Duke Harewood

A concrete, synthesis checklist for moving AI-generated code from prototype to production safely — the practical close to this series, not another abstract warning.

Treating security review as a checklist bolted on before launch guarantees it gets rushed. The teams doing this well embed it as a running skill alongside generation, not after it.

The tradeoff between shipping fast and shipping safe is real, not a strawman — here's an honest look at where the line between 'ship it' and 'review it' should actually sit.

A concrete, practical review workflow for AI-generated code at a small team — scoped by risk tier, not by pretending every line deserves equal scrutiny.

86% of AI-generated code samples failed to defend against XSS, and 88% were vulnerable to log injection. Both share one root cause: unescaped input treated as safe.

Security teams running internal 2026 workshops on AI-generated code aren't teaching new skills — they're teaching engineers to recognize a small, repeating set of failure patterns fast.

Industry-wide estimates put AI-code cleanup costs between $400M and $4B, with per-team maintenance costs up roughly 300% in 18 months. Here's what that means at your team's scale.

An 8.1-million-pull-request study found technical debt climbing after AI adoption. The cause isn't worse code per line — it's review capacity that never scaled with volume.

AI-assisted commits leak credentials at roughly double the baseline GitHub rate. The mechanism isn't a smarter attacker — it's a faster typist with no memory of what it just wrote.

Andrej Karpathy's 'vibe coding' term describes a technique, not a defect. The defect is treating AI-generated code as if it doesn't need the review any other code gets.

Review processes, ranking algorithms, and trust infrastructure — the App Store solved these once. Skill marketplaces are solving them again, with real differences.

A practical checklist for marketplace operators, built directly from why 82% of MCP servers failed on path traversal and 67% on code injection.

Cross-agent portability is the skill economy's biggest leverage argument — and its friction points are more specific than the pitch usually admits.

Most skills get installed, run once, and quietly abandoned. The ones that stick share specific, identifiable traits that have nothing to do with cleverness.

Writing a good SKILL.md file takes an afternoon. Getting the right developer to find it before they build their own version takes much longer.

82% of audited MCP servers are vulnerable to path traversal, 67% to code injection. That's the security backdrop every skill marketplace has to reckon with.

Curated skill marketplaces are converging on a 70/30 creator-platform split. That specific number reveals more about the market than either side admits.

Four primitives, one confusing acronym soup. A clarifying breakdown of when to reach for a skill versus a subagent, a hook, or an MCP server.

A plain markdown file plus optional scripts beat every fancier format for one reason: it's the lowest-friction thing a model and a human can both read.

Prompt engineering was a job title for eighteen months. Packaged, reusable, sellable skills are turning out to be the durable unit of AI leverage instead.

When part of what's on call is autonomous, the rotation needs a different runbook — one built around judgment calls, not just error codes.

The instinct to let a stuck agent 'figure it out' is usually wrong. Good kill-switch design is a judgment call made in advance, not under pressure.

If your guardrail config is a live setting someone edits in place, you can't answer 'what was the policy last Tuesday' — and eventually someone will ask.

One agent doing the same expensive thing on every bot visit can quietly burn a shared infrastructure budget before anyone notices the trend line.

Gartner expects AI-related legal claims to top 2,000 by end of 2026, mostly from thin risk guardrails. The audit log is what determines which side you're on.

Everyone agrees hallucination is a risk. Almost nobody has a number for it. Here's how to turn a vibe into a dashboarded production metric.

You can't guarantee an agent's judgment is always right. You can guarantee that when it's wrong, the damage is small, visible, and reversible.

Detection, classification, containment, root cause, remediation — the same five steps as any incident, applied to a system that can improvise its next move.

A guardrail without observability stops the bad action once and tells you nothing. Observability without a guardrail just narrates the damage as it happens.

Traditional APM tells you a request was slow. It can't tell you why an agent chose to call the refund tool instead of asking a human first.

You don't need a graph database on day one. A concrete, minimal memory stack for shipping your first agent without over-building or under-building it.

A model can be copied in a weekend. A year of accumulated, structured memory about a user cannot. That asymmetry is where real agent moats live.

Postgres plus embeddings can get a small team surprisingly far. Here's where that stops working and a dedicated framework starts paying for itself.

Zep and its Graphiti engine treat memory as a timeline, not a snapshot — so an agent can know a fact was once true without confusing it for current.

Persistent agent memory means storing personal facts indefinitely. Most teams ship it before deciding what consent, deletion, and ownership actually mean.

Not every agent needs a permanent memory store. A practical framework for deciding when ephemeral context is enough and when you need durable memory.

Persisting memory is the easy half. Deciding what to decay, contradict, or delete is where most agent-memory systems quietly fall apart.

Embedding similarity finds facts that sound alike. It doesn't know they're connected. Graph-native memory layers like Cognee are built for the difference.

Letta splits agent memory into always-in-prompt core blocks and queryable archival storage. It's a budget allocation problem disguised as an architecture choice.

