Source: raw/Build_BEAUTIFUL_Diagrams_with_Claude_+Excalidraw_Full_Workflow.md, raw/Claude_Just_Made_NotebookLM_10x_More_Powerful_NEW_Skill.md, raw/Claude+Canva=_Design_on_Autopilot_Beginners_Course.md, raw/Someone_Just_Rebuilt_Adobe_for_Free_and_Claude_Can_Run_Them_All_7_apps_tested.md
Creator: Robo Nuggets (host Jay) | Platform: YouTube
- Excalidraw: https://www.youtube.com/watch?v=Gf-yFsbxgyo | Duration: 13:48 | Published: 2026-04-10
- NotebookLM: https://www.youtube.com/watch?v=7sInxhTDA7U | Duration: 15:35 | Published: 2026-04-07
- Canva + GPT Image 2: https://www.youtube.com/watch?v=fT3rsZ22C4A | Duration: 14:19 | Published: 2026-04-29
- ArtCraft: https://www.youtube.com/watch?v=Fuo1i_-9Frc | Duration: 8:39 | Published: 2026-10-08
Four Robo Nuggets tutorials that follow one pattern: give Claude Code a skill or CLI for an outside creative tool, test the connection, iterate on the output, then save your taste back into the skill and optionally put the job on a schedule. The tools are Excalidraw (diagrams), Google NotebookLM (research decks, podcasts, mind maps), GPT Image 2 with Canva Magic Layers (posters), and ArtCraft’s open-source Adobe-style apps. The three April videos predate the September 2026 Chat and Cowork merge, so their desktop-app walkthroughs (three tabs: Chat, Cowork, Code) are historical; the integration steps still apply.
Key Takeaways
- The shared loop. (1) Install the tool’s skill, MCP server or CLI. (2) Run a cheap test (“list my last three notebooks”, “open the canvas”). (3) Ask for several variations, because more volume gives you more to choose from. (4) Give feedback. (5) Ask Claude to update the skill and to “ask me numbered questions so the intent is clear”. (6) Schedule it. Jay uses step 5 in all three April videos.
- Agent-drivable tools beat screenshots. Each tool here exposes something the agent can call and inspect: a live canvas it can screenshot, a CLI, or an MCP server. That is what makes the iterate-and-check loop possible.
- Delivery to your phone. Jay runs long-lived Claude Code sessions connected to Telegram (see Channels), so scheduled diagrams and slide decks arrive there.
- Excalidraw gives hand-drawn diagrams on a live local canvas Claude can see and refine.
- NotebookLM builds the notebook, loads the sources and generates the deck, podcast or mind map, using Google’s tool for the grounded reading. Google renamed NotebookLM to Gemini Notebook in July 2026, according to the
notebooklm-pyREADME. - Canva + GPT Image 2 produces finished posters in one shot. Canva Magic Layers then splits the flat image into editable layers for small fixes.
- ArtCraft apps give layer-by-layer control. Jay’s argument: AI generation gives a first draft that is hard to polish, while a layered app gives an “okay first draft” that is much easier to refine.
Excalidraw: diagrams Claude can see
- What it uses. The open-source
yctimlin/mcp_excalidrawproject (MIT; 2,575 stars on 2026-10-27; Jay said “a couple of hundred” in April). It runs a local canvas server athttp://127.0.0.1:3000and offers a CLI plus agent skill, an MCP server (26 tools) and a REST API. - Why not the official Excalidraw MCP. The project’s README says the official MCP is a one-shot chat widget, while this one gives element-level create/read/update/delete,
describeandscreenshotso the agent can see the canvas, plus layout tools,.excalidrawexport, snapshots and Mermaid conversion. - What a run looks like. Claude reads the skill’s design guide, clears the canvas, plans the layout, draws, zooms to fit, then screenshots its own work to check it.
- Iterate on taste. Jay’s first diagram was too technical. His follow-up asked for plain language, more visual elements, varied font styling and three variations. He then asked for text wrapped inside boxes and everything centred. His saved preferences were plain language, multiple options every time, emoji icons and numbered badge circles, no text overflow, and an orange/gray/green palette.
