Example Apps — Curation Spec

Example Apps — Curation Spec

Status: draft, 2026-07-26. Companion to applications-canonical.md T11. Scope: which of the 36 shipped templates in packages/base-nodes/nodetool/examples/nodetool-base/ get a shipped example app (an ApplicationBundle, T12), and what each app contains. No code.

Today every template gets a generated app from the CURATION table in scripts/generate-template-apps.mjs — 36 variations of one form. This spec replaces that with 11 curated apps binding 23 workflows; the remaining 13 templates ship as workflows only.

Selection criteria

Applied in priority order, recorded per app as C1/C2/C3:

  1. C1 — a real person wants the app, not the graph. A form, dashboard, or gallery is the natural interface for this job.
  2. C2 — it exercises the platform. Several workflows behind one surface, variables carrying state between operations, streaming, live property bindings.
  3. C3 — it runs on a fresh install, or degrades legibly. Keyless, local, or an explicit note naming the key and the cost.

The apps

App Workflows bound Operations Key widgets Criteria Needs
Photo Studio Image Enhance, Photo Enhancement Suite enhance, batch ImageInput, 5 Sliders, Image, Table C1 C2 C3 none for enhance; FAL key for batch
Meeting Room Transcribe Audio, Meeting Transcript Summarizer, Private Assistant transcribe, summarize, ask AudioInput, Markdown, Table, TextInput C1 C2 FAL (Whisper) + OpenAI; ask needs Ollama
Concept Studio Concept Art Iteration Board, Pokemon Maker, Image Enhance concepts, creatures, polish Select, Slider, Image gallery, Button C1 C2 OpenAI + FAL; polish keyless
Research Desk Research Agent, Hacker News Agent, Summarize RSS brief, pulse, digest TextInput, Markdown ×3, Progress C1 C2 OpenAI (+ web search tool)
Ask Your Documents Chat With Your Documents, Private Assistant ask, askLocal TextInput, Switch, Markdown ×2 C1 C2 C3 Ollama (embeddings, askLocal); OpenAI for ask
Brand & Social Brand Asset Generator, Hook & Thumbnail Factory, Social Media Calendar Filler kit, thumbnails, calendar ColorInput, Image, Table, Columns C1 C2 OpenAI + FAL
Product Launch Kit Product Mockup Generator, Product Video Generator mockups, video ImageInput, Slider, Image, Video C1 C2 C3 OpenAI + FAL; Gemini (Veo) for video, cost note
Film Studio Script to Screen, Directed Film to Timeline, Movie Posters produce, cut, poster TextInput, Select, Progress, Video, Image C1 C2 C3 OpenAI + FAL + Gemini (Veo) + Replicate; cost note
Study Buddy Flashcard Generator, Prompt Template cards, explain NumberInput, Table, Markdown C1 C2 OpenAI
Model Arena Model Arena compare TextInput, Columns, Markdown ×3 C1 C2 OpenAI + Anthropic + Gemini keys
Dataset Builder Data Generator generate TextInput, NumberInput, Table C1 OpenAI

Nine of eleven bind more than one workflow. Two are usable with no cloud key (Photo Studio’s enhance, Ask Your Documents in local mode).

Per-app detail

Output names below are the ones in the shipped graphs. Names marked † come from the Output-augmentation table in generate-template-apps.mjs; T12 must bake those Output nodes into the bundled graphs (see Prerequisites).

1. Photo Studio

Generate-free, refine-heavy: the flagship for live property bindings.

  • enhance — Image Enhance. Inputs: image from an ImageInput widget; five node properties bound to sliders — denoise-node.radius, tone-node.brightness, tone-node.contrast, color-node.saturation, sharpen-node.amount, all pace: "release". Policy replace. Output enhanced_image† → display.
  • batch — Photo Enhancement Suite. Inputs: photos (ImageListInput), brightness_adjust / color_boost sliders. Output enhanced_photos† → variable batchResults, rendered as a grid.
  • Variables: sourceImage (instance) shared by both operations so the single photo and the batch use the same drop zone.
  • Degradation: enhance is pure GPU filters (lib.image.*) and runs with no keys; batch adds a FAL ImageToImage grade pass and says so in a note.

2. Meeting Room

Recording in, minutes out, then question the transcript — the transcribe → summarize → interrogate chain across three workflows.