Model quality has converged across vendors. What users actually feel now is whether a product remembers them — and that's a UX problem, not a benchmark.

MCP is closing in on 10,000 registry servers with a stateless architecture and neutral governance behind it. What has to happen next for it to hold?

AI tool-integration standards have failed before. MCP's design and governance choices explain why this one reached 9,652 servers when others stalled.

With 82% of audited MCP servers vulnerable to path traversal, picking one to connect to your systems needs a real checklist. Here's a practical one.

MCP's Linux Foundation handoff wasn't symbolic. It changed how competing labs, enterprises, and long-tail builders actually behaved. Here's the mechanism.

950+ of MCP's 9,652 registry servers are built for customer service, sales, and internal ops. That's the signal a protocol is maturing past hobbyist adoption.

A 2026 audit of 2,600+ MCP servers found 82% vulnerable to path traversal, 67% to code injection. Why it happened, and what a safe server looks like.

The MCP Registry hit 9,652 latest server records by May 2026. A tour of what's actually in there — the enterprise wave, the long tail, and what it means.

The 2026-07-28 MCP spec traded persistent connections for a stateless architecture. Here's the scaling problem that forced the trade, and what it cost.

MCP's SDKs pull ~97M monthly downloads. That number is real, impressive, and mostly the wrong thing to be looking at. Here's what to check instead.

In 14 months MCP went from an internal Anthropic spec to a Linux Foundation standard. Here's why handing away control was the move that made it win.

A concrete, ordered starting eval stack for a small team shipping their first production agent — what to build first, and what to skip until it hurts.

The 0.85 correlation threshold everyone cites is easy to state and tedious to actually reach. Here's the real, unglamorous process of validating a judge.

Offline eval suites catch what already shipped. Runtime guardrails catch bad outputs before a user ever sees them. Here's how the two connect.

Reading five outputs and deciding a prompt 'feels better' doesn't scale past the first week. Eval-driven development writes the test before the fix.

Agent-as-judge gives evaluators tools, memory, and multi-step reasoning to grade full trajectories — but it inherits every bias a simple judge has, plus new ones.

Running LLM-as-judge on every production request doesn't scale on cost or latency. Here's the sampling architecture teams actually run instead.

You can't mine production logs for eval cases before you have users. Here's how solo builders and small teams bootstrap a golden dataset from zero.

An agent that stumbles into the right answer through wasted tool calls isn't working. Trajectory evals grade the reasoning path, not just the output.

Position bias, verbosity bias, and self-preference bias distort every LLM-as-judge score. Here's what each looks like and the concrete fixes that work.

LLM-as-judge solved the eval bottleneck in 2024. By 2026 it's table stakes — here's what production teams layer on top of it now.

Fleet size isn't a vibe — it's a function of decomposability, token budget, and how much compounding failure risk you can tolerate.

In a single-agent system, the bug is in the code. In a five-agent pipeline, the bug is in the seam — and most teams don't know how to look there.

Developer tools converged on lead-agent-with-subagents before enterprise AI did. That order of adoption is a clue about why the pattern works.

Every approval gate you add slows a fleet down. Every gate you skip is a bet that nothing downstream needed a human. Where to place them.

The weakest point in a multi-agent pipeline is never the agents. It's the seam where context passes from one to the next — or doesn't.

The models were capable enough. A postmortem look at why 40% of multi-agent pilots still didn't survive production — and it wasn't the AI's fault.

Context-inheriting forks and stateless fresh subagents solve opposite problems. Picking wrong either wastes tokens or loses critical context.

Parallel subagents cut research time up to 90% and cost roughly 15x the tokens of a single chat turn. That ratio is a budgeting decision, not a red flag.

Hub-and-spoke and peer-to-peer aren't rivals fighting for the same job — they solve different shapes of problem. Here's how to tell which one you have.

Flat agent swarms looked elegant on a whiteboard. In production, hierarchy won because someone has to own the failure.

Nine questions to run through before shipping an agent, pulled from the mechanisms that actually cause most agent failures — attention budget, tool sprawl, memory, and compaction.

A bigger context window doesn't fix the attention problem — it just gives you more room to spend badly. Structured, retrievable memory does what raw scale can't.

The best agent skill libraries look small from the model's perspective even when they're huge — because almost nothing is loaded until the task actually needs it.

Spinning off a sub-agent isn't just a way to parallelize work — it's a way to keep a noisy investigation out of the main agent's context window entirely.

Every long-running agent eventually has to summarize or evict something. The teams that get this right decide deliberately what survives compression — not by default.

OpenAI's own guidance caps recommended agent toolsets under 20, with accuracy degrading past 10. Here's the mechanism behind that ceiling, and how to prune a toolset that's already grown past it.

Retrieval, packaged skills, and tool calls all solve the same problem — getting the right information in front of the model at the right moment — but they trade off very differently.

When an agent gives a wrong answer, the reflex is to blame the model. In 2026 the more accurate diagnosis, most of the time, is that it never saw the right information.

