Examples Revamp Plan

Status: draft, 2026-07-17. Scope: the 37 shipped templates in packages/base-nodes/nodetool/examples/nodetool-base/ and their generated mini apps (scripts/generate-template-apps.mjs, web/public/app-preview/).

Why

The examples are the product’s first impression three times over: the template gallery in the editor, the mini apps on /apps/* marketing pages, and the first thing a new user runs. Today they undersell on all three surfaces. The workflows are mostly linear prompt pipes that don’t justify a node editor, and the generated apps are 37 copies of the same form. Meanwhile the platform just gained the features that would make them compelling — node-property bindings, reactive subgraph runs, run pacing, container-responsive app layout — and no example uses any of them.

Analysis

Measured across all 37 templates (graph JSON + generated app_doc):

Workflow shape

15 of 37 graphs are purely linear (max fan-out 1); median size ~7 nodes. The dominant shape is Input → Prompt template → LLM → Output — a prompt wrapper that a chat box does better.
Control flow is nearly absent: 14 nodetool.control nodes and 6 nodetool.boolean nodes across the whole set, almost all in one teaching example (Conditional Logic Engine).
Node namespace census: nodetool.input 72, nodetool.text 60, nodetool.agents 24, nodetool.generators 19, image 34, video 11, lib.sqlite 3. No vectorstore/RAG usage, no workflow-as-tool composition, no streaming showcase.

Models

gpt-5-mini appears 41 times — effectively every text node. Images are all fal-ai/flux/schnell (16). Video is veo-3.1 (5, expensive).
Providers: openai 43, fal 22, gemini 6, claude_agent_sdk 1. Zero local models (Ollama/MLX). A fresh install without an OpenAI key has ~30 broken examples and no private/offline example at all.

Content hygiene

Shipped prompt text includes typos and one-liners (“Create a 3 compelling Pokémons”). No system prompts, no output-format instructions. Prompts are the actual product of these workflows and read as first drafts.
Hardcoded content that should be inputs: RSS feed URL (BBC Europe), search topics, folder paths. FolderPathInput appears in two templates and is meaningless in a hosted app.
Node titles/descriptions are sparse; comments decorate rather than teach.

Outputs

30 of 37 templates have exactly one Output; nearly everything funnels into a text/markdown blob. Flashcards, calendars, SEO fields, and generated data are prose instead of structured objects — which caps what the apps can render (widget census over all 37 apps: 28 Markdown + 37 Text vs 13 Image, 7 Video, 1 Audio, 1 Json).

Apps

All 37 app docs share one generated shape: heading, tagline, input form, Run button, results panel. Zero Slider/Select/Switch widgets anywhere (71 WorkflowInput widgets, 37 Buttons).
8 templates have no inputs at all — their app is a Run button (Agent Google Search, Color Boost Video, Data Generator, Fetch Papers, Hacker News Agent, Image To Audio Story, Summarize RSS, YouTube Research Agent).
Apps open with empty fields (seeded demo values exist only in the marketing preview bundles), and none use pacing, node-property bindings, variables, or an iteration loop (regenerate/variants/gallery).

Marketing surface

The examples feed the marketing site directly: every template gets a /apps/<slug> landing page (marketing/scripts/generate-miniapp-entries.mjs → miniAppEntries.generated.ts, 37 pages live), a screenshot (web/scripts/screenshot-app-previews.mjs → marketing/public/apps/<slug>.png, 37 present), and a /templates/<slug> page via the generated template catalog. Consequences of the current state:

All 37 screenshots show the same empty form layout — a grid of visually identical pages that read as one product with 37 names. Conversion surface, currently underselling.
The 8 no-input apps are marketed as apps while being a Run button.
Landing pages exist for near-duplicate templates, splitting search intent (three YouTube/research agents, two thumbnail pipelines, three audio summarizers) instead of concentrating it on one strong page each.

Root cause

The generator (generate-template-apps.mjs) is sound infrastructure with a timid curation table, and the workflows were authored to demo single nodes, not to be products. Both are fixable without new platform features — the platform is now ahead of its own examples.

Goals

Every shipped example runs on first try and passes nodetool validate.
Every example demonstrates at least one thing a chat box can’t do (parallel comparison, media pipeline, control flow, RAG, live parameters, structured data).
A keyless install has at least 3 fully working local examples.
The featured apps feel like tools: seeded defaults, real controls, visible progress, an iteration loop.
Fewer, better: target ~20 curated examples instead of 37 near-duplicates.

