Obsidian Excalidraw CLI: GPT-6 Luna vs 6.1 Sol  🔗 **Source:** [Obsidian Excalidraw CLI: GPT-6 Luna vs 6.1 Sol](https://www.youtube.com/watch?v=H64AW1l-ae4) > [!note]+ Document Notes > tags: [obsidian-excalidraw, excalidraw-cli, coding-agents, model-benchmarks, infographic-vectorization] > author: Zsolt's Visual Personal Knowledge Management > source: Zsolt's Visual Personal Knowledge Management (YouTube) > reference: https://www.youtube.com/watch?v=H64AW1l-ae4 > date created: 2026-10-13 08:37 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Obsidian Excalidraw CLI: GPT-6 Luna vs 6.1 Sol (Full Workflow & Benchmark) > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--obsidian-excalidraw-cli-gpt-6-luna-vs-61-sol-full-workflow--benchmark) ## 📝 Summary - **What happened.** Zsolt's Visual Personal Knowledge Management published a 13:11 workflow demo on 2026-10-12. It shows the new command line interface (CLI) in the Obsidian Excalidraw plugin (version 2.29.0 per the description) driving two coding agents at once. Both agents get the same prompt: recreate a ChatGPT-generated raster infographic as native, editable Excalidraw elements. The auto-transcript garbles most proper nouns: the plugin appears as "Excalibro," "Excalibraw" and "Excal," the creator's name as "Jolt," Codex as "Codeex," and Sol as "Soul" and "six months sold." Quotes below keep that spelling. - **The claim.** A cheap, fast model (GPT-6 Luna) finishes a usable first version quickly at a stated cost of 6.6 credits, or 6.6 cents. A stronger model (GPT-6.1 Sol, run "on high") takes 16 minutes, costs about 10 times more (roughly 60 cents), and produces cleaner shapes, better text placement and reusable objects. The creator's verdict is that "intelligence does make a difference." - **The workflow.** Update the plugin to 2.29.0 or higher, turn on Settings > Excalidraw > Command Line Interface, create a sandbox vault, copy the source image in, and point the agent at the creator's Excalidraw CLI skill repository on GitHub with a one-line prompt. - **Why it matters.** The result is a repeatable path from a static AI-generated image to an editable vector canvas, and from there to reusable icons saved in the Excalidraw stencil library. - **Key stakeholders.** The creator (who also maintains the plugin and sells courses), Obsidian users who draw in Excalidraw, GitHub Copilot and OpenAI as the named agent and model providers, and anyone making AI infographics they need to edit afterward. - **Risks and tradeoffs.** The benchmark is one image, one run per model, judged by eye by the person who wrote the CLI. No scoring rubric, no repeated trials, no Luna wall-clock time and no token counts are given. The creator says outright to root for Sol before Sol finishes. Treat the comparison as an anecdote, not a measurement. The agent edits files in a vault, so the sandbox-vault advice is the main safety control and the video never mentions backups or permissions. - **Commercial interests.** The description links the creator's own plugin and CLI repositories, a paid course ("Master Obsidian-Excalidraw"), a Visual Thinking Workshop, a book and community ("Sketch Your Mind"), a newsletter, and a Ko-fi donation page. The video promotes a feature the creator built, so a favorable framing is expected. - **Bottom line.** The CLI workflow is concrete and reproducible from the description. The Luna-versus-Sol conclusion (cheap and fast versus slow and refined, 10x cost gap) is plausible but rests on a single unreplicated trial. ## 📖 Main content **The core idea.** Excalidraw is a vector whiteboard that runs inside Obsidian as a plugin. Version 2.29.0 adds a CLI that lets an external coding agent create and modify drawings in a vault. The demo applies it to a specific pain: ChatGPT makes good-looking infographics, but they're flat images. You can't move one element, restyle one icon or fix one word. The fix shown is to hand the image to an agent and have it rebuild every part as Excalidraw shapes and text. **Setup, as stated.** Four prerequisites: a coding agent (the speaker names Claude, Codex and GitHub Copilot), Obsidian with the Excalidraw plugin, the CLI switch turned on in settings, and a sandbox vault. The sandbox separates "what the agent is doing from all your other stuff." The speaker adds that once you're comfortable you may point the agent at your main vault. Image files get copied into the sandbox vault by hand. The prompt is short: read the GitHub repository address (the creator's skill repo, linked in the description), then recreate the attached infographic. **The race.** Two agents run side by side, each in its own sandbox vault (the screen shows "CLI demo one" and "CLI demo two"). Left: GPT-6 Luna in GitHub Copilot, described as "a few cents" per run and "not so intelligent." Right: GPT-6.1 Sol "on high" (spoken as "Codeex GPT61 Soul on high"). Both receive an identical prompt, which the creator accidentally pasted twice. **What each agent did.** - **Luna** produced a first version first, wrote a PNG export to read its own drawing back, iterated twice, and declared itself done. Result: every part of the infographic present, a robot made of real Excalidraw elements, weaker cog, magnifier and folder. Cost 6.6 credits. - **Sol** took longer to reach a first draft. The first draft had a cleaner robot but no text. The next pass added text, but some objects vanished. Later passes fixed text placement, redrew the cog, magnifier (with a shadow the creator likes), folder and light bulb, and corrected a plus sign and paper lines. The creator believes Sol builds objects once and reuses them, but marks that as a guess ("I'm pretty sure"). Total time 16 minutes, about 10 times Luna's cost. **The editing payoff.** Because the output is native, the creator selects the robot, duplicates it, alters parts, edits text, and saves the robot to the stencil library with right-click > "add to library." The robot is then reusable as an icon. **Contested and unverifiable in this source.