Treating the context window as free storage is the single most common design mistake in agent building. Every token spends attention. Here's how to think about it like a P&L instead.

Prompt engineering optimized a sentence. Context engineering architects everything the model sees — memory, tools, retrieval, state. That shift happened in under a year, and most teams haven't caught up.

Stripe processes payments; Paddle becomes the legal seller. For a solo founder selling digital products across borders, that distinction decides who handles VAT, GST, refunds, and chargebacks — you or them.

You created a Stripe account, sent a payment link, and hit 'your account cannot currently make live charges.' Here's what actually activates live mode — and the shortest path from zero to a real first dollar.

Digital Applied found 86% of AI citations come from sites with 5+ interlinked pieces on one topic. Yext found a bidirectionally-linked cluster gets cited 2.7x more than an isolated page. AI crawlers read internal links as a map — here's how to structure content as one knowledge node.

Schema markup alone does nothing — Ahrefs tested 1,885 pages and adding it with no content change produced no citation lift. But Organization and FAQPage JSON-LD on top of real content earn 2.7x more AI citations. Here's the code and the deployment checklist.

Reddit is Perplexity's #1 source at 46.7% of top citations and #1 in Google AI Overviews at 21%. One genuinely useful 30-minute answer keeps getting pulled into AI responses indefinitely. Here's the minimal-effort playbook — and why this window won't stay open.

AI engines trust what others say about you far more than what you say about yourself. Muck Rack found 82% of AI citations come from earned media, under 6% from brand-owned pages. The real work of GEO isn't polishing your copy — it's getting mentioned elsewhere.

Search just split into two paths: the ranked blue-link list and the AI-generated answer that quotes a few sources nobody clicks. SEO races for rank; GEO races to be the source the AI draws from. Here's why the distinction reorders how solo founders get found.

Most 'form a US LLC' guides stop at the EIN. The part they skip can cost $25,000: Form 5472, required annually from foreign-owned LLCs even at zero revenue. Plus the BOI rule that changed in 2025 and the state report you still owe.

No SSN? You can still get an EIN by phone or fax — the widely-repeated claim that the IRS killed the fax channel is wrong. Here's the SS-4 prep that keeps you from getting rejected: Line 7b, Line 11, Line 15, and the Responsible Party rule.

Why an LLC beats a C-Corp for most solo founders, and which state to form in. Wyoming's $60-a-year default, Delaware's investor pull, and California's $800 franchise-tax trap — plus the registered-agent requirement nobody explains until you're stuck.

The real payoff of a personal mesh VPN: run an AI coding agent on a home machine and drive it from your phone while the task keeps running. Here's the latency math, the security details that matter, and the one hard limit to plan around.

Tailscale turns every device you own into one private network — no public IP, no firewall rules, no forwarded ports. Here's the connection model that makes it work, and the one thing that quietly leaves you stuck on relay latency.

A concrete 6-step path from empty folder to a live email-capture page — frontend, backend, and database — deployed worldwide on Cloudflare for $0, before a single line of the real product exists. The fastest way to test if anyone wants the thing.

Cloudflare Workers and Pages replace 'rent a VPS and configure nginx' with 'push code, it runs in 335 cities.' Here's why serverless-at-the-edge is a different model, where the free tier ceiling actually sits, and how Pages and Workers are the same primitive.

Oracle's free tier has one real string attached: it reclaims idle instances. Here's what 'idle' actually means for a dev box, the keep-alive pattern that works, why community scripts are a fragile bet, and the fix that ends the problem for good.

Oracle Cloud quietly gives away a permanent 4-core, 24GB-RAM ARM server that would cost $80-100/month on AWS. Here's why signups get rejected, the region choice you can't undo, and how to actually get the machine.

The four-piece kit a non-resident founder needs before a US bank says yes: ITIN, a CMRA-clean address, a US SIM, and a Wise account, plus the card application.

What an ITIN is, why a Certifying Acceptance Agent beats mailing your passport to the IRS, the 8-12 week timeline, and why it's the first US-entity domino.

Why GitHub Pages beats Vercel, Netlify, and a VPS for a solo founder's first landing page: zero cost, no third-party login, and the deploy is a git push.

Registrar tradeoffs, why to route DNS through Cloudflare, the DNSSEC gotcha that takes domains offline, and the orange-vs-grey cloud rule for GitHub Pages.

Eligibility, verification, cost, data privacy, and what happens if you change schools — the practical FAQ for ChatGPT for Teachers, and a look back at the series.

ChatGPT for Teachers is free only through June 2027. OpenAI hasn't said what comes next. Here's what's known, what's speculation, and what to plan for regardless.

ChatGPT for Teachers is a teacher-and-staff offer, not a student one — and several big districts ban student ChatGPT outright. Untangling the confusion.

ChatGPT, Claude, Gemini, or Copilot for your district AI pilot? A practical framework built on the stack you already run, real per-seat prices, and priorities.