- Export. PNG or other image formats, or a
.excalidrawfile you can open and edit later. Dark mode is a canvas toggle. - Daily diagram. He asked Claude Code in plain language to make a diagram every day at 12:00 Sydney time explaining one new AI development, export it as an image and send it to Telegram. Claude registered a scheduled task and suggested a test run first.
NotebookLM: research decks without opening the app
- What it uses.
teng-lin/notebooklm-py(MIT; 19,943 stars on 2026-10-27), an unofficial Python API, CLI and agent skill. Jay credits the developer and wraps it in his own skill, which adds slide-style guidance. - Setup. The skill installs the package, then a browser window opens for Google sign-in and the session is saved. Verify it with “list the latest three notebooks I have”. The current README recommends
uv tool install "notebooklm-py[browser]"thennotebooklm login. - What Claude can do with it (per Jay’s capability listing): create, list, rename and delete notebooks; add web, YouTube and local-file sources; chat with a notebook; generate audio overviews (podcast MP3s), video overviews, slide decks, mind maps, reports, quizzes, flashcards and infographics.
- Branded slides. Hand Claude a markdown script and say “turn this into slides with your NotebookLM skill”. The skill writes the slide prompt for you, specifying a seven-slide presenter deck, titles, design and font. To change the look, ask Claude to edit the skill (“more blue tones, still dark mode, more corporate”) or give it brand-book images, then save the result as a second named style.
- Scheduling (April 2026 UI, historical). He used the desktop Schedule tab (“cybersecurity research daily”, noon Sydney time). He noted that a local task only runs while the computer is on, and that the then-new remote task option needed a GitHub-hosted workspace. Today’s equivalents are covered in Scheduled Tasks and Routines.
- Risk. The README warns that the library uses undocumented Google APIs that can change or be rate-limited without notice.
Canva + GPT Image 2: posters, then editable layers
- Generation path. Jay’s poster skill calls OpenAI’s GPT Image 2 through fal.ai (captions say “fall”). He prefers aggregators like fal.ai over subscription platforms such as Higgsfield because you top up only what you use. The skill also holds a starter library of poster styles.
- Brief to three directions. Paste the client brief and ask for three styles: two from the skill’s library and one that isn’t in it, so you don’t get locked into the templates. Attach the client logo by file path, and generate at low resolution first.
- Parallel briefs. In the VS Code extension, fork the conversation and ask Claude to spin up two subagents, one per brief, each making one library-style design and one new design end to end. Jay notes that doing it step by step gives better quality.
- Text accuracy. Jay says GPT Image 2 renders text correctly “98 to 99% of the time”; that is his estimate. His three test posters had no typos he could find.
- Resize and upscale. Ask for the highest resolution plus landscape and square versions, and have Claude summarise the planned dimensions before generating. Claude flagged GPT Image 2’s per-side pixel ceiling and suggested ESRGAN or Topaz upscalers on fal.ai for A2 print.
- Canva Magic Layers. Upload the flat poster at Canva’s Magic Layers page (free to try, per Jay). It separates text and elements into editable layers and detected the font. Edit the copy, download, and give the file back to Claude for resizing.
- Bulk variants. Give Claude one approved reference design and a folder of photos. Ask it to show the prompt pattern before running. In his example it kept the layout, typography, logo and green palette and swapped the hero photo, then produced 10 real-estate posters.
ArtCraft: open-source Adobe-style apps an agent can drive
- What it is. ArtCraft (
github.com/storytold, apps atgetartcraft.com/apps) released clean-room reimplementations of creative apps in pure Rust. The seven in Jay’s video: PhotoCraft (Photoshop), FilmCraft (Premiere; captions say “Filmora Craft”), EffectCraft (After Effects), VectorCraft (Illustrator), DesignCraft (InDesign), LightCraft (Lightroom) and PdfCraft (Acrobat). - More apps have shipped since. The GitHub org listed WordCraft (Word), GridCraft (Excel), DeckCraft (PowerPoint), CADCraft (AutoCAD-style), SoundCraft (Pro Tools), FigCraft (collaborative interface design) and AniCraft (2D animation) as of 2026-10-27.
