https://mk-comics.vercel.app/blog/ Weekly posts on the AI models and tools that matter to people shipping products. en https://mk-comics.vercel.app/blog/building-a-motion-comic-intro-page-curls-trailer-score-autoplay/ https://mk-comics.vercel.app/blog/building-a-motion-comic-intro-page-curls-trailer-score-autoplay/ Tue, 13 Oct 2026 15:00:00 GMT How the site's opening became a motion comic: rejected loaders, a 14-strip CSS page curl, recorded music and effects timed to their peaks, and why it waits for a click. Web AudioCSS 3Dautoplay policymotion designAstrocase file https://mk-comics.vercel.app/blog/chasing-100-on-pagespeed-comic-portfolio-case-file/ https://mk-comics.vercel.app/blog/chasing-100-on-pagespeed-comic-portfolio-case-file/ Sun, 11 Oct 2026 09:30:00 GMT Twelve deploys, one score at a time: contrast fixes, llms.txt, a Tamil font stall worth 600 ms, content-visibility, and the Lighthouse noise that caps mobile at 99. LighthousePageSpeed InsightsCore Web VitalsaccessibilityAstroGSAPllms.txtcase file https://mk-comics.vercel.app/blog/hiring-manager-seo-what-a-portfolio-should-rank-for/ https://mk-comics.vercel.app/blog/hiring-manager-seo-what-a-portfolio-should-rank-for/ Sat, 10 Oct 2026 02:40:00 GMT A portfolio does not need to rank for 'software engineer'. It needs to win your own name, then earn long-tail pages that prove the claim. What Google's guides say to do. SEOportfoliocareersstructured dataGoogle Search https://mk-comics.vercel.app/blog/measuring-core-web-vitals-on-an-animation-heavy-page/ https://mk-comics.vercel.app/blog/measuring-core-web-vitals-on-an-animation-heavy-page/ Sat, 10 Oct 2026 02:39:00 GMT LCP, INP and CLS on a page full of GSAP and Lenis scroll animation: what each metric counts, which patterns it ignores, and a PageSpeed workflow that does not mislead. Core Web VitalsperformanceGSAPLenisPageSpeed InsightsAstro https://mk-comics.vercel.app/blog/build-time-data-github-contributions-with-fallback/ https://mk-comics.vercel.app/blog/build-time-data-github-contributions-with-fallback/ Sat, 10 Oct 2026 02:38:00 GMT How this site fetches its GitHub contribution calendar in the Astro build, falls back to a committed snapshot, renders SVG, and a parser bug I found in my own code. AstroGitHubbuild-time dataGraphQLSVG https://mk-comics.vercel.app/blog/gsap-and-lenis-scroll-choreography-without-hiding-content/ https://mk-comics.vercel.app/blog/gsap-and-lenis-scroll-choreography-without-hiding-content/ Sat, 10 Oct 2026 02:37:00 GMT How to animate on scroll with GSAP ScrollTrigger and Lenis without pre-hiding content, with reduced-motion handling, using this site's real script and its one flaw. GSAPLenisScrollTriggeranimationaccessibilityreduced motion https://mk-comics.vercel.app/blog/comic-book-portfolio-design-decisions-in-public/ https://mk-comics.vercel.app/blog/comic-book-portfolio-design-decisions-in-public/ Sat, 10 Oct 2026 02:36:00 GMT Why this portfolio is drawn as a comic issue: the audience, the worlds it replaced, the DESIGN.md rules, and what a hiring manager sees on the first screen. designportfolioAstrodesign systemDESIGN.mdGSAP https://mk-comics.vercel.app/blog/agentic-coding-on-a-legacy-codebase-opus-5-5/ https://mk-comics.vercel.app/blog/agentic-coding-on-a-legacy-codebase-opus-5-5/ Sat, 10 Oct 2026 02:35:00 GMT Anthropic and early testers report Opus 5.5 running audits and migrations for hours. What is reported, what is unverified, and a review discipline for legacy code. Claude Opus 5.5legacy codecoding agentscode migrationcode reviewagentic workflows https://mk-comics.vercel.app/blog/skills-and-plugins-for-coding-agents/ https://mk-comics.vercel.app/blog/skills-and-plugins-for-coding-agents/ Sat, 10 Oct 2026 02:34:00 GMT What Claude Code skills and plugins are, where they live, what they cost in context, and how the impeccable plugin keeps design rules out of this site's prompts. Claude Codeskillspluginscoding agentsimpeccabledeveloper workflow https://mk-comics.vercel.app/blog/writing-a-claude-md-that-actually-helps/ https://mk-comics.vercel.app/blog/writing-a-claude-md-that-actually-helps/ Sat, 10 Oct 2026 02:33:00 GMT What belongs in a CLAUDE.md and what does not, from Anthropic's docs, with a 30-line example for a real Astro repo and the rules, imports and hooks that replace the rest. Claude CodeCLAUDE.mddeveloper experiencecontextAstro https://mk-comics.vercel.app/blog/claude-code-for-a-solo-engineer-the-workflow/ https://mk-comics.vercel.app/blog/claude-code-for-a-solo-engineer-the-workflow/ Sat, 10 Oct 2026 02:32:00 GMT How this portfolio was built with Claude Code: the plan and task files, per-section commits, the impeccable plugin, and what the repo shows did and did not work. Claude CodeworkflowAstroimpeccablesolo engineer https://mk-comics.vercel.app/blog/prompt-caching-for-coding-agents-stable-context/ https://mk-comics.vercel.app/blog/prompt-caching-for-coding-agents-stable-context/ Sat, 10 Oct 2026 02:31:00 GMT How Claude Code's prefix cache works, what breaks it mid-session (model switches, MCP changes, compaction), and the arithmetic of a 100k-token session on Sonnet 5.5. Claude Codeprompt cachingcoding agentsClaude Sonnet 5.5cost optimisation https://mk-comics.vercel.app/blog/ga4-gtm-clarity-on-a-static-site-without-wrecking-lighthouse/ https://mk-comics.vercel.app/blog/ga4-gtm-clarity-on-a-static-site-without-wrecking-lighthouse/ Sat, 10 Oct 2026 02:30:00 GMT Env-var-gated GA4, Tag Manager and Clarity on an Astro site, with a measured PageSpeed run: the third-party bill, its effect on blocking time, and what I would cut. Google AnalyticsTag ManagerClarityLighthouseAstroperformance https://mk-comics.vercel.app/blog/blog-in-astro-content-collections-schema-rss-og/ https://mk-comics.vercel.app/blog/blog-in-astro-content-collections-schema-rss-og/ Sat, 10 Oct 2026 02:29:00 GMT How this site's blog is built on Astro 7: a Zod-checked posts collection, hand-written RSS and satori OG images. Real code, plus where it departs from the docs. Astrocontent collectionsRSSsatoriOpen Graphstatic sites https://mk-comics.vercel.app/blog/open-graph-images-generated-at-build-with-satori/ https://mk-comics.vercel.app/blog/open-graph-images-generated-at-build-with-satori/ Sat, 10 Oct 2026 02:28:00 GMT How this blog makes one 1200x630 share image per post at build time with satori and resvg in an Astro endpoint: fonts, layout limits, caching traps and the code. Open GraphsatoriAstroresvgsocial cards https://mk-comics.vercel.app/blog/structured-data-for-a-personal-site-person-article-schema/ https://mk-comics.vercel.app/blog/structured-data-for-a-personal-site-person-article-schema/ Sat, 10 Oct 2026 02:27:00 GMT The Person and Article JSON-LD this site ships, what Google says it uses, three things I would fix, and how to validate it with Google's and schema.org's tools. structured dataJSON-LDschema.orgSEOAstro https://mk-comics.vercel.app/blog/llms-txt-what-it-is-who-reads-it-template/ https://mk-comics.vercel.app/blog/llms-txt-what-it-is-who-reads-it-template/ Sat, 10 Oct 2026 02:26:00 GMT llms.txt is a Markdown index for AI agents. Google says Search ignores it, log studies show few fetches, but it is cheap. The spec, the evidence and a template. llms.txtSEOAI visibilityAI crawlersMarkdown https://mk-comics.vercel.app/blog/ai-crawler-guide-robots-llms-txt-structured-data/ https://mk-comics.vercel.app/blog/ai-crawler-guide-robots-llms-txt-structured-data/ Sat, 10 Oct 2026 02:25:00 GMT Which robots.txt tokens decide whether OpenAI, Anthropic, Perplexity and Google can train on your site or cite it, from each vendor's docs, plus this site's file. SEOAI crawlersrobots.txtGPTBotClaudeBotAI visibility https://mk-comics.vercel.app/blog/review-checklist-for-ai-generated-code/ https://mk-comics.vercel.app/blog/review-checklist-for-ai-generated-code/ Sat, 10 Oct 2026 02:24:00 GMT What to check when the author is an agent: a nine-point review checklist, the published numbers on AI-written code, and a REVIEW.md you can copy. code reviewAI-generated codeClaude Codepull requestsREVIEW.md https://mk-comics.vercel.app/blog/sole-engineer-on-a-1-25m-user-platform-what-i-automated-first/ https://mk-comics.vercel.app/blog/sole-engineer-on-a-1-25m-user-platform-what-i-automated-first/ Sat, 10 Oct 2026 02:23:00 GMT EF Academy's marketing platform serves 1.25M+ monthly users in 23+ languages with one engineer. How I rank automation: bottlenecks first, silent failures next, licences. solo engineerautomationownershipStoryblokAWS Glue https://mk-comics.vercel.app/blog/chrome-extensions-that-save-admins-hours/ https://mk-comics.vercel.app/blog/chrome-extensions-that-save-admins-hours/ Sat, 10 Oct 2026 02:22:00 GMT How Manifest V3 extensions for BSBI and GISMA automated student application workflows in tools we could not change, and the service worker rules that decide reliability. Chrome extensionautomationManifest V3PDF.jsservice workers https://mk-comics.vercel.app/blog/ai-document-processing-17-types-at-95-percent/ https://mk-comics.vercel.app/blog/ai-document-processing-17-types-at-95-percent/ Sat, 10 Oct 2026 02:21:00 GMT How GPT-4 Vision and OCR classified 17+ admissions document types at 95%, and what a number like that must include: per-type results, confidence routing, review paths. document processingGPT-4 VisionOCRevaluationstructured outputs https://mk-comics.vercel.app/blog/leading-four-engineers-to-react-react-native-parity/ https://mk-comics.vercel.app/blog/leading-four-engineers-to-react-react-native-parity/ Sat, 10 Oct 2026 02:20:00 GMT How Edvanza shipped each module to web and mobile together with a four-person front-end team: shared logic, platform splits, review discipline and a parity checklist. leadershipReact Nativeteamcode reviewReact https://mk-comics.vercel.app/blog/real-time-iot-status-with-signalr/ https://mk-comics.vercel.app/blog/real-time-iot-status-with-signalr/ Sat, 10 Oct 2026 02:19:00 GMT Why Planet SIM and Planet Business push live utility and water-service status over SignalR, and what the docs say about reconnects, keep-alives, tokens and scale. SignalRreal-timeIoTdashboardsFirebase https://mk-comics.vercel.app/blog/100-lighthouse-on-astro-for-a-real-marketing-site/ https://mk-comics.vercel.app/blog/100-lighthouse-on-astro-for-a-real-marketing-site/ Sat, 10 Oct 2026 02:18:00 GMT How the Incresco and Camped sites reached 100% Lighthouse on Astro and Storyblok, and what Astro's docs say about images, fonts and islands for the same goal. AstroLighthouseperformanceStoryblokpage builders