A free, general teacher assistant is genuinely useful — and structurally incapable of three things a purpose-built skill does. Here's the line between them.

ChatGPT for Teachers and Claude for Teachers are both teacher-facing productivity tools. The student-facing Socratic tutor is a niche neither giant is actually filling.

OpenAI's education playbook — narrow vertical, distribution through a trusted intermediary, free to build the habit — is a GTM lesson for anyone building a Claude Code skill.

Paid lesson-planning and differentiation tools now compete with something free from the model vendor itself. What the squeeze looks like, and which tools escape it.

When the model vendor ships a free, polished, education-specific product directly to teachers, the ecosystem built on its general-purpose API has to rethink what it's for.

A free tool still costs you a rollout. How to pilot ChatGPT for Teachers across a district — cohort selection, a real feedback loop, and the scale-or-stop decision.

Who should be an admin, who should be a member, and the RBAC misconfigurations that quietly hand a whole district's AI workspace to the wrong people. A practical guide for K-12 IT.

OpenAI names itself a FERPA 'School Official' and says student data stays under the school's control. Here's what those terms actually commit — and what a privacy officer should still ask.

The mechanics of a managed ChatGPT for Teachers deployment: claim the domain, wire up SAML SSO, enforce MFA, and onboard teachers in phases without flooding the help desk.

It's free, so districts skip procurement. That's the mistake. Here's the checklist an IT director should run before ChatGPT for Teachers reaches a single classroom.

Voice mode isn't a gimmick for teachers. It wins in the exact moments a classroom leaves your hands full and your keyboard out of reach. Here's when to talk instead of type.

How to use ChatGPT for Teachers' image generation for worksheets, slides, and classroom visuals — and why every AI-generated diagram needs an accuracy check first.

How to use ChatGPT for Teachers to draft differentiated materials and IEP-aligned supports — and exactly where FERPA caution around student-identifying data has to stop you.

How ChatGPT for Teachers helps draft difficult parent emails — setting tone, structuring hard news, and translating — with the rule to always personalize and review first.

How file uploads in ChatGPT for Teachers speed up rubric-building and feedback drafting — without letting AI grade for you or auto-send anything to students.

How ChatGPT for Teachers' shared projects turn department co-planning from a Friday scramble into a durable, shared workspace. A realistic PLC workflow.

A custom GPT turns a good prompt into a reusable tool your whole department shares. Here's how to build a rubric-grader or differentiation GPT that actually gets used.

OpenAI, Anthropic, Google, and Microsoft are all racing for the same classrooms — with four different strategies. Here's how the board actually looks.

OpenAI could have shipped a discounted teacher tier. It shipped a fully free one through June 2027. The go-to-market logic behind that choice is worth reading.

Token context sounds abstract until you try uploading a whole class set of essays. Here's what 400K vs 1M–2M tokens actually means for a teacher's day.

ChatGPT for Teachers is free per seat — but adoption isn't. Here's the true cost of rolling it out versus paid Gemini, Copilot, and Claude for schools.

Same GPT-5.1 engine, different chassis. A free standalone workspace versus an Office-native paid tier — and why the model isn't what decides this one.

For a Google-shop district, the AI question isn't which model is smarter. It's whether Workspace gravity and NotebookLM beat a free standalone ChatGPT.

OpenAI isn't the only company giving teachers AI for free. Anthropic's education play — a free K-12 offer plus a paid, partner-rich tier — is the quieter alternative.

AI vendors are funding the union that trains teachers to use their tools. That's a real conflict worth naming — and the safeguards deserve a fair hearing too.

OpenAI is courting teachers with a free workspace while districts ban ChatGPT for students. Is that a contradiction — or two different problems wearing one name?

An Oregon NEA rank-and-file caucus wants an end to what it calls 'AI grift.' Here's what that grassroots pushback represents — and what it doesn't, within a mixed union landscape.

An MIT study found students who used ChatGPT to write essays showed weaker cognitive engagement. Here's what that implies for student use — and why teacher use is a different case.

Major districts banned ChatGPT over integrity and quality fears. Now OpenAI is courting those same districts with free teacher accounts. Are the two positions contradictory?

AI-writing detectors run roughly 80-90% accurate — and most don't publish rates at all. Here's why that's not good enough to be a teacher's primary defense.

A Study.com survey found over a quarter of teachers have caught a student cheating with ChatGPT. Here's what that number actually measures — and what it doesn't.

The AFT's academy promises to train 400,000 teachers on AI in five years. But scale and depth pull against each other — can standardized training build real literacy?

A fair critical review of OpenAI's free teacher course — what it teaches well, and the gaps around AI grading, hallucination, and detection it glosses over.

OpenAI's AI Skills Jam took the teacher training on the road — in-person workshops across US cities that drew 1,600+ educators. Here's what the format looks like.

OpenAI Academy runs two free courses for K-12 teachers: AI Foundations and ChatGPT Foundations for Teachers. Here's the actual syllabus, section by section.

