Skip to content
Changelog

Content updates.

Dated notes on course content. Newest first. Each line names the course and the module.

Entries

20 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Claude Code's experimental advisor. The main model does the work and tends to consult a stronger model before committing to an approach, when an error keeps recurring, and before declaring a task done. The advisor reads the full conversation, including every tool call and result. Turn it on with /advisor opus, with advisorModel set to opus, or with claude --advisor opus. Use it as the critic. A read-only subagent can map the repo, or run tests and report only failures, and your instructions say when to consult. The advisor must rank at or above the main model. A weaker one is not attached, and the main model runs without it. It isn't available on Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry. On Claude Platform on AWS or Claude Models on Google Cloud (Anthropic managed), it needs Claude Code v2.1.300 or later. On the Anthropic API, the session also needs the feature flag fetched from Anthropic. On a subscription plan, advisor usage counts toward plan limits. With API billing it is charged as extra tokens. Kleos's post is 18 November 2026, 16:19 UTC (19 November 2026, 03:19 Sydney).

  • The spoken line on the pass-or-fail check now names Addy Osmani, then the figure: Addy Osmani says flaky tests train a rerun instead of a fix, and slow tests mean fewer runs. He says that on the Claude Code team, about ninety percent of the code is written and reviewed by Claude.

19 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Give every agent task a check that returns pass or fail. Close the loop with a prompt such as run the auth tests after, and keep going until they pass; a /goal condition; a Stop hook that sends Claude back when the script exits 2; /code-review in a fresh subagent in the terminal; and /verify, which builds the app and looks at the result. Ask for the command and its output, or a screenshot. Addy Osmani's article, posted 17 November 2026, 07:20 UTC (17 November 2026, 18:20 Sydney), says flaky tests train the agent to run them again rather than fixing the failing code, and that slow tests mean fewer runs across batches. The same article says that on the Claude Code team, the vast majority of their code (~90%) is written and reviewed by Claude.

  • Claude Code's experimental Code Mode: Claude writes JavaScript that chains several tool calls in one step. Turn it on with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 in the shell or in ~/.claude/settings.json. Project settings are ignored, so a cloned repo cannot turn it on. It needs v2.1.323 or later and a terminal session. Claude in Chrome and computer use are unavailable, so a visual QA loop that needs Claude in Chrome will not run. Code stops after 10 minutes unless it runs in the background, and then you can keep talking while it runs. Background code needs the session to have started in auto or bypassPermissions mode. A token rename is one example of what the docs call a change that repeats across many files. It may change or be removed. ClaudeDevs' post is 17 November 2026, 21:18 UTC (18 November 2026, 08:18 Sydney).

  • Firecrawl Local scrapes, crawls, and maps on your own machine and works with Firecrawl's MCP server and SDKs. The post is 18 November 2026, 16:49 UTC (19 November 2026, 03:49 Sydney) and says they rebuilt the repo so agents can scrape locally. The local package licence is AGPL-3.0. Save pages into the research packet, cite each URL, and respect the site's terms of use and robots.txt. Local routes need no API key. Search, and a page the machine cannot fetch, can still go to Firecrawl Cloud.

18 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Install Modern Web Guidance 1.0 with npx modern-web-guidance@latest install, or with claude plugin marketplace add GoogleChrome/modern-web-guidance and then claude plugin install modern-web-guidance@googlechrome. Chrome for Developers says the copy in Claude Code's official plugin marketplace is not up to date, and that install is a pre-1.0 version. Check the copy you have loaded. A CLI install updates with npx modern-web-guidance@latest update.

  • A mobile visual-QA exercise on the Claude Code Desktop iOS Simulator pane, in public beta on macOS. Plans are Pro, Max, Ultra, Team, and Enterprise, except a HIPAA Enterprise organization, and it is not available with a third-party provider. The pane needs a Mac, Claude Desktop v1.24012.0 or later, and Xcode with the iOS platform. Xcode 26.x or Xcode 27 is recommended. An iPhone Duo needs Claude Desktop v4.29.0 or later and Xcode 27.1 or later (iOS 27.1 runtime for Xcode 27.1). Fold the device closed, half open, and open, then walk every screen and report state and layout bugs with screenshots. Local sessions only.