** The 6.6-credit and "closer to 60 cents" figures are the creator's reading of GitHub Copilot usage; the 60-cent number is an estimate "if you count by the tokens." Luna's elapsed time is never stated, only that it was "quicker." Which agent harness ran Sol is ambiguous: the transcript says "Codeex," while the cost comparison refers to GitHub Copilot tokens. The quality verdict is visual and subjective. Whether Luna's file or Sol's file better preserves editability (grouping, layer structure, element count) is not examined. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=H64AW1l-ae4&t=0) Introduction: Excalidraw CLI is here - Announces the CLI feature in Obsidian Excalidraw "229" (2.29.0 in the description), implemented by the creator. - The CLI lets a coding agent such as Claude, Codex or GitHub Copilot automate Excalidraw, for example by creating diagrams. - The video focuses on one use case: converting a ChatGPT infographic about how the CLI process works. > [!quote] [00:00:00] > I implemented CLI, the command line interface for Excalibraw, which allows your coding agent claude, Codeex, GitHub Copilot or whatever you're using to automate Excal #### ⏱️ [00:00:32](https://www.youtube.com/watch?v=H64AW1l-ae4&t=32) The challenge: making AI infographics truly editable - Problem: generated infographics are hard to modify, and you might not like the full style. - Proposed fix: convert the image into an Excalidraw drawing using the CLI. - Framing device: a race between two models, introduced before the setup walkthrough. #### ⏱️ [00:01:04](https://www.youtube.com/watch?v=H64AW1l-ae4&t=64) Setup prerequisites: agents, vaults and plugin - Four needs: a coding agent, Obsidian, the Excalidraw plugin, and the CLI setting enabled. - A sandbox vault keeps agent activity apart from the main vault. Images are copied in by hand. - Later, the creator says users may grant access to the main vault. - A GitHub repository holds a skill so the agent starts quickly. The link is in the description. > [!quote] [00:00:45] > Then you need to enable the CLI functionality on Obsidian. #### ⏱️ [00:01:58](https://www.youtube.com/watch?v=H64AW1l-ae4&t=118) The race: fast and cheap (Luna) versus high-end (Sol) - Left contender: GPT-6 Luna, run in GitHub Copilot, "super low cost" and "a few cents" per job, but "not so intelligent." - Right contender: GPT-6.1 Sol on high, run as spoken in Codex, described as powerful and "super intelligent." - The source image sits in each sandbox vault as the recreate target. #### ⏱️ [00:02:52](https://www.youtube.com/watch?v=H64AW1l-ae4&t=172) Enabling CLI in Obsidian Excalidraw settings - Path: Settings, then search "command line interface," then toggle it on. - Two sandbox vaults are visible on screen, named CLI demo one and CLI demo two. - Both prompts are submitted together; the agents need "a couple of minutes." #### ⏱️ [00:03:33](https://www.youtube.com/watch?v=H64AW1l-ae4&t=213) The simple agent prompt and GitHub skill repository - Prompt shape: "read" plus the GitHub repository address, then "recreate the attached infographics." - The creator accidentally pasted the prompt twice. Both models received the same text. - The agents will create files in the vault whose iterations can be opened and compared. > [!quote] [00:03:05] > So my prompt is super simple. It says read and I give it the address of the GitHub repository in the CLI demo 2 Obsidian Walt. #### ⏱️ [00:04:32](https://www.youtube.com/watch?v=H64AW1l-ae4&t=272) Round 1: Luna generates a first draft quickly - Luna finishes a first version before Sol does, though the creator says it isn't yet as good as wanted. - Luna writes a PNG file and uses it to read the drawing back, then updates the drawing from what it sees. - This is the first sign of the render-and-look feedback loop that both agents use. #### ⏱️ [00:05:15](https://www.youtube.com/watch?v=H64AW1l-ae4&t=315) Round 2: Sol focuses on clean geometry and vectors - Sol's first version is cleaner than Luna's, but has no text yet. - The robot Sol drew looks better than Luna's robot. - Creator corrects the naming slip: the model is 6.1 Sol, not "Codex" (the transcript renders it "61 Soul"). - Luna posts an update the creator calls "actually pretty good." #### ⏱️ [00:06:29](https://www.youtube.com/watch?v=H64AW1l-ae4&t=389) Iteration and vision feedback loop - Sol adds text, though "not really yet too good." Objects from the earlier draft disappear. - Luna appears to stop: the creator reads that as Luna considering the work final. - Luna's cost is shown on screen and read aloud: 6.6 credits, which is 6.6 cents. > [!quote] [00:06:48] > So this ate 6.6 credits. That's 6.6 cents. That is not a high amount of money. #### ⏱️ [00:07:08](https://www.youtube.com/watch?v=H64AW1l-ae4&t=428) Evaluating Luna's output: speed, quality and cost - Zooming in shows the robot is built from real Excalidraw elements. It can be selected, duplicated and altered. - Text is editable. The creator rates the result "not so perfect" but a usable infographic. - The creator says to still be cheering for Sol, expecting a better outcome. > [!quote] [00:07:33] > robot is actually a proper excol element. So you can see here I can select the robot. #### ⏱️ [00:08:26](https://www.youtube.com/watch?v=H64AW1l-ae4&t=506) Sol's progress: object reuse and detail refinement - Sol's text is now in the right place and the Obsidian logo looks more refined than Luna's. - Luna's cog is weak. Sol's cog is rated "pretty awesome." - The creator suspects Sol builds robot, cog and logo as objects and reuses them, flagged as an inference. - Remaining Sol problems at this point: the plus sign and lines on the paper. - Magnifier: the creator prefers Luna's rectangle but likes Sol's light and shadow on the glass. #### ⏱️ [00:10:06](https://www.youtube.com/watch?v=H64AW1l-ae4&t=606) Side-by-side visual comparison - A new Sol update corrects the plus sign, the paper lines and the text. - The creator switches both canvases to view mode to compare them side by side. - Folder and light bulb are rated better in Sol's