OpenAI, Microsoft, and Anthropic compete everywhere else — but co-funded the same $23M union program to train 400,000 teachers. Here's why.

Inside the UFT's Manhattan flagship for the National Academy for AI Instruction — the format, who attends, and why a physical building anchors a national AI program.

The AFT accepted $23M from OpenAI, Microsoft, and Anthropic to train teachers on AI. Its rationale — 'teachers in the driver's seat' — and the skepticism it drew.

The AFT's National Academy aims to train one in ten U.S. teachers on AI by 2030. What does standardizing AI literacy at that scale gain — and quietly cost?

A full breakdown of the $23M National Academy for AI Instruction: who pays what, OpenAI's $10M split, the five-year scope, and the Manhattan hub at its center.

OpenAI calls itself a FERPA "School Official" for ChatGPT for Teachers. Here's what that legal designation actually guarantees for student data — and what it doesn't.

The admin's walkthrough for rolling out ChatGPT for Teachers district-wide: claiming your domain, role-based access, SSO, MFA, and the single-district rule.

Free flagship AI for millions of K-12 teachers with no revenue looks like charity. It isn't — here's the habit, procurement, and competitive logic underneath.

OpenAI now sells four different ChatGPT plans. Here's what makes the free K-12 teacher tier genuinely different from Plus, Edu, and Team — and who each is for.

Before ChatGPT for Teachers unlocks, SheerID has to confirm you're a real educator. Here's the verification flow, the fallbacks, and what happens if you fail.

OpenAI is giving verified US K-12 teachers unlimited GPT-5.1 Auto, file uploads, connectors, and image gen free through June 2027. Here's the full list.

Fifty articles in, the takeaway is simple: free AI raises the floor for every teacher. The ceiling is set by the tools and skills you build on top of it.

Vendor competition happens at the platform layer. But the real classroom differentiation lives in the skills layer on top — and it sits on either model.

You don't have to wait for a vendor to ship the perfect tool. Here's how a reusable lesson-planning skill layers your curriculum onto any capable model.

A student asking general ChatGPT for help often just gets the answer. A tutor built with deliberate pedagogy withholds it — and that gap decides who learns.

A general chatbot forgets your class between sessions, enforces no rubric, and guarantees no standards alignment. Here's the exact gap a skill closes.

A general chat window is powerful but undifferentiated. It doesn't encode your curriculum standards, rubrics, or pedagogy the way a purpose-built skill can.

OpenAI gives ChatGPT for Teachers away free and funds a $23M AFT academy. Follow the money: here's the business logic and what to stay skeptical about.

AI as a shortcut can erode the critical thinking it should build. Here's the fair case for classroom AI limits and the guardrails a teacher can actually set.

AI-text detectors flag real student writing as fake and miss lightly edited AI. Here's why AI-proofing fails and what assignment redesign does instead.

Baltimore, LA, NYC and Seattle banned ChatGPT in 2023. Many later reversed course. Here's what changed to move districts from bans to ChatGPT for Teachers.

In the same survey, 43% of teachers said AI makes their job harder while nearly 58% said easier. Unpacking the split — and what it means for district training.

Roughly 1 in 4 teachers have caught a student using AI to cheat. A clear-eyed look at the integrity problem — and what shifts when teachers hold the same tool.

Free isn't automatically good. The steelmanned case against ChatGPT for Teachers: data exposure, vendor lock-in, and the incentive question Futurism raised.

EdWeek framed ChatGPT for Teachers as a boon, a bust, or just 'meh.' We lay out all three cases fairly, then offer a considered synthesis of which is closest.

A free workspace can draft, differentiate, and grade fast — but it can't build trust, manage a room, or know your students. An honest look at the hard limits.

Six named, copy-pasteable ChatGPT for Teachers prompt templates — differentiation, rubric-first grading, parent-email tone, and more — in one reference.

How connectors let ChatGPT for Teachers work directly from your Drive files instead of copy-paste — plus the access-scope questions worth asking first.

Sick and out tomorrow? A ChatGPT for Teachers workflow that turns your existing lesson context into a clear, self-contained substitute plan in ten minutes.

A step-by-step ChatGPT for Teachers workflow using file upload to convert a textbook chapter into a varied, standards-aligned quiz you can trust and reuse.

A repeatable ChatGPT for Teachers workflow for progress updates, behavior notes, and conference summaries that protects student privacy and keeps your voice.

A practical workflow for using ChatGPT to speed up grading — rubric-based first passes and spotting error patterns — while keeping the judgment human.

Adapt classroom materials to match IEP accommodations with ChatGPT — a careful workflow that assists the process without replacing it or risking data.

A ChatGPT workflow for adapting one passage across five reading levels for a mixed-ability class — the prompt technique and the calls only you can make.

A concrete workflow for turning one curriculum standard into five days of ready-to-teach lesson plans in ChatGPT — the prompt sequence and what to check.

Most districts won't pick just one AI vendor. Here's the realistic three-vendor stack — Gemini, ChatGPT, Claude — and the governance tax it quietly adds.