  • transcribe — Transcribe Audio. Input audio from an AudioInput widget, event on change (auto-run on upload). Output transcript† → variable transcript. Policy replace.
  • summarize — Meeting Transcript Summarizer. Input Transcript from variable transcript (the graph’s IsEmpty/If pair already skips its own ASR branch when a transcript is supplied). Outputs Summary → Markdown, Action Items → Table (it emits a dataframe), Transcript unused.
  • ask — Private Assistant. Inputs document from variable transcript, question from a TextInput, tone from a Select. Output answer → Markdown. Local Ollama, so the follow-up Q&A costs nothing and works offline once the transcript exists.
  • Variables: transcript (instance) is the whole point — one value written by op 1 and read by ops 2 and 3.

3. Concept Studio

Generate a gallery, pick one, polish it. The creative iteration loop.

  • concepts — Concept Art Iteration Board. Inputs creative_brief, art_style, mood_keywords (TextInput/Select), variations (Slider 1–8). Output Concept Art → append-disposition gallery.
  • creatures — Pokemon Maker. Inputs animals (TextInput), style (Select fed from the graph’s existing SelectInput options: cel-shaded, 3D render, watercolor, 16-bit, holographic). Output pokemon† → the same gallery.
  • polish — Image Enhance, bound a second time with different mappings: input image from variable picked, sliders as in Photo Studio. Policy replace.
  • Variables: picked (instance) — the gallery selection feeds polish.
  • Both generators use ListGenerator → TextToImage, so results stream in one at a time; a Progress widget shows the batch filling.

4. Research Desk

One topic, three sources, three streamed briefings side by side — the case where a dashboard beats a chat box.

  • brief — Research Agent. Inputs topic (variable topic), audience (Select: engineers / executives / general). Output brief → Markdown, streams as the agent works.
  • pulse — Hacker News Agent. Input topic from the same variable. Output analysis → Markdown.
  • digest — Summarize RSS. Input feed_url (TextInput with a shipped default). Output summary† → Markdown.
  • All three run parallel from one Button, each into its own panel with its own Progress label.

5. Ask Your Documents

RAG with a legible offline fallback.

  • ask — Chat With Your Documents. Inputs question, search, and the three document text areas doc_specs / doc_charging / doc_warranty (shipped with the bundled Aurora sample text). Outputs Answer → Markdown, Retrieved Passages → Markdown (the citation panel).
  • askLocal — Private Assistant. Inputs document from variable pastedDoc, question from the same question variable. Output answer → Markdown.
  • Widgets: a Switch bound to variable localOnly decides which operation the Run button triggers, and visibleWhen swaps the results panel.
  • Note: the vector collection embeds with Ollama nomic-embed-text, so ask needs Ollama running and an OpenAI key; askLocal needs only Ollama. The app note must say this — it is the one place a user can be surprised.

6. Brand & Social

The marketing dashboard: one brand identity drives three deliverables.

  • kit — Brand Asset Generator. Inputs brand_name, brand_description, tagline (TextInputs, written into variables), primary_color (ColorInput). Outputs social_assets → Image grid, brand_brief → Markdown.
  • thumbnails — Hook & Thumbnail Factory. Inputs Video Topic (TextInput), Target Audience from variable audience, Number of Hooks (Slider 3–8). Output thumbnail_gallery → gallery.
  • calendar — Social Media Calendar Filler. Inputs brand_name, target_audience, brand_voice from variables; monthly_theme TextInput; posts_per_week Slider. Outputs content_calendar → Table (dataframe), post_images† → gallery, captions† → Markdown.
  • Variables: brandName, audience, voice (user-scoped, persist: true) — fill them once, every operation reuses them across sessions.

7. Product Launch Kit

Photo in, mockups and a launch video out; the expensive step is opt-in.

  • mockups — Product Mockup Generator. Inputs product_image (ImageInput → variable productPhoto), product_name, product_description, target_audience, num_scenes (Slider 1–6). Outputs mockup† → gallery, scene† → Markdown shot list.
  • video — Product Video Generator. Inputs image_input_1 from variable productPhoto, campaign_brief, target_audience (shared variable), key_features. Output product_video† → Video. Its own Button with a cost note; policy queue, timeoutMs generous — Veo runs are long.
  • Runs mockups on a fresh install without touching Veo, so the app is useful at the FAL tier alone.

8. Film Studio

The showcase run: brief → direction → storyboard → cut → key art.

  • produce — Script to Screen. Inputs Brief (variable brief), Visual Style (Select), Shot Count (Slider 3–8). Outputs direction → Markdown, storyboard → gallery, film → Video.
  • cut — Directed Film to Timeline. Input Brief from the same variable. Output film → Video, presented as “the editable rough cut”.
  • poster — Movie Posters. Inputs Movie Title, Genre, Visual Style (shared Select). Output Poster → Image.
  • Progress widget per operation; policy queue throughout; a standing note that this app spends real money (Veo 3.1 + Replicate MusicGen) and should be run once, deliberately.