https://mk-comics.vercel.app/blog/prospect-uploader-10000-leads-per-csv/ https://mk-comics.vercel.app/blog/prospect-uploader-10000-leads-per-csv/ Sat, 10 Oct 2026 02:17:00 GMT Why EF Academy's lead uploads moved from direct CSV imports to a Lambda and EventBridge pipeline with multilingual validation, email alerts and a Prospect Viewer. AWS LambdaEventBridgeSalesforcedata validationserverless https://mk-comics.vercel.app/blog/23-language-marketing-site-nextjs-storyblok-sole-engineer/ https://mk-comics.vercel.app/blog/23-language-marketing-site-nextjs-storyblok-sole-engineer/ Sat, 10 Oct 2026 02:16:00 GMT EF Academy's marketing site serves 1.25M+ monthly users in 23+ languages with one engineer. The decision that mattered: let content teams run localisation themselves. Next.jsStorybloki18nlocalisationsole engineer https://mk-comics.vercel.app/blog/replacing-tibco-scribe-with-aws-glue-before-the-licence-ran-out/ https://mk-comics.vercel.app/blog/replacing-tibco-scribe-with-aws-glue-before-the-licence-ran-out/ Sat, 10 Oct 2026 02:15:00 GMT How I moved EF Academy's Salesforce pipelines off a $24,000-a-year Tibco Scribe licence onto AWS Glue before it expired: the options, the plan, the cutover, $840. AWS GlueTIBCOmigrationcostSalesforce https://mk-comics.vercel.app/blog/text-to-speech-product-tours-gemini-tts-nextjs/ https://mk-comics.vercel.app/blog/text-to-speech-product-tours-gemini-tts-nextjs/ Sat, 10 Oct 2026 02:14:00 GMT Generate product-tour narration with Gemini 3.8 Flash TTS at build time, cache the audio by content hash, and ship it accessibly. Cost maths and a copyable script. Gemini TTSNext.jsaccessibilitytext to speechbuild-time generation https://mk-comics.vercel.app/blog/voice-agents-with-gemini-live-turn-taking-barge-in/ https://mk-comics.vercel.app/blog/voice-agents-with-gemini-live-turn-taking-barge-in/ Sat, 10 Oct 2026 02:13:00 GMT How the Gemini Live API handles turns, interruptions and sessions, which latency knobs the docs give you, and a minimal browser client with ephemeral tokens. voice agentsGemini LiveWebRTCbarge-inWebSocketephemeral tokens https://mk-comics.vercel.app/blog/hosted-apis-vs-open-weights-for-a-product-team/ https://mk-comics.vercel.app/blog/hosted-apis-vs-open-weights-for-a-product-team/ Sat, 10 Oct 2026 02:12:00 GMT Data residency, cost at volume, who is on call, quality and speed of change: five questions that decide hosted API or open weights, with a break-even sketch to copy. open weightsself-hostingdecisiondata residencyLLM pricingvLLM https://mk-comics.vercel.app/blog/model-migration-playbook-sonnet-4-5-to-5-5/ https://mk-comics.vercel.app/blog/model-migration-playbook-sonnet-4-5-to-5-5/ Sat, 10 Oct 2026 02:11:00 GMT Sonnet 4.5 retires on 30 November 2026. What breaks, what the bill does once tokens and thinking change, how to eval the move, and how to roll back. migrationClaude Sonnet 5.5evalsprompt cachingcost optimisation https://mk-comics.vercel.app/blog/ai-feature-flags-rolling-out-model-upgrades/ https://mk-comics.vercel.app/blog/ai-feature-flags-rolling-out-model-upgrades/ Sat, 10 Oct 2026 02:10:00 GMT A model ID is a config value that behaves like a code deploy. Bucket by conversation, shadow first, watch tokens and refusals, and keep the old model until retirement. feature flagsrolloutmodel upgradesOpenFeatureClaude Sonnet 5.5observability https://mk-comics.vercel.app/blog/prompt-injection-through-tool-results-and-documents/ https://mk-comics.vercel.app/blog/prompt-injection-through-tool-results-and-documents/ Sat, 10 Oct 2026 02:09:00 GMT Indirect prompt injection arrives inside tool results, emails and OCR text, not from your user. What OWASP, Anthropic and Willison advise, plus a screening call to copy. prompt injectionsecurityagentsOWASPtool useleast privilege https://mk-comics.vercel.app/blog/red-teaming-your-own-ai-feature-in-an-afternoon/ https://mk-comics.vercel.app/blog/red-teaming-your-own-ai-feature-in-an-afternoon/ Sat, 10 Oct 2026 02:08:00 GMT A four-hour plan, a 30-case attack set and a small Python harness with canary tokens, mapped to the OWASP LLM Top 10 and Anthropic's guidance on prompt injection. red teamingsecurityprompt injectionOWASPClaude Haiku 5.5Python https://mk-comics.vercel.app/blog/testing-llm-features-snapshot-tests-eval-sets-ci/ https://mk-comics.vercel.app/blog/testing-llm-features-snapshot-tests-eval-sets-ci/ Sat, 10 Oct 2026 02:07:00 GMT What to run per pull request and what to run nightly, plus how to cap the bill: snapshot the prompt, not the answer, and gate on a small eval set. With GitHub Actions. testingCIevalsGitHub ActionsLLMClaude Haiku 5.5 https://mk-comics.vercel.app/blog/caching-model-outputs-when-to-memoize-an-answer/ https://mk-comics.vercel.app/blog/caching-model-outputs-when-to-memoize-an-answer/ Sat, 10 Oct 2026 02:06:00 GMT Prompt caching does not cache answers. When memoizing the output itself pays, what to put in the key, how to pick a TTL, and a Python wrapper for Claude Sonnet 5.5. cachingmemoizationLLMprompt cachingClaude Sonnet 5.5Python https://mk-comics.vercel.app/blog/rate-limits-and-backoff-across-llm-apis/ https://mk-comics.vercel.app/blog/rate-limits-and-backoff-across-llm-apis/ Sat, 10 Oct 2026 02:05:00 GMT How Claude, Gemini and OpenAI document rate