Google and Microsoft bundle education AI into existing contracts. OpenAI gave it away. An analysis of the free-through-2027 offer as a competitive land-grab.

ChatGPT, Claude, Gemini, or Copilot? A practical decision framework for districts evaluating all four AI assistants, with a full side-by-side comparison table.

Two rival AI labs, one identical pitch: free for K-12 educators. A head-to-head on offer structure, what's confirmed vs. not, and the ecosystem difference.

For districts invested in Microsoft 365, the AI choice is subtle. A head-to-head on standalone workspace vs. Copilot living inside Word, Outlook, and Teams.

A head-to-head on the two AI assistants K-12 educators actually reach for: free offer structure, integration depth, context window, and best-fit use case.

A union runs teacher AI training instead of the tech giants funding it. Futurism smells a motive. Here's the skeptic's case — and why the union still matters.

AI fluency for teachers isn't prompt engineering. It's four practical skills: judging output, knowing what to keep human, data basics, and teaching AI literacy.

The AFT's flagship AI training campus opens in New York City as the hub before a national rollout. Here's what a physical teacher-AI campus actually does.

400,000 teachers trained by 2030 means roughly 220 educators a day, every day. We run the scale math on the AFT AI Academy's ambitious five-year target.

Three fierce AI rivals — OpenAI, Microsoft, and Anthropic — co-fund the same teachers' union training program. Here's the coopetition logic behind the move.

OpenAI, Microsoft, and Anthropic put $23 million behind one teachers' union to train 400,000 educators in AI by 2030. Here's what the Academy actually is.

No breach has hit ChatGPT for Teachers — but K-12 ed-tech is a proven target. Here's what a breach of an AI workspace holding student data could expose.

Deep LMS integration for ChatGPT for Teachers is an open question, not a confirmed feature — an honest look at the SSO, rostering, and Classroom gap to watch.

A numbered checklist for IT directors piloting OpenAI's ChatGPT for Teachers before district-wide rollout — vendor review, opt-in, metrics, and escalation.

OpenAI says workspace data isn't used for training 'by default.' Here's what that two-word qualifier really implies for teachers — and how to weigh opt-outs.

FERPA in plain terms: what OpenAI's stated support for it means in practice — and what FERPA does NOT guarantee just because a vendor claims to support it.

A practical, hedged playbook for IT teams provisioning OpenAI's free ChatGPT for Teachers at scale — accounts, SheerID opt-in, and pilot-to-scale sequencing.

ChatGPT for Teachers is free — but only for verified U.S. K-12 educators. Here's who's left out, why, and what alternatives exist for those it excludes.

ChatGPT for Teachers is free — but only through June 2027. Here's the business logic behind the sunset date and what a teacher or district should plan for.

OpenAI's ChatGPT for Teachers includes image generation. Here's what it's actually good for in a K-12 classroom — and where teachers must review before use.

Web search, connectors, and file uploads are the three most classroom-useful features in ChatGPT for Teachers. Here's what each does, with classroom examples.

GPT-5.1 Auto picks the right model for every task so teachers don't have to. Here's what model routing means for a busy educator — and why that's the point.

ChatGPT for Teachers, ChatGPT Edu, ChatGPT Plus, and ChatGPT Team overlap and confuse. Here's a clear comparison to help teachers and IT buyers pick right.

OpenAI outsources teacher verification to SheerID instead of building its own. Here's why that choice matters, what it signals about fraud, and what to expect.

OpenAI's ChatGPT for Teachers is free for verified US K-12 educators through June 2027. Here's exactly what's included and who qualifies for the offer.

The open K-12 teaching skills were written to use the Learning Commons Knowledge Graph connector. When your environment doesn't have it, here's how to bring your own standards and curriculum so lessons stay aligned.

Anthropic is working with the American Federation of Teachers on a Gold Standard of K-12 safety and privacy best practices. Why a union-shaped standard is the right way to set the floor for classroom AI.

Anthropic's free AI Fluency course, co-created with Teach for America, is model-agnostic and Creative Commons-licensed. Why fluency — not access — is the real bottleneck in classroom AI.

The evidence on student-facing AI is mixed and implementation-dependent. Teacher-facing AI is a different bet entirely — support the craft, protect the time. Why Claude for Teachers points the tool at the adult in the room.

The Claude for Teachers class-data workflow: give Claude a folder of rosters, diagnostics, attendance, and notes, and it builds a picture of where each student is. You control what's shared, and nothing shared is used for model training.

The ASSISTments connector in Claude for Teachers generates auto-scored, standards-aligned math problems for practice and assessment. The auto-scoring is the point — it closes the feedback loop at the speed of the lesson, not the speed of your grading stack.

Claude for Teachers ships more than a product: open-source skills, a public evals write-up, ecosystem connectors, a Detroit Public Schools pilot, and a Gates Foundation partnership. The bigger picture.

