Interactive Notebook Demo

Overview

A demo notebook does two jobs at once, and must do both:

• Orient — it doubles as semi-interactive documentation for a feature. Read top to bottom, it should leave the reader understanding the motivating problem, the design, the mechanism, and the data flow — the narrative layer over the code.
• Interact — a reader can run one operation, read the real output, edit a literal, and re-run, building hands-on understanding by poking at it. The cells are the examples; git holds the original, so they can change anything and revert.

The orientation is what to teach; the interaction is how to teach it. A notebook that interacts beautifully but explains nothing is a toy; one that explains but only runs as opaque blocks is a slideshow. Clear both bars.

Two failure modes make a notebook performative (a script that happens to have cell markers) rather than interactive:

1. Wrapper functions. Grouping operations into async def demo_function(...) and calling them as opaque blocks. The reader can only run the whole block; they cannot step into a single call and tweak it.
2. Terse labels. Printing f"created {name}" instead of showing the evidence — the object the call returned, the before/after, the read-back. The reader has to take the markdown's word for it.

This skill is the bar a demo notebook must clear. It is language-agnostic in principle; examples use Python because most demo APIs are. For asyncio event-loop ownership (top-level await, kernel-owns-the-loop), defer to python-notebooks-async.

When to Use

• Writing a #%% / percent-format or .ipynb notebook that demonstrates a library, API, or service so a maintainer can experiment with what was built.
• Reviewing a demo notebook that "looks like a script," wraps steps in functions, or prints labels instead of proving effects.
• Converting a working script into a genuinely interactive walkthrough.

When Not to Use

• Production pipelines, ETL, or analysis notebooks where reusable logic should live in functions/modules — extract to .py and test it.
• Pure asyncio mechanics (RuntimeError: event loop already running, asyncio.run() in a cell) — see python-notebooks-async.
• One-shot throwaway exploration no one else will read.

Invocation Notice

• Tell the user when this skill is running: interactive-notebook-demo. Skip the notice when the user asked for the skill by name or slash command; spelling and spacing need not match. A phrase from this skill's own trigger list is not a name — naming the work is not naming the skill.

What the demo must orient the reader to

A reader should come away understanding the feature, not just a list of callable methods. Convey each of these through the prose and the evidence — never by labeling a cell with the lens name (see Write it, don't announce it, below):

• The motivating problem. What pain drove this and who it's for — the job it's hired to do. A reader who knows why it exists reads every later cell with the right frame; open with it and keep each section anchored to it.
• The design. The core components, how they layer, and the non-goals — what it deliberately doesn't do. Sequence the notebook to mirror this shape (see narrative arc, below) so the structure itself teaches the architecture.
• The mechanism. What actually happens under the hood for a representative operation — make the invisible visible. Don't just create a record and print "ok"; surface the value the system computed so the reader sees the rule fire.
• The data flow. What moves where, especially across boundaries — input → which layer → storage → response; what crosses a trust or process boundary and what deliberately stays behind. Trace one representative path end to end.

Pull these through as well — they are usually the difference between a tour and real orientation:

• Vocabulary. Define each domain term on first use. A newcomer drowns in unexplained nouns; one sentence where the term first appears is the cheapest, highest-leverage orientation you can give.
• Guarantees and invariants. Make the system's promises observable — a rejected stale write, an unauthorized call refused, a retried write that stays idempotent. This is where prove-with-data earns its keep: don't assert the guarantee, demonstrate it.
• Failure modes. Show how the system says no — its errors made tangible (the error-demo cells). The shape of failure orients as much as the happy path.
• Pointers into the code. Close each section with the real source, spec, or test behind it, so the documentation links back to what it describes rather than dead-ending. This is what keeps the demo anchored to the code it documents.

Core principles (the bar)

1. One claim per cell. Each cell demonstrates one idea, and its output stands on its own. Usually that's one call; group calls only when the claim needs them together (a setup→read pair, a before/after), never to save space. No main(), no multi-call wrapper functions (one exception below).
2. Idempotent cells, or marked. Aim for cells a reader can re-run any number of times with the same outcome, so re-running one never desyncs the others (whether or not they've been run). Reads, lookups, and set-to-a-fixed-value calls are naturally idempotent — leave them unmarked. When the underlying operation can't be — it bumps a version, appends, increments, or has an external side effect — that's fine, but mark the cell with a short, consistent note (# not idempotent: re-running adds a revision). Marking is an expectation, not a judgment call: an unmarked cell is then a trustworthy promise that re-running it is safe. The reset for accumulated drift is the whole-notebook contract — Restart→Run-All must pass. Keep the reader's knob in plain view, not buried in setup().
3. Markdown teaches before each cell. Precede each code cell with a # %% [markdown] cell of 2–4 sentences explaining what it demonstrates and why it matters — teach, don't label. This prose carries the orientation: tie the operation to the motivating problem, surface the mechanism or invariant it shows, define any new term.
4. Write it, don't announce it. Make relevance evident in the writing; never prepend a meta-label. No **Why this exists:**, **What you'll learn:**, or **Story:** prefixes — the first sentence already carries the point. Give each section a heading that states its claim ("## Edits version — they don't overwrite"), not a bare noun ("## Editing"), so the headings read as the story and the notebook is navigable. See antislop-writing.
5. Prove the effect with real data. Every claim the markdown makes must be visible in the cell's output — show the evidence, not a summary:
   • Dump the whole object the call returned when its shape is instructive (e.g. print(result.model_dump_json(indent=2)) for a pydantic model), not one cherry-picked field — but skip dumps so large they bury the point.
   • Show before and after for a mutation, side by side — the values, never a bare True/False.
   • For an invariant, assert it with the values in the message (assert rb == expected, (rb, expected)): the claim is visible and a wrong value fails the kernel run instead of quietly printing False.
   • Read the record back after a write and print what came back, to prove it landed.
6. Interactive-first. Top-level await in cells (the kernel owns the loop). No if __name__ == "__main__" script guard — straight-line python file.py is not a supported mode. Make cleanup its own clearly-marked final cell the reader runs when done.
7. kwargs everywhere. client.update(record_id=rid, data=payload, expected_revision=3) — explicit names aid comprehension and survive signature drift. Verify against the real signatures; naive guesses drift (a parameter you'd call path is really path_prefix, version is really target_version). Read the source before you write the call.
8. try/except only where the error is the point. Keep it for cells whose purpose is to show a failure (a version conflict, an authorization denial, a validation error). Delete defensive try/except Exception around calls that just work — let the reader see the real call and its real result.