9. Study Buddy

Structured data an app renders better than a graph does.

  • cards — Flashcard Generator. Inputs topic (variable topic), num_cards (NumberInput). Output Flashcards → Table, with front/back columns; the graph persists them in SQLite, so re-opening the app shows the stored set.
  • explain — Prompt Template. Inputs Topic from the same variable, Audience (Select). Output Explanation → Markdown, shown beside the cards as “the concept behind this deck”.
  • Variables: topic (user-scoped, persisted) is the study session.

10. Model Arena

Single workflow, but the app is the point: three answers in three columns is unreadable on a canvas.

  • compare — Model Arena. Inputs brief, context (TextInputs). Outputs openai, anthropic, gemini → three Markdown panels inside a Columns layout, each streaming independently with its own Progress label.
  • Degradation: a missing provider key fails one column, not the run. The note names all three keys.

11. Dataset Builder

  • generate — Data Generator. Inputs topic (TextInput), row_count (NumberInput 5–50). Output generated_data → Table.
  • The smallest app in the set, kept because a table view of a dataframe is a genuinely better surface than a Preview node, and it is the reference for the Table widget.

Templates that stay workflow-only

13 templates ship without an app:

Template Reason
Ad Creative Factory Terminal nodes are Previews, not Outputs — nothing for widgets to bind; the video matrix is also the most expensive run in the catalog.
Audio To Image One input, one image; the template card already tells the whole story.
Cold Outreach Co-Pilot Preview-only terminals, and the structured draft fields need an app-side email composer this spec does not define.
Color Boost Video Per-frame grading of an uploaded video is minutes long with one slider — a bad first-run experience; the sliders belong in the editor.
Conditional Logic Engine Teaching example. Its value is the graph.
Image To Audio Story One image in, one audio clip out — a Run button with extra steps.
Image to Video Animation Text-to-image → Veo; it has no image input, so it cannot be the “animate this” step of any gallery app (see Findings).
Movie Trailer Generator Overlaps Film Studio’s produce at higher cost and with a weaker output set.
Music Video Visualizer Strong candidate deferred: one upload, one long paid run, one video. Revisit once a Podcast/Audio app justifies the surface.
Podcast Repurposing Studio Preview-only terminals; needs Output nodes and structured quote-card output before an app can render it.
Research Paper Summarizer Overlaps Research Desk’s brief; its 22-node fetch/filter chain is the interesting part and lives in the editor.
SEO Content Engine Preview-only terminals; the typed article fields want a per-article editor, not a results panel.
Workflow As A Tool Composition demo for graph authors. An app hides exactly what it teaches.

Prerequisites and findings

Things T12 must handle, found while reading the graphs:

  • Output augmentation must move into the graphs. Eight bound workflows have no Output node today and rely on the AUGMENT table in generate-template-apps.mjs: Transcribe Audio, Summarize RSS, Pokemon Maker, Image Enhance, Photo Enhancement Suite, Product Mockup Generator, Product Video Generator, Social Media Calendar Filler (2 of its 3 outputs). Bundles carry graphs, not a generator, so those nodes must be committed to the example JSON.
  • Preview-only templates cannot back an app. Ad Creative Factory, Cold Outreach Co-Pilot, Podcast Repurposing Studio, and SEO Content Engine end in nodetool.workflows.base_node.Preview with zero Outputs. All four are in the workflow-only list for that reason; promoting any of them later is a graph change first.
  • Image to Video Animation has no image input. The obvious “animate the concept I picked” operation for Concept Studio is not buildable against the shipped graph (TextToImage → ImageToVideo). Concept Studio uses Image Enhance for its second step instead; adding an ImageInput to that template is a separate, small change if the animate step is wanted.
  • Chat With Your Documents needs Ollama even in “cloud” mode. Its vector.Collection embeds with nomic-embed-text on Ollama while the answer agent is OpenAI gpt-5-mini. Either the app note says both, or the template switches to a cloud embedding model.
  • Multi-binding is real here. Image Enhance is bound by two apps (Photo Studio, Concept Studio) and Private Assistant by two (Meeting Room, Ask Your Documents) — with different input mappings each time. The bundle format’s workflows[].key indirection covers this; the installer must not create duplicate workflow rows when both bundles are installed.
  • Marketing impact. /apps/* drops from 36 generated pages to 11 curated ones. Every retired slug needs a redirect to the app that absorbed it (or to its /templates/<slug> page, which survives) — the redirect map is T13 work, but the mapping is this table.
Edit this page on GitHub