Plan

Phase 0 — Guardrails (do first, half a day)

CI job: run nodetool validate over every example JSON; --warnings-as-errors for unknown types, dangling edges, unselected models. This is the rot detector for everything that follows.
Add a cheap smoke tier: nodetool debug on 3–4 keyless/local examples. Skip paid-model examples in CI; run them in a manual pre-release pass.

Phase 1 — Hygiene pass on all survivors (1–2 days)

For each kept template:

Rewrite prompts properly: system prompt, task structure, explicit output format, no typos. Prompts are the product; treat them like code review.
Parametrize hardcoded values into inputs (feed URL, topic, style).
Name and describe every node; make comments explain the why of the graph.
Set sensible input defaults so Run works with zero typing.
Model pass: keep a cheap fast default for text, but vary providers deliberately where it shows breadth; flag expensive nodes (veo) in the description with a cost note.

Phase 2 — Restructure the catalog (2–3 days)

Disposition of all 37 (keep = hygiene only; fix = keep + targeted change; merge/retire = fold into another or drop):

Template	Disposition	Change
Ad Creative Factory	keep	showcase; structured variant outputs
Agent Google Search	merge	into Research Agent (topic input, streaming)
Audio To Image	keep	starter; seed demo audio
Brand Asset Generator	keep	hero app
Cold Outreach Co-Pilot	fix	structured outputs (draft objects)
Color Boost Video	fix	VideoInput + slider-bound intensity (live grading)
Concept Art Iteration Board	keep	hero; variant gallery loop
Conditional Logic Engine	keep	teaching example for control flow
Creative Story Ideas	merge	into one “Prompt Template” starter
Data Generator	fix	schema input → dataframe output → table app
Fetch Papers	merge	into Research Paper Summarizer (topic input)
Flashcard Generator	fix	emit card objects, card-flip app rendering
Hacker News Agent	fix	topic input, streaming markdown
Hook & Thumbnail Factory	keep	hero (absorbs YouTube Thumbnail Pipeline)
Image Enhance	fix	FLAGSHIP: sliders bound to node properties, pace: release
Image To Audio Story	fix	add ImageInput
Image to Video Animation	keep
Learning Path Generator	merge	into Prompt Template starter
Meeting Transcript Summarizer	fix	action items as dataframe (absorbs Summarize Audio)
Movie Posters	keep	hero
Movie Trailer Generator	keep	showcase; cost note
Music Video Visualizer	keep	showcase
Photo Enhancement Suite	fix	replace FolderPathInput (multi-image) or mark local-only
Podcast Repurposing Studio	keep	showcase
Pokemon Maker	fix	hero; rewrite prompt, style Select, variant gallery
Product Mockup Generator	keep	hero
Product Video Generator	keep
Research Paper Summarizer	fix	absorbs Fetch Papers
SEO Content Engine	fix	typed field outputs
Social Media Calendar Filler	fix	calendar as dataframe
Story to Video Generator	retire	overlaps Movie Trailer; FolderPath-bound
Summarize Audio	merge	into Meeting Transcript Summarizer
Summarize RSS	fix	feed URL input; digest by topic
Transcribe Audio	fix	instant tool: auto-run on upload (change event, paced)
Wikipedia Agent	merge	into Research Agent
YouTube Research Agent	merge	into Research Agent
YouTube Thumbnail Pipeline	merge	into Hook & Thumbnail Factory

Net: 37 → ~24, minus retire → ~23, plus new below → ~28; prune to ~20 at final curation if the merged families hold up.

Marketing follow-through for every merge/retire: keep the old /apps/<slug> and /templates/<slug> URLs as redirects to the absorbing template’s page (the slugs are already indexed), and grep docs/tutorials for retired template names in the same pass.

Phase 3 — New differentiator examples (2–3 days)

Each maps to a marketing story the catalog currently can’t tell:

Chat With Your Documents — RAG: index a folder into the vectorstore, ask questions with cited answers. (First vectorstore usage in examples.)
Research Agent — the merged agent family: topic input, real tool use (search + browse), streaming markdown brief with citations.
Model Arena — one brief fanned out to 3 providers side-by-side. Earns the graph visually and doubles as a great comparison app.
Private Assistant (local) — Ollama/MLX end-to-end, no API keys. Along with 2 other local-capable starters, satisfies the keyless goal.
Workflow as a Tool — a workflow exposed via tool_name, consumed by an agent in another workflow. Shows composition, which nothing demos.