limits, 429s and retry signals, and a shared TypeScript client with jittered backoff that knows which errors never to retry. rate limitsbackoffTypeScriptClaude APIGemini APIOpenAI API https://mk-comics.vercel.app/blog/event-driven-ai-eventbridge-retries-idempotency/ https://mk-comics.vercel.app/blog/event-driven-ai-eventbridge-retries-idempotency/ Sat, 10 Oct 2026 02:04:00 GMT EventBridge and Lambda retry by design, so a model call in the pipeline will run twice. Where duplicates come from, what stacked retries cost, and an idempotent handler. EventBridgeidempotencyAWSAWS LambdaLLMretries https://mk-comics.vercel.app/blog/aws-glue-for-ai-data-pipelines-lessons-from-tibco/ https://mk-comics.vercel.app/blog/aws-glue-for-ai-data-pipelines-lessons-from-tibco/ Sat, 10 Oct 2026 02:03:00 GMT What replacing a $24,000-a-year TIBCO Scribe licence with AWS Glue taught me, how Glue pricing really works, and how Glue jobs prepare data for model pipelines. AWS Gluedata pipelinesTIBCOSalesforcecost optimisationETL https://mk-comics.vercel.app/blog/serverless-llm-pipelines-aws-lambda-step-functions/ https://mk-comics.vercel.app/blog/serverless-llm-pipelines-aws-lambda-step-functions/ Sat, 10 Oct 2026 02:02:00 GMT Lambda's 15-minute cap is rarely the real limit for LLM work. Queue sizing, partial batch failures, Step Functions retries and idempotency for pipelines that call models. AWS LambdaStep FunctionsSQSLLMidempotencyserverless https://mk-comics.vercel.app/blog/astro-for-ai-content-sites-static-first/ https://mk-comics.vercel.app/blog/astro-for-ai-content-sites-static-first/ Sat, 10 Oct 2026 02:01:00 GMT Why an AI content site should ship static HTML: Astro 7 content collections, islands, build-time model calls and an honest look at the JavaScript this site ships. Astrostatic sitesperformancecontent collectionsislandsbuild-time AI https://mk-comics.vercel.app/blog/nextjs-and-ai-server-actions-streaming-caching/ https://mk-comics.vercel.app/blog/nextjs-and-ai-server-actions-streaming-caching/ Sat, 10 Oct 2026 02:00:00 GMT Where AI calls belong in Next.js 16.4: Route Handlers for streaming, Server Actions for mutations, Cache Components for data, and what to cache around the model itself. Next.jsstreamingAI SDKServer ActionsCache ComponentsVercel https://mk-comics.vercel.app/blog/ai-in-salesforce-integrations-enrichment/ https://mk-comics.vercel.app/blog/ai-in-salesforce-integrations-enrichment/ Sat, 10 Oct 2026 01:59:00 GMT Let an LLM enrich Salesforce records safely: external-ID upserts, schema-checked output, confidence gates and a human review lane, with Prospect Uploader lessons. SalesforceintegrationAIupsertdata validationstructured outputs https://mk-comics.vercel.app/blog/playwright-and-ai-cutting-automation-latency-40-percent/ https://mk-comics.vercel.app/blog/playwright-and-ai-cutting-automation-latency-40-percent/ Sat, 10 Oct 2026 01:58:00 GMT Where Playwright automation time goes, the documented levers (state reuse, waiting, blocking, parallelism), and where a model helps with selectors and where it hurts. Playwrightautomationlatencyflaky testsAI agentsadmin tooling https://mk-comics.vercel.app/blog/automating-admin-work-with-ai-the-3-5-hours-pattern/ https://mk-comics.vercel.app/blog/automating-admin-work-with-ai-the-3-5-hours-pattern/ Sat, 10 Oct 2026 01:57:00 GMT How to find the hours worth automating in a browser workflow, what 99.9% uptime means for an extension, and where AI belongs. From case file 08, with a worked example. automationadmin toolingChrome extensionManifest V3reliability https://mk-comics.vercel.app/blog/manifest-v3-and-ai-service-workers-offscreen-documents/ https://mk-comics.vercel.app/blog/manifest-v3-and-ai-service-workers-offscreen-documents/ Sat, 10 Oct 2026 01:56:00 GMT Idle shutdown after 30 seconds, a 5-minute event cap, one offscreen document, runtime messaging only: the MV3 limits that decide where an AI call can run, with code. Manifest V3Chrome extensionservice workersoffscreen documentsAI features https://mk-comics.vercel.app/blog/chrome-extension-that-calls-claude-safely/ https://mk-comics.vercel.app/blog/chrome-extension-that-calls-claude-safely/ Sat, 10 Oct 2026 01:55:00 GMT Why an API key cannot live in a Chrome extension, the proxy that should hold it, what the Web Store requires before page content leaves the browser, and 429 handling. Chrome extensionManifest V3Claude APIAPI keysrate limitsprivacy https://mk-comics.vercel.app/blog/fraud-signals-in-uploaded-documents/ https://mk-comics.vercel.app/blog/fraud-signals-in-uploaded-documents/ Sat, 10 Oct 2026 01:54:00 GMT A vision model reads a payslip well but never sees its PDF metadata. Which fraud checks belong in the model, which in code, with a Claude API example. fraud detectiondocument processingClaude APIPDFstructured outputvision models https://mk-comics.vercel.app/blog/cost-per-decision-framework-for-classification/ https://mk-comics.vercel.app/blog/cost-per-decision-framework-for-classification/ Sat, 10 Oct 2026 01:53:00 GMT A formula for classification cost: tokens, cache, batch, human review and error cost, filled with published Haiku 5.5, Sonnet 5.5 and Gemini prices, no invented accuracy. costclassificationHaiku 5.5Gemini 3.8 FlashLLM pricingevaluation