Ultra Lab
cat /blog/*

Ultra Lab Blog

Social automation, SaaS development, and AI integration, with practical guides and deep analysis

🎓 New to AI? Start Here

Free tutorial tracks, zero coding required. Got a computer? Take the desktop track. Phone only? That works too; read in order.

🎓Beginner Onramp · Free AI Tutorials
🛡️AI Security & Defense
📊Solo Founder Build Log
ClaudeClaude CodeAI VideoAI DevelopmentField Notes

Claude Fable 5.2 wrote the code that renders a 110-second cat's-eye film, with no video generation model

The awesome-fable-5.2 repo on GitHub has a 110-second, single-shot film seen through a cat's eyes. Claude Fable 5.2 wrote about 5000 lines of code in Claude Code, including a path tracer, a gait and gaze simulation for the cat's head, focus that follows whatever the cat stares at, synthesised sound and a QC pass, with no AI image generation, photos or assets. We go through what the README says it did, why this is a different thing from video generation, how it compares with the films we render from code ourselves, and where the limits are.

· 16 min read
Google UpdateAI ModeAI OverviewsGEOAEOSEOSearch Console

Google Search now runs on Gemini 4 Hydrogen: what actually changes for SEO, GEO and AEO

Google put Gemini 4 Hydrogen into Search on November 12, 2026, for all users worldwide. The parts that matter for site owners are a stronger query fan-out and more comparison tables and images inside answers. We checked every claim against Google's announcement and its AI optimization guide: five changes for SEO, GEO and AEO, why llms.txt does nothing for Google, why you should not read your charts before November 24, and five things worth doing this week.

· 11 min read
Claude CodeMCPAI developmenttutorialfield notes

How to connect MCP servers to Claude Code (2026): the command, .mcp.json and remote OAuth

There are three ways to connect an MCP server to Claude Code: one claude mcp add command, a .mcp.json file in the project root, or a remote server you sign in to with OAuth. Based on the official docs and our own test run: adding stdio and HTTP servers, where the local, project and user scopes are stored, checking status and re-authenticating in /mcp, and six common errors, three of which we reproduced.

· 12 min read
Claude CodeClaudeAI developmenttutorialfield notes

Claude Fable 5.2 is here: how to upgrade Claude Code and check which model you are running

Anthropic released Claude Fable 5.2 on November 9, 2026. Paid plans in the claude.ai app offered it the same day, but Claude Code only recognizes it from version 2.1.321. Based on the official docs and our own early-morning test run (Taipei time): how to check your version, upgrade, confirm where the fable alias points, read the unrecognized_model line, and upgrade a setup with one coordinator session and several overnight workers without interrupting anything.

· 8 min read
MindThreadMCPClaude CodeThreadscontent strategy

Connect MindThread to Claude: pull your best posts first, then draft

Connect the MindThread MCP server to Claude Code, use get_top_posts to pull the 20 best-performing posts from your own Threads account, have Claude turn them into a content document (core ideas, voice, worldview), then draft 3 new posts from it. Both prompts, the output of every step of our test run, a read-only key setup, and the error we actually ran into.

· 9 min read
AI AgentLLMpermissionsMindThreadFirestore

Letting an AI assistant act without overreaching: action cards, a confirm gate and risk levels

In the upcoming MindThread 2.0 (on our preview site for now), we let Xiaomai, the in-app assistant, schedule posts, change posting times, toggle cloud patrol and pause accounts for members. The model never touches a tool: rules pick the intent, the model only fills card fields, the server stores the card, every card needs a click, high-risk ones need two, state is re-read before running, and everything is logged. Includes a confirm gate you can adapt.

· 10 min read
AI agentsLLMhallucinationPrompt InjectionMindThread

Making an AI assistant quote only real numbers: the faithfulness check behind our in-app helper

The worst thing an in-app AI assistant can do is get a number wrong. For Xiaomai, the helper in MindThread 2.0, we built four layers: the model never gets tools, data is precomputed, every number and quote in the answer is checked against the data, and anything that does not match triggers one retry, then a plain "not found". Includes a ready-to-use answerIssues function.