version. > [!quote] [00:10:36] > I think intelligence does make a difference. #### ⏱️ [00:11:09](https://www.youtube.com/watch?v=H64AW1l-ae4&t=669) The verdict: speed, intelligence and cost breakdown - Luna was "much faster." Sol took 16 minutes. - Sol cost 10 times more, estimated near 60 cents by Copilot token count. - The creator notes a monthly or weekly quota makes that difference small in practice. - Opinion on Sol's result: "pretty damn good," with the creator "super happy." > [!quote] [00:11:23] > 61 soul cost 10 times more. So this is probably around closer to 60 cents if you count by the tokens in GitHub Copilot. #### ⏱️ [00:12:04](https://www.youtube.com/watch?v=H64AW1l-ae4&t=724) Editing the diagram and saving vectors to your stencil library - Elements can be moved around because the canvas is native Excalidraw. - Two stated uses: convert ChatGPT infographics into editable drawings, and extract icons. - Saving an icon: select the robot, right-click, choose add to library. > [!quote] [00:12:07] > You just select the robot like that and you right click and you select add to library. #### ⏱️ [00:12:56](https://www.youtube.com/watch?v=H64AW1l-ae4&t=776) Wrap up and what's next - Closes with excitement about use cases and an invitation to share them in the comments. - The transcript ends at about 12:52 of speech. The remaining seconds to 13:11 carry no captioned words. - The description links two more videos (youtu.be/6BjhUyfS4iM and youtu.be/am2HOlbYsxI) with no title or context. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [Obsidian-Excalidraw CLI (GitHub)](https://github.com/zsviczian/obsidian-excalidraw-cli): the skill repository the prompt points the agent at. The speaker says its link goes in the description. - [Excalidraw Automate script template (GitHub)](https://github.com/zsviczian/ea-script-template): listed under "AI Scripting / Excalidraw Automate Template." The transcript doesn't mention it. - [Obsidian Excalidraw plugin (GitHub)](https://github.com/zsviczian/obsidian-excalidraw-plugin/): the plugin that ships the CLI in 2.29.0. - [Obsidian](https://obsidian.md): the host application. - [Sketch Your Mind community](https://community.sketch-your-mind.com/welcome), [book](https://community.sketch-your-mind.com/sym), [Visual Thinking Workshop](https://community.sketch-your-mind.com/vtw), [Master Obsidian-Excalidraw course](https://community.sketch-your-mind.com/em) and [newsletter](https://sketch-your-mind.com/newsletter): the creator's paid and free programs, listed in the description. - [Ko-fi donation page](https://ko-fi.com/zsolt): plugin support link in the description. - [NotebookLM notebook for plugin questions](https://notebooklm.google.com/notebook/42d76a2f-c11d-4002-9286-1683c43d0ab0): described as a place to ask questions about the plugin. - GitHub Copilot, Codex and Claude: the agents named in the transcript and description. Model names "GPT-6 Luna" and "GPT-6.1 Sol" come from the title and description. The transcript spells them "GPT6 Luna," "GPT61 Soul," and "six months sold." No vendor documentation for those model names is cited. - [Obsidian Members Group Discord handle @zsviczian](https://twitter.com/zsviczian) and [Mastodon](https://pkm.social/@zsviczian): contact channels in the description. ## 🗣️ Paraphrase Zsolt wants a ChatGPT infographic to be editable. A flat image can't be restyled one piece at a time, so the video shows how to rebuild it as native Excalidraw shapes by letting a coding agent drive Obsidian through the new CLI in plugin version 2.29.0. The agent reads a skill repository, then redraws the image in a sandbox vault, checking its work by exporting a PNG and looking at it. To make the demo watchable, Zsolt runs two models in parallel with the same prompt. Luna, the cheap one, delivers something usable fast and spends 6.6 cents. Sol, the expensive one, grinds for 16 minutes at roughly 10 times the price and ends up with cleaner geometry and nicer small details like the cog, the folder and the light bulb. The takeaway offered is that spending more on the model gets you a better drawing, though the creator also says Luna's version is already a usable infographic. Both outputs are real vector objects, so any element can be duplicated, edited or saved to the stencil library as a reusable icon. ## 📰 Newsletter **Turn a flat AI infographic into something you can edit.** Zsolt's 13-minute demo is the clearest walkthrough yet of the new Excalidraw CLI (plugin 2.29.0). You switch on the CLI in settings, give a coding agent the creator's skill repo and a screenshot, and it rebuilds the picture out of real Excalidraw elements. Luna, a budget model, finished first at 6.6 cents. Sol, a stronger model, took 16 minutes at about 10 times the cost and drew cleaner cogs, folders and light bulbs. The creator's takeaway: "intelligence does make a difference." What's worth borrowing for documentation and community work: the render-and-look loop. Both agents export a PNG, view it and fix what they see. That pattern applies to any generated diagram. Two cautions. The benchmark is one image judged by the person who built the feature, and the creator sells courses around the plugin. Run your own test on a diagram you care about before you pick a model. Links: [the video](https://www.youtube.com/watch?v=H64AW1l-ae4), [the CLI skill repo](https://github.com/zsviczian/obsidian-excalidraw-cli), [the Excalidraw plugin](https://github.com/zsviczian/obsidian-excalidraw-plugin/). ## ✅ Action items #### DevRel - Developer advocacy: reproduce the demo on one Pantheon or WordPress diagram and publish the per-run cost and elapsed time for two models, using the same prompt, so readers can compare against Zsolt's numbers. - Education: run an office-hours session on turning generated infographics into editable vectors, using the sandbox-vault pattern as the safe starting point. - Developer experience: write down the render-export-inspect loop (agent exports a PNG, reads it back, edits) as a reusable pattern for any agent that draws. - Developer