https://mk-comics.vercel.app/blog/ocr-plus-vlms-document-processing-then-and-now/ https://mk-comics.vercel.app/blog/ocr-plus-vlms-document-processing-then-and-now/ Sat, 10 Oct 2026 01:52:00 GMT The 2024 GPT-4 Vision plus OCR pipeline at 95% across 17+ document types, and what current Claude and Gemini document APIs change, with per-page cost maths. document processingOCRVLMClaudeGeminicase file https://mk-comics.vercel.app/blog/multilingual-ai-features-23-languages-without-23-prompts/ https://mk-comics.vercel.app/blog/multilingual-ai-features-23-languages-without-23-prompts/ Sat, 10 Oct 2026 01:51:00 GMT One prompt with a locale slot, language detection only where needed, per-language token costs and per-language evals. A design for 23 languages, not 23 prompt files. multilinguali18nLLMevaluationprompt cachinglanguage detection https://mk-comics.vercel.app/blog/pii-and-prompts-redaction-before-the-model/ https://mk-comics.vercel.app/blog/pii-and-prompts-redaction-before-the-model/ Sat, 10 Oct 2026 01:50:00 GMT Redact in code before the model call: regex, NER and review, reversible placeholders, what GDPR and India's DPDP Rules imply, and how caching and retention interact. PIIprivacyredactionGDPRDPDPprompt cachingPresidio https://mk-comics.vercel.app/blog/guardrails-that-actually-fire/ https://mk-comics.vercel.app/blog/guardrails-that-actually-fire/ Sat, 10 Oct 2026 01:49:00 GMT Run cheap deterministic checks first, a small model screen second, and log how often each guard fires. A layered design with Haiku 5.5, structured outputs and Noul. guardrailssafetyvalidationstructured outputsClaude Haiku 5.5TypeSafe Noul https://mk-comics.vercel.app/blog/prompt-versioning-and-rollback/ https://mk-comics.vercel.app/blog/prompt-versioning-and-rollback/ Sat, 10 Oct 2026 01:48:00 GMT OpenAI is retiring prompt objects and Anthropic's Playground keeps no history. Version prompts like code, hash them into logs, and roll back without wrecking your cache. promptsversioningCIrollbackprompt cachingDatadog https://mk-comics.vercel.app/blog/cost-observability-for-llm-apps/ https://mk-comics.vercel.app/blog/cost-observability-for-llm-apps/ Sat, 10 Oct 2026 01:47:00 GMT Usage fields differ across Claude, OpenAI and Gemini. Normalise them, price each call with a dated rate table, and send cost and cache hits to Datadog per feature. observabilitycostDatadogLLM pricingClaude APIusage tracking https://mk-comics.vercel.app/blog/latency-budgets-for-ai-features/ https://mk-comics.vercel.app/blog/latency-budgets-for-ai-features/ Sat, 10 Oct 2026 01:46:00 GMT Split an AI feature's latency into time to first token and decode time, then spend a budget on model tier, caching, output length and parallel calls, with a worksheet. latencyperformanceLLMprompt cachingTypeSafe JevClaude Haiku 5.5 https://mk-comics.vercel.app/blog/streaming-ux-for-ai-answers/ https://mk-comics.vercel.app/blog/streaming-ux-for-ai-answers/ Sat, 10 Oct 2026 01:45:00 GMT What users should see in the first second of an AI answer, how to render half-written Markdown without flicker, and how to cancel cleanly. Includes a React example. streamingUXReactMarkdownClaude APIserver-sent events https://mk-comics.vercel.app/blog/function-calling-claude-gpt-6-gemini-same-tool-three-apis/ https://mk-comics.vercel.app/blog/function-calling-claude-gpt-6-gemini-same-tool-three-apis/ Sat, 10 Oct 2026 01:44:00 GMT One lookup tool defined for Claude, OpenAI and Gemini: where the call comes back, how results go in, what parallel and strict mode look like, and a copyable adapter. function callingtool usemulti-providerClaude Sonnet 5.5GPT-6Gemini 3.8 Flash https://mk-comics.vercel.app/blog/structured-output-json-schema-vs-system-one/ https://mk-comics.vercel.app/blog/structured-output-json-schema-vs-system-one/ Sat, 10 Oct 2026 01:43:00 GMT Claude and Gemini schema modes guarantee the shape of a reply; TypeSafe's Jev returns typed answers with probabilities. What each documents, and when to use which. structured outputJSON schemaTypeSafe JevClaudeGeminiSystem One https://mk-comics.vercel.app/blog/human-in-the-loop-by-design-review-queues/ https://mk-comics.vercel.app/blog/human-in-the-loop-by-design-review-queues/ Sat, 10 Oct 2026 01:42:00 GMT Use model confidence to route work to people, then size the queue, set the SLA and feed corrections back into evals. Managed review tools are closing, so build your own. human in the loopreview queuesconfidenceTypeSafe Jevevals https://mk-comics.vercel.app/blog/agentic-workflows-enterprise-control-loop/ https://mk-comics.vercel.app/blog/agentic-workflows-enterprise-control-loop/ Sat, 10 Oct 2026 01:41:00 GMT Plan, act, verify, escalate: budgets, idempotency keys, permission hooks and audit logs for an agent that touches enterprise systems, built on documented SDK features. agentsenterpriseworkflowidempotencyaudit logsClaude Agent SDK https://mk-comics.vercel.app/blog/agent-tool-design-schemas-models-call-correctly/ https://mk-comics.vercel.app/blog/agent-tool-design-schemas-models-call-correctly/ Sat, 10 Oct 2026 01:40:00 GMT Naming, descriptions, parameter types and error returns for agent tools, checked against the Claude and Gemini docs, plus the forced tool-call change on Claude 5.5. tool usefunction callingagentsClaudeGeminiJSON