  • Paper's remote MCP is in early access, so a cloud agent can read and write a Paper file when Paper Desktop is closed. The launch post is 16 November 2026, 19:13 UTC (17 November 2026, Sydney), and its example is a CI update on every commit. Add the Paper MCP server and sign in. The older local HTTP endpoint works only while Desktop is open. Setup may change. A write stays on ask.

  • Starting in macOS 27.3, when some apps ask for Full Disk Access, you can choose specific protected folders. When some apps ask for System & App Protection Override, you can choose specific apps. When a terminal, an agent app, or a desktop agent asks and macOS offers that choice, choose the folders for Full Disk Access or the apps for System & App Protection Override. The support article was published 16 November 2026.

  • Emil Kowalski's skills pack as review gates you can load again. Install with npx skills@latest add emilkowalski/skills. The page, last updated 15 November 2026, calls the pack free and open source and heads the list Available skills 19. The gates named here are /break-ui, /remove-animations, /find-animation-opportunities, and /emil-design-eng. Emil Kowalski's post is 14 November 2026, Sydney time (13 November 2026, 17:49 UTC).

16 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Lesson note on Chrome's Modern Web Guidance 1.0, stable on the blog dated 10 November 2026, with Firefox and Edge backing it alongside Chrome. Install from the project directory with npx modern-web-guidance@latest install. search and retrieve supply a component brief. Before handoff, run npx modern-web-guidance@latest legacy-scan . on the prototype. The modernization page says the scan covers about 20% of roughly 200 guides, that each hit still needs keepWhen and dropWhen, and that there is no --fix flag. This note quotes no pass rate.

  • Lesson note on the open-source improve skill in the shadcn/improve repository. The most capable model audits and writes self-contained plans into plans/. The skill never edits source. /improve execute runs one plan in an isolated worktree and reviews the diff. Merging stays with a person. /improve quick and /improve review-plan are the cost-aware path. LICENSE.md is the MIT licence. This note has not run the skill.

  • Lesson note on the google-labs-code/stitch-skills repository. The README title is Stitch Design Skills. The plugins are stitch-design, stitch-build, and stitch-utilities, over the Stitch MCP server, from Claude Code, Cursor, Codex, and the other agents the README names. The README says this is not an officially supported Google product. An exercise generates two or three directions inside the agent, takes the chosen screen into code, and compares that with this course's wireframe test. generate-design's variantCount runs from 1 to 5 and defaults to 3, and that count is variants of one existing screen. This note does not state a skill count.

  • Further reading on Anthropic's Building always-on agents, published 13 November 2026. The shape is sleep, wake on an event, one run, persist state, and sleep. Events can be deduplicated by the source id, one worker handles that inbox, and a new worker can retry a crashed run. The article delivers to the user. This course maps that onto the scheduled reviews: the cadence is the timer, the findings file is the memory, and the named owner is this course's mapping of delivery. This note quotes no cost figure from the article.

  • Case study on licence-aware stills in Cinematic Skill. The README's example is a one-line request for a 90-second documentary, drawn in code rather than generated as video. fetch_commons.py records licence and credit, refuses a Commons file that fails the free-licence check unless --allow-nonfree is passed, and the credits command writes credits.md. A url download is not refused the same way. The README's line that credits.md is written for you is that credits command, not get by itself.

  • OpenFilm added beside Remotion and Hyperframes, labelled newer and less proven. Each shot is an HTML page with frame(t). film.html places those shots on tracks. Local Studio edits that file. npx openfilm render writes an MP4 beside it. The repository LICENSE file is MIT. The docs say Node.js 22 or later. The repository README install line says Node 22.2 or later, and ffmpeg. This note states no star count.