marketing: don't cite "10 times cost" or "6.6 cents" without noting it is one run on one image by the feature's author. - Community management: when members share model comparisons, ask for the prompt, run count and cost source. - Internal/external: if the team adopts Excalidraw for docs diagrams, check the CLI repo's license and update cadence before depending on it. #### OSPO - Policy and governance: decide whether agents may write into a shared documentation vault, and require a sandbox copy first, as the video recommends. - License compliance: read the licenses of the CLI repo and the plugin before bundling either into a docs workflow. - Security management: an agent with CLI access to a vault can create and edit files. Document what the CLI exposes and limit the vault scope. - Software inventory: if the workflow enters a pipeline, record the plugin version (2.29.0 or higher) and the agent harness as dependencies. - Developer support: write a troubleshooting note for the setting path (Settings, Excalidraw, Command Line Interface). - Community engagement: invite the creator to share the skill repo as a worked example of an agent skill. - Strategic alignment: track agent spend per finished diagram so model-tier choices rest on data. #### Community - Maintainers: test the CLI on one contributor-guide diagram and record whether the editable output saves review time. - Contributors: add a convention for storing the source raster beside the editable canvas so changes stay traceable. - Sponsors: if a vendor funds an AI-diagram pilot, publish cost and quality results from more than one image. - Ecosystem: compare the same infographic across at least three models and publish the rubric with the results. #### WordPress - Core: no action. This is a note-taking and diagram workflow, not core code. - Core Performance: none from this source. - Core AI: note the render-inspect-edit loop as a pattern for agents that work on visual output. - **Core Program Team:** none directly. Nearest relevance: editable diagrams for release and roadmap communication. - Mobile: none. - Playground: none directly. A sandbox-first rule maps to Playground's disposable-site philosophy. - Community: none from this source. - Training: build a lesson on converting a generated infographic to editable vectors and saving icons to a library. - Marketing: none directly; editable infographics could ease localized or restyled campaign graphics. - Support: none. - Design: test whether the stencil-library flow suits shared icon sets. - Documentation: evaluate editable Excalidraw diagrams (versus flat PNGs) for handbook pages, and keep the editable source in the repo. - Photos: none. - TV: none. - Polyglots: editable text layers would let translators swap labels in infographics without redrawing. Test it. - Accessibility: none from this source. Nearest relevance: any exported infographic still needs alt text and a text equivalent. - Meta: none. - CLI: nearest relevance: an application exposing a CLI for agents, with the setting off by default. Compare with WP-CLI's permission model. - Hosting: none. - Tide: none. - Openverse: none. - Themes: none. - Plugins: none. Nearest relevance: a plugin shipping an agent-facing CLI is a pattern worth tracking. #### The WPOCC - Programs: none directly. If The WP Open Community Collective funds documentation tooling, editable diagrams are a candidate line item. - Projects: none from this source. - Funding experiments: if a microgrant covers AI-assisted diagramming, require the cost and quality data from more than one image. ### FAIR - Technical independence: the workflow depends on a hosted model plus a local plugin. Document which part would break if a vendor changed its model line-up. - Policies and covenants: none directly from this source. - Productized governance artifacts: none from this source. Nearest relevance: a checklist for agent write access to shared content. ## 🧪 Hypotheses 1. On a set of 10 different infographics, GPT-6.1 Sol ranks above GPT-6 Luna on blind-rated fidelity for at least 8, matching the single-case result. 2. The cost ratio of about 10 to 1 and the 16-minute run time hold within a factor of 2 across the same 10 infographics. 3. Giving Luna a second pass (re-prompt with Sol's critique list) closes more than half of the quality gap at under twice Luna's cost. 4. Sol's output contains fewer distinct drawing elements per visual part than Luna's, because it reuses objects, and that makes it easier to edit. 5. A blind rater who doesn't know which model made which file picks Sol's version over Luna's no more often than the creator did, once rater and feature author differ. 6. The sandbox-vault step prevents unintended edits to the main vault in 100 percent of runs, while pointing an agent at a main vault produces at least one unintended change in 20 runs. ## 📊 Methodology - **Design for H1 and H5.** Collect 10 infographics of varying density. Run each model once per image with the creator's skill repo and the same one-line prompt. Export each result as PNG, shuffle, strip labels, and have three raters score fidelity (layout, text, icons) from 1 to 5. Report the mean and rater agreement. - **Design for H2.** Record wall-clock time from submit to "done" for every run, and read cost from the Copilot usage panel after each run rather than estimating from tokens. - **Design for H3.** For Luna, add a second prompt that includes a numbered list of defects from the first output. Compare scores and total cost with Sol's single-pass runs. - **Design for H4.** Parse each saved Excalidraw file and count elements and unique repeated shapes per visual part. Compare Luna and Sol. - **Design for H6.** Prepare a vault with 50 notes, hash all files before and after runs, and diff. Repeat in sandbox and main-vault conditions across 20 runs each. - **Controls throughout.