schema https://mk-comics.vercel.app/blog/golden-datasets-200-labelled-examples-in-a-week/ https://mk-comics.vercel.app/blog/golden-datasets-200-labelled-examples-in-a-week/ Sat, 10 Oct 2026 01:39:00 GMT A day-by-day plan for 200 labelled examples: sampling, labelling guidelines, adjudication, splits and versioning, plus the arithmetic on what 200 can and cannot prove. evalslabellingdatasetsgolden datasetLLM-as-judge https://mk-comics.vercel.app/blog/evals-before-agents-smallest-useful-harness/ https://mk-comics.vercel.app/blog/evals-before-agents-smallest-useful-harness/ Sat, 10 Oct 2026 01:38:00 GMT An 85-line Python harness with a dataset, code checks, an LLM judge and a report, plus what to measure first and how much a pass rate on 100 cases can tell you. evalstestingLLM https://mk-comics.vercel.app/blog/hybrid-search-bm25-embeddings-reranker/ https://mk-comics.vercel.app/blog/hybrid-search-bm25-embeddings-reranker/ Sat, 10 Oct 2026 01:37:00 GMT Why keyword search still earns a place, how reciprocal rank fusion merges it with embeddings, what rerankers cost, and an eval harness to run on your own data. hybrid searchBM25reranking https://mk-comics.vercel.app/blog/milvus-vs-pgvector-for-a-team-of-four/ https://mk-comics.vercel.app/blog/milvus-vs-pgvector-for-a-team-of-four/ Sat, 10 Oct 2026 01:36:00 GMT Milvus or pgvector? Index types, filtering behaviour, the 2,000-dimension index limit and the operational footprint, read from both projects' own docs. Milvuspgvectorvector database https://mk-comics.vercel.app/blog/chunking-is-a-product-decision/ https://mk-comics.vercel.app/blog/chunking-is-a-product-decision/ Sat, 10 Oct 2026 01:35:00 GMT Chunk size sets what the model reads, what you store and what each query costs. The evidence on size and overlap, a priced table, and a harness to test your own. RAGchunkingembeddings https://mk-comics.vercel.app/blog/rag-in-2026-when-retrieval-is-optional/ https://mk-comics.vercel.app/blog/rag-in-2026-when-retrieval-is-optional/ Sat, 10 Oct 2026 01:34:00 GMT Every current Claude model has a 1M-token window and cached reads cost 5% of input. When to skip retrieval, when not to, with a priced 150k-token example. RAGlong contextretrieval https://mk-comics.vercel.app/blog/ai-coding-agents-compared-claude-code-codex-antigravity/ https://mk-comics.vercel.app/blog/ai-coding-agents-compared-claude-code-codex-antigravity/ Sat, 10 Oct 2026 01:33:00 GMT As of 10 October 2026: what Claude Code, OpenAI Codex and Google Antigravity are, which models and plans sit behind them, and how to run a fair bake-off. Claude CodeCodexAntigravitycoding agentspricing https://mk-comics.vercel.app/blog/open-weight-models-late-2026-llama-mistral-qwen/ https://mk-comics.vercel.app/blog/open-weight-models-late-2026-llama-mistral-qwen/ Sat, 10 Oct 2026 01:32:00 GMT As of 10 Oct 2026: Qwen3.8 and Mistral Small 4 are open, Mistral Large 4 is API-only for now, and Meta has nothing newer than Llama 4. Licences and when to self-host. open weightsLlamaMistralQwenlicencesself-hosting https://mk-comics.vercel.app/blog/claude-api-speed-fast-setting-when-to-pay/ https://mk-comics.vercel.app/blog/claude-api-speed-fast-setting-when-to-pay/ Sat, 10 Oct 2026 01:31:00 GMT As of 10 October 2026, speed: fast gives Opus 5.5 up to 2.5x output tokens per second at twice the price. What it does, what it breaks, and when it pays. Claude APIfast modelatencypricingClaude Opus 5.5prompt caching https://mk-comics.vercel.app/blog/1m-context-128k-output-what-you-can-build/ https://mk-comics.vercel.app/blog/1m-context-128k-output-what-you-can-build/ Sat, 10 Oct 2026 01:30:00 GMT Every current Claude model has a 1M window and 128k output. The costs, the streaming rule for long generations, the 300k batch beta, and a worked bill. context windowmax outputClaudestreamingcost https://mk-comics.vercel.app/blog/claude-haiku-5-5-rejects-temperature-sampling-controls/ https://mk-comics.vercel.app/blog/claude-haiku-5-5-rejects-temperature-sampling-controls/ Sat, 10 Oct 2026 01:29:00 GMT As of 10 October 2026, Haiku 5.5, Sonnet 5.5 and Opus 5.5 return 400 for non-default temperature, top_p and top_k. What to do about determinism and variety. Claude Haiku 5.5temperatureAPI designsampling parametersmigration https://mk-comics.vercel.app/blog/prompt-cache-read-pricing-compared-across-labs/ https://mk-comics.vercel.app/blog/prompt-cache-read-pricing-compared-across-labs/ Sat, 10 Oct 2026 01:28:00 GMT Cache read multipliers, write fees, minimum prompt sizes and lifetimes for Claude, Gemini and GPT-6 in one table, plus break-even maths and a rule by traffic shape. prompt cachingLLM pricingClaudeGeminiGPT-6cost optimisation https://mk-comics.vercel.app/blog/price-per-million-tokens-october-2026-all-labs/ https://mk-comics.vercel.app/blog/price-per-million-tokens-october-2026-all-labs/ Sat, 10 Oct 2026 01:27:00 GMT Input, cached input and output per million tokens for the current Claude, Gemini and OpenAI models, with a source for every row and the footnotes that change the bill. LLM pricingClaudeGeminiGPT-6cost optimisation https://mk-comics.vercel.app/blog/model-deprecation-calendar-q4-2026/ https://mk-comics.vercel.app/blog/model-deprecation-calendar-q4-2026/ Sat, 10 Oct 2026 01:26:00 GMT Every dated