The Brisk Teaching connector turns a teacher's idea into interactive student activities and standards-aligned lessons in seconds. This piece is about the thing that usually dies between the idea and the artifact — and why closing that gap changes what you're willing to try.

Anthropic open-sourced two K-12 teaching skills. Read them as a blueprint: domain grounding in references/, real artifacts from scripts/, rubric-based evals, and honest connector-awareness.

The Canva for Education connector turns lesson materials into classroom-ready designs and interactive experiences. This is about the last mile of lesson-making — the gap between correct content and a thing students will actually engage with.

Claude for Teachers ships with Claude Code and Cowork, so Claude can carry multi-step work forward on its own. The shift from chatbot to agent is what turns AI from a faster search box into a colleague who finishes the job.

Claude for Teachers ships with a directory of K-12 connectors — ASSISTments, Brisk, Canva, Coteach, Diffit, Eedi, MagicSchool, Snorkl, TeachFX — that unify a fragmented edtech landscape inside one conversation.

Claude for Teachers is a product for the adult in the room, not the students. Here's why a teacher-only tool consistent with Claude's 18-and-over policy is the safer design decision.

Claude for Teachers is entirely free once you verify — and signing up by June 30, 2027 locks in a full year of access. Here's exactly who qualifies, what verification means, and why the deadline is the part not to sleep on.

Claude for Teachers is for individual verified educators, with a schools-and-districts offering coming soon. Until then, districts run through Claude for Nonprofits. Here's which path fits whom, and how to choose today.

The Coteach connector creates high-quality math diagrams grounded in K-12 curriculum. In math, the diagram isn't a decoration on the concept — it often is the concept. This is why 'grounded in curriculum' is the phrase that matters.

Claude for Teachers includes Cowork and Claude Code, so Claude can run multi-step work autonomously. A practical guide to what a teacher should hand off, what to keep, and how to delegate work that carries forward without you.

How the Claude for Teachers differentiation workflow turns one lesson into tiered versions — below, at, and above proficiency — with a differentiation plan and per-level student materials that keep the core content consistent.

Differentiation and personalization get used interchangeably, and the confusion matters. One keeps the destination fixed and varies the road; the other risks moving the destination itself. Here's which one Claude for Teachers actually does — and why the distinction protects students.

The Diffit connector creates and adapts instructional materials for every student. The interesting verb is 'adapt' — reworking material you already trust for the range of readers in front of you, which is the highest-value, most time-expensive job a teacher does.

Eedi's Claude for Teachers connector generates diagnostic questions that reveal student thinking in English and Spanish — turning assessment from a backward-looking grade into a forward-looking plan.

Claude for Teachers ships its own teacher terms and a K-12 Data Processing Addendum written to comply with FERPA. Here's what that actually protects when you hand Claude your class data.

Anthropic open-sourced not just the K-12 teaching skills but the evals/ framework and rubric CSVs that judged them. Here's how to fork that harness to test your own teaching skills for rigor, alignment, and usability.

The k12-lesson-planning skill doesn't stop at chat text. Its scripts/ folder renders a structured lesson JSON into a Word doc and a styled web page — because the real finish line is a file you can print, hand out, and project.

Playlab will support a national network of lab schools where teachers become builders of classroom AI, not just users. Why teacher-as-builder breaks the ceiling of any off-the-shelf tool.

Lesson planning is a second, unpaid shift that spills into teachers' evenings — and it lands hardest in under-resourced schools. Claude for Teachers goes after that tax directly: it targets time, the one thing teachers value most.

The danger of AI feedback isn't sloppy grading — it's feedback so complete it does the student's thinking for them. Good feedback diagnoses without dissolving the struggle where learning happens. Here's where AI belongs in the grading loop, and where it doesn't.

The Claude for Teachers repo ships its evaluation framework — per-subject CSV rubrics, a clarifying-question rubric, and a differentiation rubric. Here's what those rubrics reveal about how you build an education skill you can trust.

Claude for Teachers plans math from IM v.360 by Illustrative Mathematics, not from a blank page. Here's why grounded math lesson design beats generic generation — and how a trusted curriculum keeps AI math honest.

A close read of the second open-source Claude for Teachers skill — how k12-lesson-differentiation turns one lesson into below-, at-, and above-level versions while keeping the core content identical, and the rubrics that keep it honest.

A close read of the k12-lesson-planning skill Anthropic open-sourced with Claude for Teachers — how it draws on standards-mapped curricula and learning progressions to draft a plan plus student-facing materials.

The two K-12 teaching skills ship free inside Claude for Teachers — but they're also open-sourced. Here's the exact git clone, plugin marketplace add, and plugin install sequence to run them in Claude Code today.

Anthropic's Claude for Teachers gives verified US K-12 educators free premium Claude, a library of teaching skills, standards for all 50 states, and Claude Code plus Cowork. Here's what's inside.

The K-12 teaching skills ship as a Claude Code plugin — bundling both skills plus teacher-focused MCP servers behind marketplace.json, plugin.json, and .mcp.json. It's a reusable packaging pattern worth copying.

