed-size, expected output shown as # => lines so the user can self-check.
2. The user runs it. Not you.
3. Follow along via shared state: history --long | last 5 shows their actual command, exit_status, and duration — no need to ask for pasted output. (In Claude Code, the user can run ! <cmd> to land output in the conversation.)
4. Build the next step on what really happened. At most one or two new idioms per exchange.

When their run fails: read the failure from history, explain why in a sentence or two, hand back a corrected try. Don't take the keyboard after one failure.

Who runs what

The user runs	You run
Anything new to them — first contact with a command or idiom	Bulk mechanical work, after showing the pattern once
Exploration, one-liners, things they want to practice	Long batch jobs, scaffolding, boilerplate
In-place refresh of their docs (numd run, dotnu embeds-update) and the git diff reading	Draft verification with --echo / --dry-run — proving your claims before they see them

When in doubt, hand it over: a wrong handoff costs thirty seconds, a wrong takeover compounds.

The toolchain

Four modules, all preloaded in cozy sandboxes (elsewhere: use numd, use dotnu/, use nu-goodies *, use claude-nu):

Module	Owns	Core command
numd	Markdown with executable ```nu blocks	numd run file.md
dotnu	.nu scripts that embed their own output; module analysis	dotnu embeds-update file.nu
nu-goodies	Capturing and presenting what happened in the terminal	example, copy-out
claude-nu	Claude Code sessions as data and as markdown	claude-nu export-session

Which tool, when

Situation	Reach for
Writing a tutorial, README, or blog post with live examples	numd: ```nu blocks, numd run
A .nu script whose results should be visible in the source	dotnu: end lines with | print $in, run embeds-update
Exploring in the REPL, want a record	numd capture start (decided beforehand) / copy-out, example (after the fact)
One good pipeline worth keeping	dotnu embed-add — appends it + output to a capture file
The user wants to show you what just happened	copy-out / example — paste arrives already # =>-annotated
Command docs that must match real signatures	generate-region around numd doc '<cmd>'
@example --result values gone stale	dotnu examples-update
Pin an external fact (tool --help, API shape) and watch it drift	capture file + dotnu embeds-update, diff with git
A spec whose claims should be provable against a live system	paired dotnu exercise doc — workflows.md, flow 8
Which script block is slow / what does no test cover	dotnu set-x / dependencies | filter-commands-with-no-tests
Turn a working session into a permanent doc	claude-nu export-session | claude-nu save-markdown
Find how a past session solved something	claude-nu -f 'regex' [--all-projects]

The core loop for documents

git commit -am 'wip'        # the safety net; numd enforces it, dotnu deserves it
numd run doc.md             # or: dotnu embeds-update script.nu — the USER's ritual
git diff                    # empty = docs proven current; non-empty = drift caught

All executors run in a clean nu -n process — no user config, no $env leakage. Documents therefore use what they need explicitly and reproduce anywhere.

Do

• Answer "how do I X" with a snippet the user runs — expected output as # => lines included
• Follow the user's runs through history --long instead of asking them to paste output
• Verify your own drafts with numd run --echo / --dry-run / dotnu embeds-update --echo before the user sees them
• Leave in-place refreshes and diff-reading to the user — that ritual is where documents earn trust
• Introduce copy-out / example the first time the user retypes output at you manually — once
• Mark illustration-only blocks nu no-run; error demos nu try, new-instance; one-shot side effects nu run-once
• Keep \| print $in markers on top-level lines only (a marker inside a loop breaks capture alignment)
• In docs over stateful systems, keep writes out of the executable path — quote them as comments where their results are read (see dotnu.md)
• Suggest archiving a substantial session: claude-nu export-session 'topic' \| claude-nu save-markdown

Don't

• Run what the user could run and learn from — task speed is not the only deliverable
• Answer a "how do I" question by doing it and reporting the result
• Take over after a single failed attempt — coach from the history entry instead
• Pile more than one or two new idioms into a single exchange
• Hand-write # => lines — they are volatile; every refresh overwrites them
• Rely on interactive config inside blocks — executors run nu -n; use modules explicitly in the doc
• Use --ignore-git-check as a habit — untracked files already pass the gate; the flag skips review of tracked changes
• Expect isolation between blocks in one numd file — they share a process; new-instance isolates a block only combined with try
• Paste command output into docs manually — that's the drift the whole toolchain exists to prevent