· 10 min read
AI writingoutreachLLMMindThreadquality checks

AI-written outreach should read like the recipient wrote it: the quality bar we set for samples

We are designing a 'send a sample first' outreach flow: three Threads drafts written with MindThread in the recipient's own voice, or three findings from their website. The first version averaged 5.7 out of 10. This post covers the four gates we set: how we build the brand brief, computing voice with code, a source for every factual sentence, and a reviewer playing the business owner who scores the email, with anything at 7 or below rewritten. We have not sent a single email; this is a process design. Voice stats and overlap check code included.

· 10 min read
DependabotnpmGitHubsupply chain security自動化open source

Dependabot alerts across 49 repos in one night: inventory, batch fixes, verification

49 repos, 1,230 Dependabot alerts, one night: inventory and batch fixes, with critical and high confirmed at zero on GitHub for 5 public repos after merging. gh api inventory, 403 means scanning is off, baseline before fixing, no --force, overrides instead of downgrades, re-check after merge. Scripts included.

· 11 min read
GeminiLLMmodel migrationMindThreadfallback chain

When the model deprecation notice lands: how we split the gemini-2.5-flash migration into four steps

Google says gemini-2.5-flash shuts down on 2026-12-01 and requests will be redirected to gemini-3.5-flash-lite. It sounds like nothing to do, but when we called 3.5-flash-lite directly with a thinkingBudget, we got a 400. We split the migration into four steps: inventory, a tested parameter matrix, per-feature switching through config, and a quality comparison on real inputs. Along the way we found two links in our fallback chain that had been broken for a while. Probe and switching code included.

· 11 min read
GitWindowsNode.jsnpm自動化

git worktree plus a junctioned node_modules: how we broke our shared dependencies, and how we stop it now

On Windows, if a git worktree's node_modules is a mklink /J junction back to the main checkout, git worktree remove follows the junction and empties the main copy. We hit this four times between early September and mid October. Retesting this week: it happens without --force too, and running npm ci inside the junction also empties the main copy. Includes our cleanup order and a check script.

· 12 min read
Gitmerge conflictsbranchingMindThreadengineering workflow

Merging 106 commits into a branch that had been open for two and a half weeks: how we sorted and resolved 44 conflicts

The MindThread 2.0 branch went a week without syncing, mainline grew by 106 commits, and the merge stopped on 44 conflicted files. We sorted them into four groups (shared code, old-only, new-only, config), gave each group a default rule, then ran two test stages, held the bundle budget and walked the preview site. Includes a conflict sorting script and a by-key three-way JSON merge script.

· 13 min read
Claude CodeAI agentsmulti-agentWindowsPowerShell自動化

We ran 10 Claude Code agents at once and froze the PC. Here is how we schedule them now

Ten headless Claude Code agents, each running npm ci, a build and headless Chrome, left a 96 GB desktop unable to start cmd.exe. Our rules now: 6 at once, 2-minute stagger, a resource guard that sets a pause flag, workers clean up after themselves.

· 11 min read
VercelCrondeployment自動化

Vercel cron only runs on the production deployment: how alias set left our schedules two months behind

The site was on the new version, but a newly added Vercel cron never ran. The project had auto-assign of production domains turned off, so vercel --prod plus alias set only moved the domain. Vercel's production deployment was still two months old, and the cron jobs stayed with it. How to check via the API, how to promote, and the side effect of auto-assign coming back on.

· 8 min read
WindowsTask Schedulermonitoring自動化VBScriptNode.js

Exit code 0 is not success: how we check 50+ Windows scheduled tasks for real

Task Scheduler showed 0x0 while our Telegram watchdog actually exited 255 and never restarted anything. We stopped trusting result codes: every task writes its own start and end markers, a checker reads them, and no success within the time limit means red. 52 tasks covered, 1 left. Includes a VBS wrapper, the cmd errorlevel trap and a Node checker skeleton.