retirement from Anthropic, Google and OpenAI between October and December 2026, with migration targets, the soft dates to ignore, and a CI check. deprecationClaudeGeminiOpenAImigration https://mk-comics.vercel.app/blog/gemini-deep-research-models-preview-deprecations/ https://mk-comics.vercel.app/blog/gemini-deep-research-models-preview-deprecations/ Sat, 10 Oct 2026 01:25:00 GMT The December 2025 Deep Research agent shuts down on 23 October 2026. What the two April replacements do, what they cost per task, and a migration checklist. Gemini deep researchdeprecationagentsInteractions API https://mk-comics.vercel.app/blog/confidence-gated-routing-second-decision-axis/ https://mk-comics.vercel.app/blog/confidence-gated-routing-second-decision-axis/ Sat, 10 Oct 2026 01:24:00 GMT Act, confirm or escalate: how to gate model answers on confidence, set thresholds by cost of error, and get a usable signal from Jev and from JSON-schema chat models. confidenceroutinghuman in the loopTypeSafe Jevstructured output https://mk-comics.vercel.app/blog/jev-noul-yes-no-probability-thresholds/ https://mk-comics.vercel.app/blog/jev-noul-yes-no-probability-thresholds/ Sat, 10 Oct 2026 01:23:00 GMT TypeSafe's Noul returns the probability that a statement is true. How to phrase questions, pick thresholds from the cost of errors, and route the middle band to people. TypeSafe JevNoulthresholdsguardrailsPython https://mk-comics.vercel.app/blog/jev-score-primitive-designing-rubrics/ https://mk-comics.vercel.app/blog/jev-score-primitive-designing-rubrics/ Sat, 10 Oct 2026 01:22:00 GMT How TypeSafe's Jev Score works: ordered levels, the score and confidence maths, why numbers in rubrics fail, and composite scoring to rank pull requests in code. TypeSafe JevScorerubricscomposite scoringPython https://mk-comics.vercel.app/blog/jev-choice-primitive-255-options-probability-distribution/ https://mk-comics.vercel.app/blog/jev-choice-primitive-255-options-probability-distribution/ Sat, 10 Oct 2026 01:21:00 GMT How to write Choice criteria for TypeSafe's Jev, read the probability distribution, set confidence thresholds per question, and search deep taxonomies with a beam. TypeSafe JevChoiceclassificationconfidencePython https://mk-comics.vercel.app/blog/typesafe-ai-the-lab-betting-on-models-that-answer/ https://mk-comics.vercel.app/blog/typesafe-ai-the-lab-betting-on-models-that-answer/ Sat, 10 Oct 2026 01:20:00 GMT TypeSafe came out of stealth on 15 September with Jev, a System One model that returns typed decisions. What the category is, what is claimed, and who should care. TypeSafeJevSystem OnefundingLLM pricingstructured output https://mk-comics.vercel.app/blog/agentic-video-understanding-gemini-88-percent-fewer-tokens/ https://mk-comics.vercel.app/blog/agentic-video-understanding-gemini-88-percent-fewer-tokens/ Sat, 10 Oct 2026 01:19:00 GMT Gemini's agentic video mode loads only the frames, audio and transcript a prompt needs. What shipped, why 88% fewer tokens is not 88% off the bill, and a calculator. video understandingGeminitokensLLM pricingGemini 3.8 Flashcost optimisation https://mk-comics.vercel.app/blog/lyria-3-5-full-length-song-generation/ https://mk-comics.vercel.app/blog/lyria-3-5-full-length-song-generation/ Sat, 10 Oct 2026 01:18:00 GMT lyria-3.5 generates songs of a couple of minutes from text or images at $0.08 each. The API call, a draft-then-render workflow, and the licensing questions. Lyria 3.5music generationGemini APIInteractions APIlicensing https://mk-comics.vercel.app/blog/antigravity-agent-preview-google-agent-platform/ https://mk-comics.vercel.app/blog/antigravity-agent-preview-google-agent-platform/ Sat, 10 Oct 2026 01:17:00 GMT antigravity-preview-09-2026 replaced the 05-2026 agent on 17 September: renamed tools, line-range edits, token budgets. A migration checklist and a minimal call. AntigravityGemini agentsGoogleInteractions APImigration https://mk-comics.vercel.app/blog/1m-context-by-default-when-long-context-beats-rag/ https://mk-comics.vercel.app/blog/1m-context-by-default-when-long-context-beats-rag/ Sat, 10 Oct 2026 01:16:00 GMT Most current Claude models include 1M tokens at standard pricing. A priced comparison of stuffing a 400k corpus versus retrieving 8k, and a decision table. long contextRAGClaudeprompt cachingcost https://mk-comics.vercel.app/blog/claude-fable-5-1-mythos-tier-who-needs-it/ https://mk-comics.vercel.app/blog/claude-fable-5-1-mythos-tier-who-needs-it/ Sat, 10 Oct 2026 01:15:00 GMT Fable 5.1 costs $10 in and $50 out per million tokens, 2.5x Opus 5.5. Where the premium can pay, how cached agent loops price out, and who can use Mythos 5.1. Claude Fable 5.1Claude MythosClaude Opus 5.5LLM pricingagents https://mk-comics.vercel.app/blog/openai-july-2026-agent-incident-safeguards-first/ https://mk-comics.vercel.app/blog/openai-july-2026-agent-incident-safeguards-first/ Sat, 10 Oct 2026 01:14:00 GMT What sources say about OpenAI's July 2026 agent incident, the UK AISI Astra tests and the GPT-6.1 shelving, plus a default-deny tool policy you can copy. AI safetyagentsOpenAIsandboxingtool permissionsGPT-6 https://mk-comics.vercel.app/blog/gpt-6-instant-intelligent-ui-widget-answers/ https://mk-comics.vercel.app/blog/gpt-6-instant-intelligent-ui-widget-answers/ Sat, 10 Oct 2026 01:13:00 GMT ChatGPT answers now arrive as buttons, charts and forms. What OpenAI has said, what it has not, and how to design chat features and Apps SDK widgets around it. GPT-6Intelligent UIgenerative UIproduct designMCP Apps https://mk-comics.vercel.app/blog/gpt-6-astra-sol-luna-instant-what-we-know/ https://mk-comics.vercel.app/blog/gpt-6-astra-sol-luna-instant-what-we-know/ Sat, 10 Oct 2026 01:12:00 GMT GPT-6 has a published API price list, but Instant is not on it. What OpenAI's pricing page, press and Wikipedia agree on, where they conflict, and how to plan around it. GPT-6OpenAImodel selectionLLM pricingGPT-6 Astra https://mk-comics.vercel.app/blog/gemini-2-5-wind-down-what-limited-access-means/ https://mk-comics.vercel.app/blog/gemini-2-5-wind-down-what-limited-access-means/ Sat, 10 Oct 2026 01:11:00 GMT Since 18 September, Gemini 2.5 access is limited to prior users. The text models have no shutdown date; the audio and image ones do. Dates, prices and a migration check. Gemini 2.5deprecationmigrationGemini APIimplicit caching https://mk-comics.vercel.app/blog/nano-banana-2-1-4k-image-generation-gemini-api/ https://mk-comics.vercel.app/blog/nano-banana-2-1-4k-image-generation-gemini-api/ Sat, 10 Oct 2026 01:10:00 GMT gemini-nano-banana-2.1 went GA on 6 October with 1K, 2K and 4K output and panoramic ratios. Per-image prices, what it replaces, and a Python call you can adapt. Nano Banana 2.1image generationGemini APIpricingOpen Graph images https://mk-comics.vercel.app/blog/gemini-3-8-flash-tts-voices-endpoint-narrated-tour/ https://mk-comics.vercel.app/blog/gemini-3-8-flash-tts-voices-endpoint-narrated-tour/ Sat, 10 Oct 2026 01:09:00 GMT Gemini 3.8 Flash TTS went GA on 22 September with a voices endpoint. Prices, limits, a Next.js route that narrates a page, and the 17 November shutdown to plan for. Gemini TTStext to speechNext.jsGemini 3.8 Flash TTSvoice designaccessibility https://mk-comics.vercel.app/blog/gemini-3-8-live-voice-agents-extended-thinking/ https://mk-comics.vercel.app/blog/gemini-3-8-live-voice-agents-extended-thinking/ Sat, 10 Oct 2026 01:08:00 GMT Gemini 3.8 Live and its extended-thinking variant went GA on 15 September 2026. What the extra reasoning does, what it costs the client, and a latency budget. Gemini 3.8 Livevoice agentsLive APIextended thinkinglatency https://mk-comics.vercel.app/blog/gemini-3-8-flash-in-production/ https://mk-comics.vercel.app/blog/gemini-3-8-flash-in-production/ Sat, 10 Oct 2026 01:07:00 GMT Since 8 October 2026, requests to gemini-3.7-flash are routed to gemini-3.8-flash at the same price, until rates double on 1 January 2027. What to test before it bites. Gemini 3.8 FlashGemini APImigrationLLM pricingmodel routing https://mk-comics.vercel.app/blog/claude-sonnet-4-5-retirement-migration-checklist/ https://mk-comics.vercel.app/blog/claude-sonnet-4-5-retirement-migration-checklist/ Sat, 10 Oct 2026 01:06:00 GMT claude-sonnet-4-5-20250929 retires on 30 November 2026. The settings that now return a 400 on Sonnet 5.5, the cost delta, and how to test before the date. Claude Sonnet 4.5Claude Sonnet 5.5deprecationmigrationLLM pricing https://mk-comics.vercel.app/blog/claude-opus-5-5-breaking-changes-migration/ https://mk-comics.vercel.app/blog/claude-opus-5-5-breaking-changes-migration/ Sat, 10 Oct 2026 01:05:00 GMT Opus 5.5 costs less per token than Opus 5 but returns a 400 for thinking disabled and forced tool_choice. The exact errors, the cost trap and a migration checklist. Claude Opus 5.5migrationAPIadaptive thinkingLLM pricing https://mk-comics.vercel.app/blog/claude-haiku-5-5-high-volume-classification/ https://mk-comics.vercel.app/blog/claude-haiku-5-5-high-volume-classification/ Sat, 10 Oct 2026 01:04:00 GMT Haiku 5.5 launched 7 October 2026 at $0.10 input and $0.50 output per million tokens. What changed from 4.5, what returns a 400, and a classification bill priced out. Claude Haiku 5.5classificationLLM pricingstructured outputsprompt caching https://mk-comics.vercel.app/blog/jev-typed-questions-instead-of-prompt-and-parse/ https://mk-comics.vercel.app/blog/jev-typed-questions-instead-of-prompt-and-parse/ Sat, 10 Oct 2026 00:00:00 GMT Jev answers typed questions with calibrated probabilities instead of writing text: how Choice, Score and Noul work, what confidence means, and a worked intake example. TypeSafe JevSystem Onestructured outputclassificationconfidencePython https://mk-comics.vercel.app/blog/prompt-caching-economics-rag-bill-2026/ https://mk-comics.vercel.app/blog/prompt-caching-economics-rag-bill-2026/ Sat, 10 Oct 2026 00:00:00 GMT Cached input now costs 5 to 10% of list price on Claude, Gemini and GPT-6. The rules, minimums and traps, plus a 10,000-call-a-day RAG bill priced on three models. prompt cachingLLM pricingRAGClaude Sonnet 5.5Gemini 3.8 FlashGPT-6cost optimisation https://mk-comics.vercel.app/blog/which-ai-model-to-build-on-october-2026/ https://mk-comics.vercel.app/blog/which-ai-model-to-build-on-october-2026/ Sat, 10 Oct 2026 00:00:00 GMT Claude 5.5, GPT-6, Gemini 3.8 and TypeSafe's Jev all shipped in five weeks. Dates, prices and limits from primary sources, and what I would actually pick for a product. Claude 5.5GPT-6Gemini 3.8TypeSafe Jevmodel selectionLLM pricing