13 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Lesson note on a disposable sandbox with Apple's open-source container tool, v1.7.1, tagged 7 November 2026, Sydney time (6 November 2026, 23:56 UTC). Apple silicon, supported on macOS 26. The technical overview still describes limited behaviour on macOS 15, so this note does not say the tool refuses to launch there. Each container runs in its own lightweight VM. A throwaway container run mounts one project folder read-only, sets a read-only root, and caps CPUs and memory, so an agent can investigate that folder without changing it. The container-machine demo the README links to shares the home directory and is not this sandbox. This note states no start time. One line in Automation and Maintenance points an unattended or scheduled read at the same sandbox.

  • Worked example for a per-screenshot scoring pass with the Gemini Decisions API and model gemini-decision-1, on API version v1beta. The Google blog is dated 11 November 2026 and prints no clock time. One shared screenshot can take yes/no, multiple-choice, and ordered rubric questions in one pass, returning probabilities and confidence and no generated text. The course's critique dimensions become score questions ordered meets, note, warn, and block. Note never blocks. Auto-pass only a high-confidence meets, after labeled examples set the threshold. This note does not adopt the docs' sample of 0.7. Low confidence, or warn or block, goes to a person. The score is not a reason or a fix, so the explaining critic stays. Module 1 points at the example.

  • A hedged lesson note, medium confidence, on World Labs Atlas as a generated set, plus QA in the motion pipeline. The Atlas blog is dated 1 September 2026. The product page read for this note prints no opening date, so the note does not say Atlas opened to everyone on 12 November. A text prompt can generate an image. Photos or generated images can become views, a camera-path video, or an SPZ Gaussian splat. The docs say Atlas can invent uncaptured areas, that estimated geometry is not a measured dimension, and that direct video upload is not a current workflow. The cited pages do not name keyframes as an output. Blender 5.3 rendering notes: import PLY and SPZ, render in EEVEE and Cycles, no export. Pipeline and I/O notes also import Gaussian splats from glTF and USD. Apply Transform does not work correctly. Rendering is not fully correct where opacity is low and radiance is high. Do not promise a client an edit or an export those notes do not support.

12 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Lesson note on Anthropic's prompting guide for Claude Fable 5.2. Start at high effort and re-sweep rather than carrying a setting over from an earlier model. An ownership sentence changed how often duplicated code was consolidated, and it also raised tokens and unrequested changes, so the smallest edit versus a restructure is a choice. A closing summary often leaves out small changes. Prune skills, memory, and notes, and review SKILL.md and the team harness when the model changes.

  • Lesson note on opening a connector beside the chat, and on embedding a vendor site. Work in the tool yourself when you are judging it, or ask the agent for a change you approve. A Verified directory listing can allow an origin. Community listings and local servers are blocked. A custom connector needs approval in that conversation, and not for every organization. Adobe, Atlassian, and Box are some of the tools available in Claude today, not the small directory group that article says it is starting with. The article is dated 10 November 2026.

  • A balanced case from Epic's 10 Nov 2026 AEDT (9 Nov UTC) post. Unreal Engine 5.9 MegaJam Preview is described as the first Unreal Engine release built primarily with agentic workflows, by a handful of people directing fleets of agents. The preview had about a week of development. Epic also names a shorter QA pass, agent-written code still being cleaned up, and parts of the web renderer are prototypes they plan to refactor. A known issue on the same thread: Unreal MCP is on by default while its toolset plugins are off, so agents get no actor, level, asset, Blueprint, or Python tools, and no explanation. That is a fail-loudly miss, and a speed versus review-capacity trade-off.

  • A hedged lesson note on LTX-2.5 finishing tools on fal: Alpha Gen mattes, SDR-to-HDR frames, and Reframe for a new aspect ratio, including a 16:9 to 9:16 cutdown. Treat the call as a finishing step with a cost cap and a QA gate for temporal consistency and invented edges. ComfyUI merged partner nodes for these tools. This note does not run them. Re-check the pages before a budget.