- Stars on 2026-10-27: PhotoCraft 73.9k, PdfCraft 18.2k, FilmCraft 16.6k, LightCraft 14.9k, VectorCraft 10.9k, EffectCraft 8.0k, DesignCraft 4.9k. Jay reported 20k for PhotoCraft a few hours after launch.
- Status. The PhotoCraft README labels it early alpha and says it is clean-room, “implemented from public specs and observed behaviour only”.
- How Claude drives it. In PhotoCraft every menu item and tool runs a command from one registry of 500+ commands, shared by the UI, a CLI (
photocraft-cli run,batch), a JSON control channel and an MCP server (photocraft-cli mcp). Each repo ships anAGENTS.mdhandbook. Jay’s install route: send your agent the repo link and let it set things up. - Jay’s tests. One prompt in EffectCraft produced a logo animation with editable layers. A DesignCraft magazine-spread layout was built by Claude. He rates both as usable first drafts, not finished work.
- Where it fits. For short logo-style animations, video models (Kling, Seedance) or HTML animation with Opus 5.5 are already enough. ArtCraft is the better choice for longer, layered production work where you need fine control.
Try It
- Excalidraw: install
mcp_excalidraw(npx -y mcp-excalidraw-server), openlocalhost:3000, and ask Claude Code to diagram one of your workflows in three variations. Pick one, give two corrections, then say “update the skill with what I liked; ask me numbered questions first.” - NotebookLM:
uv tool install "notebooklm-py[browser]",notebooklm login, then ask Claude to “create a notebook on [topic], load 10 web and YouTube sources, and generate a slide deck and an audio overview.” - Posters: have Claude generate three low-res poster directions from a real brief through fal.ai GPT Image 2. Take the winner through Canva Magic Layers for one copy edit, then ask for square and landscape versions.
- ArtCraft: point Claude Code at the
storytold/photocraftrepo, have it readAGENTS.md, and run one headlessphotocraft-clibatch edit on a folder of images.
Implementation
Tool/Service: yctimlin/mcp_excalidraw; teng-lin/notebooklm-py; GPT Image 2 via fal.ai plus Canva Magic Layers; ArtCraft apps (storytold/*).
Setup: Excalidraw needs Node 20+ and runs locally with no API keys (per its README). notebooklm-py downloads Chromium (~170 MB) on first login (per its README). fal.ai needs an API key with credit. ArtCraft apps build with cargo or download from getartcraft.com/apps.
Cost: Excalidraw MCP, notebooklm-py and ArtCraft are free and open source (MIT; MIT; Apache-2.0 or MIT). NotebookLM was free to try at recording time. fal.ai is pay per generation. Canva Magic Layers is free to try per Jay; plan limits weren’t covered.
Integration notes: Jay’s own skills for Excalidraw, NotebookLM and posters are distributed through his community; the open-source projects underneath work without them. Cloud routines can’t attach MCP servers configured locally in Claude Code (v2.1.251 note in Scheduled Tasks), so a scheduled Excalidraw job that relies on the local canvas likely needs a desktop or local scheduled task.
Open Questions
- Whether ArtCraft used Opus 5.5 to write the apps is Jay’s claim; the PhotoCraft README doesn’t mention Claude or Opus. His statement that the clean-room approach makes them “perfectly legal” is his opinion, not legal advice.
- Whether notebooklm-py still works the same after the Gemini Notebook rebrand isn’t tested here; the README says it does.
- How Canva Magic Layers handles posters with several photos wasn’t shown; Jay only demoed one-photo posters.
- How often the April Excalidraw and NotebookLM flows break as those upstream projects ship new versions isn’t known.
Related
- Claude Design Power-User Tricks — same creator’s design tips
- CLI vs MCP Tool Selection — why CLI-plus-AGENTS.md tools like ArtCraft suit agents
- Skill Design Patterns — the skill-refinement loop used throughout
- Claude Code Scheduled Tasks — current scheduling options
- Channels — the Telegram delivery path
- GPT Image 2 launch — the poster model
- Reverse-Engineering Software With Coding Agents — wider context on agent-built software clones
- HyperFrames — the code-first alternative for short animations
- NotebookLM for Marketers (Grace Leung) — NotebookLM used directly for marketing work