** Same plugin version, same agent harness for both models, same reasoning setting stated for both, same starting image. Pre-register the scoring rubric before the first run. ## ❓ Questions - What exact agent harness ran Sol? The speech says "Codeex," while the cost estimate cites GitHub Copilot tokens. - How long did Luna take? The video says only that it finished first. - What does "on high" refer to: a reasoning-effort setting, and what were the settings for Luna? - Are "6.6 credits" and "closer to 60 cents" billing figures or the creator's estimates from the Copilot usage panel? - How many elements does each output contain, and are they grouped and named for later editing? - How did the original source image get chosen, and would the ranking change on a denser or hand-drawn infographic? - What does the CLI expose to an agent, and can it be limited to one vault or one folder? - Is the skill repository version-pinned, or can the agent pick up changes to it mid-run? ## 🔮 Future research - **Blind, repeated model comparisons for vector recreation.** The field has little published work on scoring image-to-editable-vector conversion; a shared rubric and image set would let comparisons stack. - **Editability metrics.** Count of reused objects, grouping depth and number of elements per visual part are all candidates for measuring whether an output is easy to edit, not only whether it looks right. - **Agent write-access safety in note vaults.** There's no standard guidance on scoping CLI access for agents in personal knowledge bases. - **Render-and-look loops.** Measure how many feedback iterations each model needs and whether more iterations cost more than a stronger model. - **Cost framing.** Quota-based pricing (Copilot credits) hides the real token cost. A cost-per-accepted-output metric would make comparisons portable. ## 💡 Implications - **Practice.** For flat-to-vector work, start with the cheaper model for a first draft, then decide whether the stronger model's 10x spend buys enough polish for your use. The video gives one data point for that choice. - **Policy.** Any workflow that lets an agent write into a vault needs a stated scope. The sandbox vault is the video's answer, and it works as a default policy. - **Security.** The CLI is a control surface for file changes. Keep it off unless needed, and review what the skill repository instructs agents to do before running it. - **Education.** Converting a generated image to editable vectors teaches both the weakness of raster output and the value of a feedback loop. It's a compact lab for a workshop. - **Funding.** The 6.6-cent figure suggests per-diagram agent cost is small next to staff time. Budget by accepted output, and track time as well as spend, since Sol took 16 minutes. - **Vendor literacy.** A demo from the author of the feature, with links to the author's courses and donation page, is a useful case for reading creator-led benchmarks critically. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch Reading Ease on the transcript body is about 78 and the Flesch-Kincaid grade level about 6 to 7: short, conversational sentences. Domain terms (CLI, sandbox vault, stencil library, agent skill) lift effective comprehension to an intermediate Obsidian user. #### 👀 Visual learners - Present: the whole video is a screen recording with side-by-side canvases, original versus Luna versus Sol, and zoom-ins on the robot. - Improve: no on-screen step list for setup, and no labeled before-and-after still at the end. A final comparison frame with three captions would anchor the verdict. #### 🔊 Aural learners - Present: continuous narration that names each change as it appears ("the plus and the lines corrected"). - Improve: long silent waits are narrated only by "we're going to take a break." A spoken recap of setup steps at the end would help listeners. #### 🗣️ Verbal learners - Present: setup is stated in order (agent, Obsidian, settings toggle, sandbox vault, prompt) and the prompt is read aloud. - Improve: the full prompt text and the setup list live partly in the description, not in the narration. Say the repository name aloud. #### 🧩 Kinesthetic learners - Present: concrete clicks (Settings, search, toggle) and a hands-on edit (select, duplicate, add to library). - Improve: no exact prompt on screen long enough to copy, and the CLI command forms are never shown. A copy-paste block in the description would let viewers follow along. #### 🧠 Logical learners - Present: a controlled setup in intent (same prompt, two models) and explicit cost and time figures. - Improve: no rubric, no repeated runs and no stated Luna time. State the criteria for "better" before showing the result. #### 🧑‍🤝‍🧑 Social learners - Present: closes by asking viewers to share use cases in comments. The description lists a Discord group, Mastodon and a community. - Improve: no prompt for viewers to post their own model comparisons, and no shared test image for a community benchmark. #### 🧘‍♀️ Solitary learners - Present: self-contained 13-minute video with chapters, so it's easy to pause, replay and try the steps alone. - Improve: no end-of-video checklist or exercise. "Convert one of your own infographics" would give a clear task. #### 🌿 Naturalistic learners - Present: nothing explicit. The subject is a digital workflow. - Improve: running a high-effort model for 16 minutes carries an energy cost. A mention of compute per run would connect the 10x price gap to resource use. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies player landmarks, a chapter list (15 chapters) and a text description with links. - Improve: the infographic, both outputs and the Excalidraw UI are visual only. WCAG 2.2 AA 1.1.1 (Non-text Content) and 1.2.5 (Audio Description, Prerecorded) aren't met for the side-by-side comparison, which carries the verdict. #### Keyboard navigation - Present: the YouTube player is keyboard operable with visible focus, and chapter links give a jump path. - Improve: the video adds no barriers. The demo shows mouse-only editing (click, right-click, drag). Mention the keyboard equivalents for duplicate and add to library. 