  • Lesson note on the HTML-in-Canvas API, stable in Chrome 157 as of the 26 October 2026 Chrome for Developers article. Real HTML drawn into a canvas can stay selectable and translatable, and on the accessibility tree. The article's 2D example is drawElementImage(). WebGL and WebGPU use the same setup and a different draw step. Three.js support is experimental. Babylon.js, PlayCanvas, and PixiJS v8.19.0 are named, as third-party libraries. The handoff says to feature-detect, name a fallback, and treat the path as Chrome-only until another browser is checked. The article does not survey other browsers.

10 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Lesson note on Astryx as a worked case study of a design system an agent can read. MIT licence, in beta, on React 19 or later, with 150+ accessible components. A command writes agent context files. A hosted server exposes two tools, search and get. A theme is CSS custom property overrides. An unrecognized transport can look connected and expose no tools. The pages read for this note do not print an announcement date.

  • Lesson note on the Elephant-Goldfish plugin. The elephant holds project context and drafts a spec. Goldfish reviewers use a fresh context that sees the document and its sources, not the earlier conversation, to expose unstated assumptions. A person resolves trade-offs and approves. If a fresh context cannot be opened, the review stops. This note does not ask anyone to install or run the plugin. The README says it is not an officially supported Google product.

  • Lesson note on an Xcode entry added to Anthropic's official plugin marketplace on 6 November 2026. The entry lists Apple Inc. as author. Apple's documentation does not announce that plugin and does not name a minimum Xcode version. Apple's documented route still needs the project open in Xcode and external agents allowed in Xcode settings. This note does not say you can leave Xcode closed.

9 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

  • Lesson note on four built-in output styles besides Default. Concise, from Claude Code v2.1.237, shortens narration. Its instructions keep errors, security warnings, and destructive confirmations in full. A style is an instruction, not a guarantee. Learning and Explanatory are the pair for a codebase you are still learning.

  • Lesson note on Cloudflare's security-audit skill as a six-phase critic pattern, from the current README. The 18 June 2026 blog described seven phases and a later fleet harness. Reports come from verified records. Without the full sandbox, a lead stays needs_validation. This note does not ask anyone to install or run the skill, and it is not a security certification.

  • Lesson note on the 6 November 2026 Chat SDK integration: a typed catalog constrains the JSON an agent may render. That date is not a json-render launch date. Schema validity is not accessibility or design quality. Adapters may degrade over-limit content.

  • Lesson note on OpenCode Browse as a real signed-in browser session, with a per-session site allowlist and a human handoff for login, 2FA, and payment. Signed-in privileges, not copied credentials. This note does not ask anyone to install the extension or open a session.

  • Lesson note on one developer's 6 November 2026 measurement of a 137k-token session: lossy summaries, a checkpoint fork, and a fresh session. His better paths were in series with retrieval, or fresh parallel sessions. Not a universal result.

  • A reading from Linear's chief executive on ambition and operating model. Vendor perspective, not a study. The sentence on screenshots, videos, and a preview build is unfinished in the source and is not completed here.

7 Nov 2026

Approval date for these lesson notes. This page records the notes. It is not a deploy record.

5 Nov 2026

  • Google's gemini-api-cli skill, as a worked example of a vendor-published skill an agent can load.

  • A reading: Addy Osmani on critical thinking when output is mostly correct and stops getting reviewed.

  • A short section on agent state machines, beside fan-out, pipelines, and critic pairs.

  • The idea of scoring past conversations and proposing skill fixes as reviewable changes.

  • The authors' reported finding that, once given tools, models fail to refuse harmful requests more often, up to 68.7% more, and a case for narrow tool access.

  • A note on running the agent in a dedicated workspace, with the least access the task needs.

  • A sidebar on recorded-footage tools: PySceneDetect, Auto-Editor, vid.stab, and rembg.