Phase 4 — App layer (2 days, extends the curation table)

Extend generate-template-apps.mjs curation entries with per-template hints instead of hand-editing app docs:

sliders: node-property bindings with ranges (autofill already derives min/max/step from metadata), pace: "release" for anything expensive.
selects: finite choices (aspect ratio, style, tone, platform) instead of free-text inputs.
seeds: shipped default values in the app itself, not just the marketing preview — first Run must work with zero typing.
gallery: append-disposition outputs rendered as a grid with a “Make another” button for the creative heroes.
featured: flag the ~8 hero apps for the gallery and /apps landing pages (Image Enhance, Pokemon Maker, Movie Posters, Brand Asset Generator, Hook & Thumbnail Factory, Product Mockup Generator, Concept Art Iteration Board, Music Video Visualizer).
Progress widget on every template whose run exceeds a few seconds.

Regenerate previews and /apps pages (screenshot-app-previews.mjs, marketing/scripts/generate-miniapp-entries.mjs) as the last step.

Phase 4b — Marketing page impact

What the revamp changes on the marketing site, and the work it implies:

Screenshots become the sell. Today’s 37 near-identical empty-form screenshots become ~20 distinct ones showing seeded results: a photo mid-enhancement with sliders, a Pokémon variant gallery, a model-arena comparison, a calendar table. The screenshot pipeline needs seeded values rendered (the preview bundles already seed; verify the new gallery/slider states render before the screenshot settles).
Fewer, stronger landing pages. ~20 pages that each answer one search intent instead of 37 splitting it. Merged/retired slugs redirect (see Phase 2); the /apps index orders by the featured flag with heroes first.
New pages with real search demand. The Phase 3 examples add landing pages for terms the site can’t currently target: “chat with your documents / RAG”, “compare AI models side by side”, “local private AI assistant”, “AI research agent”. These are stronger queries than most existing template names.
Copy pass per page. Landing copy currently describes the workflow; rewrite to describe the outcome and name the differentiator (“drag the slider, watch the photo update — runs the graph live”), reusing each template’s new description field so app, gallery, and landing page say the same thing.
Mechanics. All of this flows through the existing generators — the work is the redirects map, the featured ordering, and the copy fields in the curation table; no new marketing infrastructure. Entries without a fresh screenshot stay noindexed by the existing rule, which safely gates half-finished templates out of search.

Phase 5 — Verification and release

nodetool validate green across the catalog (CI from Phase 0).
nodetool app debug per hero app: bindings resolve, run trigger exists, display widgets receive values.
Manual pass on paid-model examples before release; screenshot refresh.

Sequencing and effort

Phases 0–1 are independent of 2–4 and land first (~2 days, pure win, no decisions needed). Phase 2 dispositions need a maintainer sign-off on the merge/retire list, then 2–4 parallelize by template family. Total ~7–9 working days; the flagship (Image Enhance live editor) is worth pulling forward as the proof piece for the whole direction.

Risks

Cost: veo/fal examples are expensive to smoke-test. Mitigation: CI validates but doesn’t execute paid models; manual pre-release pass only.
Model drift: pinned model ids rot. Mitigation: Phase 0 CI catches unknown models at validate time; prefer provider-default aliases where the registry supports them.
User familiarity: retired/merged templates may be referenced in docs and tutorials. Mitigation: grep docs for template names during Phase 2; keep slugs redirecting on the marketing site.
Scope creep: structured outputs may reveal missing table/gallery rendering in app widgets. The Json/Output widgets cover the minimum; a dedicated Table widget is a fast follow, not a blocker.

Success criteria

Fresh install, no keys: 3+ examples run end-to-end locally.
Every example passes validate in CI; hero apps pass app debug.
Every kept example demonstrates a named differentiator (recorded in its description).
Featured apps: seeded first run, at least one real control (slider/select), progress on long runs, and an iteration loop on the creative ones.
Marketing: no dead /apps or /templates URLs (redirects in place), every indexed page has a distinct seeded-result screenshot, heroes lead the /apps index.

Edit this page on GitHub