· 14 min read
Threadssocial mediaviral formulaMindThreadfield test

Threads Formula 30-Day Test, Wrap-Up: 13 AI-Written Posts, One Hit 100K Views, and the AI-vs-Handwritten Question We Can't Answer

The wrap-up of our 30-day stance-formula test. All 13 formula posts were AI-generated, 10 of them landed below the median of the same accounts' everyday posts in the same window, and the formula missed the +50% bar we set in advance. One post on day 28 reached 108,000 views, but two weeks later the same account's non-formula post on the same topic (the cost of smoking) reached 119,000. The opening post's third question, AI-generated versus handwritten, can't be answered: the experiment never had a handwritten control group. This post explains why.

· 15 min read
Threads automationThreads auto replycomment patrolMindThread

Threads Auto Reply and Comment Patrol: How to Catch the First Few Minutes (2026)

Threads auto reply answers comments on your own posts through the official API; comment patrol comments on other people's posts. Risks and a five-step setup.

· 12 min read
PHPPHP2ShellCVE-2026-109618CVE-2026-109619SecurityWordPressWebsite SecurityAI 安全

What Is PHP2Shell? The PHP Flaws Under 70% of the Web, and the Three Things Your Site Needs Now

What is PHP2Shell? Two PHP flaws (CVE-2026-109618, 109619, CVSS 9.2) that may allow unauthenticated code execution, fixed in 8.5.12, 8.4.27, 8.3.36, 8.2.35.

· 21 min read
AI agentsClaude Codecostone-person companybuild in public

7 months after my "$0/month AI agents" post: what it costs now

What AI agents cost a one-person company 7 months after the $0/month Gemini setup: two Claude subscriptions at about $400 a month, MindThread MRR, open issues.

· 9 min read
Node.jsfile lockingGPUClaude CodeAI agents自動化

Several Claude sessions, one GPU: a cross-process render lock in a single file

Several Claude sessions, one GPU: a 183-line Node file lock with atomic 'wx' create, queue not skip, heartbeat, stale-lock cleanup, and first-night results.

· 8 min read
ClaudeAI agentscostblind testingvideo automationbuild in public

Three experiments to cut our AI video pipeline's cost: one passed, one passed with conditions, one failed

We classified every turn of our AI Shorts pipeline's Claude usage and found the money was in rework, not writing. Then we ran three experiments: record the voice first (US$0.76 down to US$0.17 per video), lower the reviewer's reasoning effort (medium saves about 30%, low lets real problems through), and draft with Sonnet (34% cheaper, factual errors 4 to 1, not adopted). Includes how we seeded errors for a blind test and the mistake we made ourselves.

· 9 min read
AI agentsClaude CodeYouTube Shorts自動化costbuild in public

Letting agents produce our YouTube Shorts: about 250k tokens and $3 a video, from four real runs

We turned our Shorts pipeline into a weekly unattended run: pick topics, render, mechanical checks, AI review, schedule. A human replies to one weekly report. Here is the token ledger from four real runs: 514k for the first video, 254k and a first-round pass for the fourth after we added mechanical checks before review, and why the savings came from fewer redos rather than a cheaper model.

· 8 min read
Prompt InjectionAI 安全OWASPMCPUltraProbe

What Is Prompt Injection? How It Reaches an AI Assistant Through Documents and Tools, and Three Layers of Defense (2026)

What is prompt injection? How hidden instructions reach an AI assistant through documents, web pages and tools, and three layers of defense.

· 8 min read
AI VisibilityWhite LabelAgenciesAVSUltraGrowth

Should an Agency White-Label AI Visibility? Four Paths and 8 Questions Before You Sign (2026)

When clients start asking about AI visibility, web and marketing agencies have four paths. This post compares cost, client ownership and delivery responsibility, lists 8 questions to ask before signing a white-label deal, and shows how to explain what the score is and is not. Written by Ultra Lab, interests disclosed.

· 8 min read
RSS

ULTRA LAB WEEKLY

Every Monday: GitHub's top 10 AI tools, plus one thing we learned

Each tool is labelled with who it suits and whether the licence allows commercial use. Stuck? Ask in our Discord.

Only used for this weekly email, never shared, one-click unsubscribe. Issues are written in Traditional Chinese.

Join Discord