Structure

• Narrative arc, not an API index. Order sections as a story that mirrors the design: open with the motivating problem, build from primitives to composed behavior, happy path before edge cases, and let the failure and guarantee cells land after the reader has seen the thing work. The sequence should teach the architecture; never sort cells alphabetically by method name.
• Header cell. Lead with the motivating problem and the mental model (the core nouns and how they relate), then the run mechanics: supported environments (VS Code Interactive / Jupyter), that the kernel handles top-level await, that the reader can edit any input and re-run to explore (git holds the original), the re-run convention (unmarked cells are safe to re-run; # not idempotent cells mutate each run, and Restart→Run-All resets), and any optional dependency the later cells need.
• Setup may stay a function. Connection or bootstrap boilerplate the reader doesn't need to inspect granularly can remain an async def setup(...), invoked in its own cell that unpacks into notebook-scope variables. Everything after setup is flattened to one cell per claim.
• Close each section with where it lives. A short pointer to the real source, spec, or test behind the section, linking the documentation to the code it describes — written as a sentence, not a **Where this lives:** label.
• Zero-infrastructure default. The primary notebook should run with no external services (local or in-memory backends) under a self-cleaning temp dir, so a reader can run it immediately.

Infrastructure-dependent notebooks (remote backends)

For a notebook that needs a database, object store, or other services:

• Probe-first. The first cell is a connectivity probe. On failure it prints actionable guidance (how to start the stack, which ports) with no traceback, and makes clear the later cells need the stack. Subsequent cells may assume the connection succeeded.
• Coexist with the test stack. Ship a demo docker-compose.yaml on non-conflicting host ports so the demo and test stacks run side by side; document the ports/DSNs in the file header.
• Graceful skip, never a wall of red. Absent services → one clear message, not a stack trace.
• README + gitignore. A short README.md (how to open in VS Code/Jupyter, how to start the stack, a port table) and a .gitignore for demo artifacts (local DB files, data/temp dirs).

Validate the real artifact — don't fool yourself

The committed .py percent-file is the artifact. Its top-level awaits are the part most likely to break interactively — so they are exactly what you must execute.

• Do not validate an async-wrapped copy. Wrapping the cells in async def main(): ...; asyncio.run(main()) for a quick python run is a trap: it exercises a different program and leaves the real top-level-await artifact unverified. (This is the mistake to catch in review.)
• Run the actual #%% file through a live kernel via jupytext + ipykernel, then assert zero unexecuted cells and zero error outputs. Never commit the generated .ipynb.
• Exact headless commands, sandbox JUPYTER_* redirects, and the ruff per-file-ignores notebooks need (F704 top-level await, E402, B018, D, S) are in references/headless-validation.md.
• Clean up any scratch files; leave nothing behind.

Common mistakes (review checklist)

• Wrapper-per-topic functions — a whole topic behind one await demo_topic(...) block instead of cell-per-claim.
• Labels instead of evidence — print("done"), or a bare True/False, rather than the returned object, the before/after, the read-back.
• Non-idempotent cell, unmarked — re-running mutates state (version bump, append) with no marker, so the reader can't tell it's unsafe to hammer.
• Meta-labels and flat headings — **Why this exists:** / **Story:** prefixes, or bare-noun headings that don't state a point.
• Operations without orientation — correct cells that never convey why, what they prove, or where the code lives.
• Method-index ordering — sections sorted by API surface, not told as a story.
• Unexplained vocabulary — domain nouns used before they're defined.
• Defensive try/except — swallowing errors around calls that work, hiding real behavior.
• Positional args from memory — guessed names that drift from the real signature.
• Validating a stand-in — an async-wrapped copy or python file.py path; neither runs the real artifact through a kernel.
• Committing the .ipynb — the #%% .py is the source of truth; the executed notebook is a byproduct.

References

• references/headless-validation.md — run the real #%% file through a live kernel, sandbox env setup, ruff per-file-ignores, gitignore/README checklist.
• references/cell-patterns.md — copy-paste cell skeletons: header, section opener, point-into-the-code closer, markdown-then-operation pairing, prove-with-data idioms, the error-demo cell, the connectivity probe, and cleanup.
• references/notebooks-async.md — event-loop ownership, top-level await, and background-task patterns for demos with async cells (shared from python-notebooks-async).