2.1.1 passes via the platform, not via the content. #### Contrast and readability - Present: unverifiable from a transcript. The demo canvas appears to be a light Excalidraw canvas with the Obsidian side panel. - Improve: small canvas text at default resolution risks 1.4.3 (Contrast Minimum). Zoom into the compared regions longer and read out the labels. #### Text resizing and reflow - Present: the player scales, and captions resize on their own. - Improve: burned-in on-screen text and the infographic's own text don't resize, so 1.4.4 (Resize Text) and 1.4.10 (Reflow) fail for text inside the frame. #### Media and captions - Present: captions exist (auto-generated), plus 15 chapters and a timestamped description. - Improve: the auto-captions garble the plugin ("Excalibro," "Excalibraw," "Excal"), the creator's name ("Jolt"), Codex ("Codeex") and Sol ("Soul," "six months sold"). That is a 1.2.2 (Captions, Prerecorded) failure in substance: captions exist but are wrong on most proper nouns, including both model names that carry the comparison. Human-corrected captions are the fix. The chapter titles and description do spell the names correctly. #### Forms and interactions - Present: the video itself has no form. Enabling the CLI is a single toggle in Obsidian settings, shown on screen. - Improve: the toggle's label and path are said aloud once ("command line interface"). Put the exact path in the description as text, which the description's quick setup summary already does. 3.3.2 (Labels or Instructions) applies to the plugin settings page, not the video. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate captions on key terms), **1.2.5** (no audio description for the visual comparison), **1.1.1** (infographic and outputs have no text equivalent), **1.4.4 / 1.4.10** (burned-in text). Corrected captions are the highest-impact fix. ## 🌍 DEIB assessment #### Cultural breadth - Present: the demo subject (an infographic about the CLI process) is tool-neutral. The creator's materials span a book, community and workshops, which suggests an international audience. - Improve: only an English infographic is tested. A non-Latin script or right-to-left layout would show whether the agents handle other text. #### Inclusive language - Present: no exclusionary terms in the transcript. Language is direct and second person. - Improve: superlatives like "amazing" and "super intelligent" for a model carry ability framing. Swapping them for measured results (6.6 cents, 16 minutes) helps. #### Neurodiversity support - Present: chapters, a clear structure (setup, race, verdict, edit) and a quick setup summary in the description. - Improve: long stretches of waiting commentary ("let's wait for it") offer little new information. Cut to a timestamped change log on screen. #### Representation - Present: a single presenter, the creator. Names in the description refer to the creator's own handles. - Improve: no outside user or second voice appears. A user who tried the CLI on a different image would add perspective. #### Psychological safety - Present: the tone is warm and low-pressure. The creator admits mistakes (pasting the prompt twice, the naming slip) and says Luna's output is usable. - Improve: telling viewers to expect the stronger model to win before the result is in can make a viewer on a free tier feel they're behind. State the cheap model's use case plainly. ## ✅ Logical validity #### Strong reasoning - **Controlled setup in intent, 00:03:05 to 00:03:50.** Both models received the same prompt, and the creator says so, including the accidental double paste. The identical-input design is the right base for a comparison. - **Direct observation, 00:07:33.** Zooming in and showing that the robot can be selected and duplicated demonstrates the key claim (output is native Excalidraw, not an image) with a visible action. - **Cost accounting with a caveat, 00:11:23.** The 10x ratio is paired with a note that a monthly or weekly quota blunts its impact. That qualifies the number instead of inflating it. - **Process transparency, 00:04:37 to 00:06:48.** The creator reports intermediate problems in both runs (Sol's missing text, vanished objects), not only the end states. #### Fallacies - **Hasty generalization, 00:10:36.** "I think intelligence does make a difference." One image, one run per model, one rater. *Fix:* run 10 or more images and report the spread, or phrase it as a single observation. - **Biased judge and confirmation framing, 00:07:33.** The creator says to be "still cheering" for Sol while Luna's result is under review, then scores Sol higher. Rooting for an outcome mid-run primes the scoring. *Fix:* set criteria first or use blind raters. - **Unquantified quality claims, 00:08:20.** "definitely more refined" and "pretty awesome" carry no rubric. *Fix:* list the features scored (text placement, shape accuracy, color match) and rate each. - **Unsupported inference stated tentatively, 00:09:04.** "I'm pretty sure that Soul is reusing objects" is a guess about internals from what's visible. The creator hedges it, so this is a flag, not a fault. *Fix:* inspect the saved file and count repeated shapes. - **Undisclosed interest, throughout.** The creator wrote the CLI and sells courses and community access, listed in the description. The video never states that interest aloud. *Fix:* one spoken line at the top. - **Ambiguity from uncorrected captions, throughout.** "Soul," "six months sold" and "Codeex" make the model identities unusable as cited facts without the description. *Fix:* corrected captions. #### Tone of voice Enthusiastic, informal and first-person, with contractions throughout ("we're going to take a break," "I'm super happy"). Superlatives ("amazing," "game-changing" in the description) cluster at the open and close, and the middle is narrated observation with honest intermediate failures. The creator audibly hedges ("I'm pretty sure," "so don't judge just yet"). That pattern means the process description is more reliable than the verdict. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response pattern to accusation or criticism. This is a product demo with no accusation, no accuser and no defensive posture. For completeness across the three prongs: **Deny**: nothing is denied; the creator names Sol's early problems and Luna's weak cog. **Attack**: neither model is disparaged; Luna's result is called "pretty okay" and usable. **Reverse victim/offender**: no such move appears. ## 🎯 Conclusion The durable content is the workflow. Turn on the CLI in Excalidraw 2.29.0, work in a sandbox vault, give the agent the skill repository and a short prompt, and let it redraw a raster infographic as native elements. The render-export-inspect loop, shown at 00:04:37 with Luna's PNG, is the mechanism that makes agent-drawn diagrams converge. The editing payoff at 00:07:33 and 00:12:07 is real and visible: select, duplicate, restyle, save to the library. The benchmark conclusion is weaker than its delivery. Luna cost 6.6 cents and finished fast. Sol took 16 minutes at about 10 times the cost and drew cleaner details. That is one image, one run each, judged by the feature's author, who said they were cheering for Sol. Elapsed time for Luna, token counts and the exact harness for Sol are missing. The creator also sells a course and community, which the video doesn't disclose aloud. **Next steps for readers:** copy the setup from the description (plugin 2.29.0 or higher, CLI toggle, sandbox vault), run both a cheap and a strong model on one of your own diagrams, and record time, cost and a written rubric before deciding which tier you need. Check the model names against the description, because the auto-captions garble them. ## 🔄 Refactor Appendix — Obsidian Excalidraw CLI: GPT-6 Luna vs 6.1 Sol (Full Workflow & Benchmark) > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied. creator promotion > and the "game-changing" description claim were dropped as unevidenced. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] A one-line prompt plus a skill repo drives the agent > [!quote] [00:03:05] > my prompt is super simple. It says read and I give it the address of the GitHub repository **Context & use**: - 00:03:05, as both agents are submitted. The skill repository carries the instructions, so the prompt stays short. - A reusable pattern: put the procedure in a linked skill, keep the prompt to a pointer plus the task. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H2] The cheap model's run costs 6.6 cents > [!quote] [00:06:48] > So this ate 6.6 credits. That's 6.6 cents. That is not a high amount of money. **Context & use**: - 00:06:48, when Luna declares itself ready. The figure is the creator's reading of Copilot credits. - Keep with the caveat that Luna's elapsed time is never stated, so cost alone isn't a full comparison. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H3] Agent output becomes real, editable Excalidraw elements > [!quote] [00:07:33] > robot is actually a proper excol element. So you can see here I can select the robot. **Context & use**: - 00:07:33, zoomed into Luna's robot. The point is that the infographic turns into objects, not a placed image. - This is the feature's core value claim and the only one the video demonstrates by direct action. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H4] The creator judges stronger models draw better > [!quote] [00:10:36] > I think intelligence does make a difference. **Context & use**: - 00:10:36, comparing folders and light bulbs. A single-rater, single-image judgment, and the rater said earlier they were cheering for Sol. - Treat as a hypothesis to test across images, not a finding. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H5] The stronger model costs ten times more > [!quote] [00:11:23] > 61 soul cost 10 times more. So this is probably around closer to 60 cents if you count by the tokens in GitHub Copilot. **Context & use**: - 00:11:23, the verdict segment. The 60 cents is an estimate from token counts, and Sol's run lasted 16 minutes (00:10:36). - Pair with the creator's own note that a monthly or weekly quota makes the gap small in practice. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] Let the agent look at its own output **Claim** (one sentence, Courtney's voice): An agent that exports what it drew and reads it back fixes visual errors that a text-only edit loop never sees. **Evidence pointer**: [[#H1]] · [[#H3]] **Scope & conditions**: Holds when the agent has a vision-capable model and a way to render its output. Breaks when renders are expensive, when the target is not visual, or when the agent can't compare against the original. The video shows the loop only qualitatively, so the size of the gain is unmeasured. **Links**: - [[Agent feedback loops]] — because extends - [[Verify delegated work]] — because applies-in render-and-inspect before calling a diagram done - [[Structured output with schema validation]] — because contrasts on checking visual output instead of schema-valid output <!-- atomicity: 1 idea ✓ --> #### [P2] A one-image benchmark is an anecdote **Claim** (one sentence, Courtney's voice): A model comparison on one input, one run and one rater who built the tool tells you what happened once, not which model is better. **Evidence pointer**: [[#H4]] · [[#H5]] **Scope & conditions**: Applies to any creator-led model comparison. It doesn't discredit the process observations (what each model got wrong at each step), which are reliable as reports. The tell is a rubric-free quality verdict plus a rooting statement before the result. **Links**: - [[Reading vendor benchmarks]] — because extends - [[Every claim carries a fact]] — because applies-in writing standards - [[Self-limiting disclosure raises credibility]] — because contrasts on direction of interest — this video admits intermediate failures but not its commercial ties <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Confirm real model names and plugin version from the description: GPT-6 Luna, GPT-6.1 Sol, Excalidraw 2.29.0. The auto-captions say "Soul," "six months sold," and "Excalibro": https://www.youtube.com/watch?v=H64AW1l-ae4 #fleeting #inbox - [F2] Find out which harness ran Sol: speech says "Codeex," but the cost estimate cites GitHub Copilot tokens at 00:11:23 #fleeting #inbox - [F3] Open question: what can the Excalidraw CLI expose to an agent, and can it be scoped to one folder or vault? #fleeting #inbox - [F4] Try the sandbox-vault pattern on one of my own diagrams and record time and cost for both a cheap and a strong model #fleeting #inbox - [F5] Article seed: "Make the generated infographic editable": the image-to-vector workflow with a render-and-look loop #fleeting #inbox #articleseed - [F6] Check whether the stencil-library flow (select, right-click, add to library) fits my icon sets #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Read the Excalidraw CLI skill repo and the 2.29.0 release notes, then confirm the setup steps before trying them: research, 20 min #things - [ ] Run a cheap-versus-strong model test on one of my own infographics and log time and cost: diagram workflow experiment #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Obsidian Excalidraw CLI GPT-6 Luna vs 6.1 Sol.md|Obsidian Excalidraw CLI: GPT-6 Luna vs 6.1 Sol by Zsolt's Visual Personal Knowledge Management]] ## Highlights - [00:03:05] to turn that on for this to work. And what you will notice here is this is CLI demo two. This is CLI demo one. But let the race begin. I'm going to submit both of these queries and the agents start to work. It is going to take a couple of minutes. So we're probably going to take a break or at least I'm going to take a break and come back once we are getting more done. But I just want to show you the prompt first. So my prompt is super simple. It says read and I give it the address of the GitHub repository in the CLI demo 2 Obsidian Walt. So that's there. Recreate the attached [00:03:50] infographics. And it seems I actually accidentally pasted this there twice. You can see here it's the same prompt. So Luna and GPT1 soul received the same item. What we're going to see in a moment is we're going to see how both tools are going to create a file here and then we can open the images and we can see the iterations. Now I'm going to stop talking. We're going to get back once we have something to look at and discuss. So it seems Luna is quicker. It has ([View Highlight](https://read.readwise.io/read/01m4tmg3hqb6xsbffafxt97tz1)) - Note: [H1] A one-line prompt plus a skill repo drives the agent - Tags: atom/literature/h1 - [00:06:04] I'm I'm going to close this site panel so we can make this even larger. But you can see here that the image that Luna has created is actually not bad. You have every part there. And I'm pretty confident that Luna is going to call this complete, but we might see one more update on this from Luna. Let's wait for it. Okay, so Codeex 61 Soul has added text. Um, the text is not really yet too good. Also, some changes happened. So, the objects from this item disappeared. So, I'm pretty sure that Codex is going to [00:06:48] or 61 soul. I keep calling it CEX. So, you know what I mean. 61 Soul, I'm pretty sure, is going to do a couple of more updates. And I think Luna here seems to think that the end result is getting to the final one. And I think yes. So Luna feels ready. And I actually think this one here looks pretty okay. And we can just take a look. So this ate 6.6 credits. That's 6.6 cents. That is not a high amount of money. And if we zoom in then you can see here that this ([View Highlight](https://read.readwise.io/read/01m4tmg5wcqj7vj7ct2qwqtghg)) - Note: [H2] The cheap model's run costs 6.6 cents - Tags: atom/literature/h2 - [00:07:33] robot is actually a proper excol element. So you can see here I can select the robot. I can actually duplicate my robot. I can uh if I want I can change certain elements of this robot. Um I mean I can just make it uglier but yeah there you go. So I can I can start to play with this also. Of course I can change the text here. And I mean from now on it's not a big surprise. But yeah, I think I think Luna is done. It's not so perfect, but but I think already a pretty usable infographics. I'm still cheering for six months sold because I'm [00:08:20] pretty confident it's going to be better. And yes, I'm sure we're going to get one more update. But you can see here now the text is in the right place. You can also see that this Obsidian here and this Obsidian, I mean, this is definitely more refined than this one right here. Overall, the look and feel. I mean, of course, soul is not yet ready, so don't judge just yet, but I think it's already becoming pretty nice. Also, if you can look here, so Luna didn't do very good on this cog. But the cog that Soul created, I think, is ([View Highlight](https://read.readwise.io/read/01m4tmg8421kg3p3503hhfgtr1)) - Note: [H3] Agent output becomes real, editable Excalidraw elements - Tags: atom/literature/h3 - [00:10:36] if you look at the folder that Saul created is much nicer than the folder that Luna created and the light bulb as well. So again, this light bulb is super nice. This light bulb is so so so yeah I think intelligence does make a difference. All right and now 61 soul is ready as well and let's compare. So Luna was definitely much faster and we can just check the total time. So this took quite some time. It took 16 minutes for soul to do its job. But if now we look at the [00:11:23] original image, this was our original image and this is what 61 soul made of this. And this is what six Luna made of this. Now the difference is 61 soul cost 10 times more. So this is probably around closer to 60 cents if you count by the tokens in GitHub Copilot. Of course, if you have a monthly quotota or a weekly quotota, then yeah, this is not a big impact on your quotota. So this is actually not bad. But yeah, I think this is pretty damn good. I'm super happy with this. Now I can edit this. Now I ([View Highlight](https://read.readwise.io/read/01m4tmgan4261vy3hadz9f8avg)) - Note: [H4] The creator judges stronger models draw better - Tags: atom/literature/h3