Claude for Teachers ships every educator a connector to the Learning Commons Knowledge Graph — academic standards across all 50 states, plus the competencies and progressions beneath them. Here's why that changes what 'standards-aligned' means.

A standard is a destination, not a route. Claude for Teachers reaches past the standard to the smaller competencies it's built from and the order students learn them — and that progression map is what makes real scaffolding possible.

MagicSchool's Claude for Teachers connector makes instructional content classroom-ready — the finishing step that turns a solid draft into something you can actually put in front of students tomorrow.

Mastery-based learning is a proven practice that's brutal to run at scale — it needs constant reassessment, a clear map of what comes next, and materials for wherever each student actually is. Here's how progressions, differentiation, and assessment loops make it feasible with Claude for Teachers.

Differentiation is the highest-value, most time-expensive thing a teacher does — which is why it's usually the first thing to get cut. Here's why Claude for Teachers targets it, and how it becomes a minutes-long job.

Claude for Teachers brings OpenSciEd into the planning loop, so science lessons are grounded in a vetted curriculum instead of generated from scratch. Here's what teaching from trusted materials buys you that raw generation can't.

The Claude for Teachers lesson-planning workflow, step by step: ask for a lesson, let Claude draw on curricula mapped to your state's standards and the progressions beneath them, and get back a plan plus student-facing materials you revise.

Differentiation, mastery-based learning, and small-group instruction reliably help students — and reliably cost teachers time they don't have. Claude for Teachers doesn't replace the practice; it becomes the planning partner that makes it feasible.

Snorkl's Claude for Teachers connector surfaces insights about classes, assignments, and student progress to inform instruction — closing the loop from what happened in class to what you plan next.

Most AI lesson tools write first and align later. Claude for Teachers draws standards and learning progressions at planning time, so lessons come out aligned and scaffolded by construction — not retrofitted with a standard code stapled on at the end.

Both open-source Claude for Teachers skills ship per-subject reference files — ela.md, math.md, science.md, social_studies.md, plus learning-commons-kg.md. Here's why subject-specific guidance beats one generic prompt.

You control what class data goes into Claude, and nothing you share trains the model. Here's a practical, teacher-tested checklist for handling rosters, diagnostics, and notes responsibly.

TeachFX's Claude for Teachers connector gives personalized instructional feedback grounded in real classroom talk — feedback on your teaching, from your actual room, not a once-a-year observation.

Scheduled tasks in Claude for Teachers let you hand off reviewing each day's exit tickets and adapting tomorrow's plan — running every school day at 4pm, so the analysis is done by the time you get home.

Anthropic's train-the-trainer module, co-created with the American Federation of Teachers, solves a harder problem than the individual course: how one prepared person brings a whole staff along.

Stop asking whether AI is good for the classroom. Ask which tasks it suits. A practical, model-agnostic framework from Anthropic's CC-licensed teacher guidance — and the three-question gut check behind it.

The single hardest thing to make an AI tutor do is not answer. Withholding the solution and asking the next guiding question — the Socratic move — is the core design constraint that separates a tutor from an answer key.

A smarter base model doesn't automatically make a better tutor. Pedagogy is a behavior you install — a SKILL.md that tells the agent when to question, when to scaffold, when to assess — not something you get from more parameters.

Point AI at assessment instead of answers and it becomes a formative-assessment engine — generating quizzes, blueprints, and retrieval practice that reveal what a learner doesn't yet know.

You don't need forty education skills. A working AI tutor is four composable behaviors — question, explain-and-ground, drill, and assess — assembled from the marketplace. Here's the stack.

LLMs are extraordinary analogy engines — the core of a good explainer. They're also confidently wrong in ways a novice can't catch. Great AI explanation is analogy plus grounding, never analogy alone.

An AI that rewrites the student's essay teaches nothing. The skilled move is rubric-based feedback that diagnoses without doing the work — preserving the struggle where the learning actually lives.

A hallucinating tutor is worse than no tutor — it teaches errors with authority. Grounding a tutor in real tools (a math engine, references, a code runner) is a safety requirement, not a nice-to-have.

The durable win from an AI tutor isn't the physics answer — it's teaching the learner how to plan, self-test, and manage anxiety. Metacognition, not content, is what compounds.

Most ed-tech fads died. Spaced repetition — testing yourself at widening intervals against the forgetting curve — didn't. Here's why it's the highest-leverage thing to wire an AI tutor into, and the skills that do it.

Bloom's 2-sigma problem — that one-on-one tutoring beats classroom instruction by about two standard deviations — was a scaling impossibility for 40 years. Personalized AI tutors are the first plausible answer, with real caveats.

How a solo builder plus an AI agent ship on-brand UIs fast using a DESIGN.md from the registry. The leverage that closes the design gap without a designer.

A guided tour of the 135+ agent-ready design systems in the aiskill.market Designs category: the sub-categories, the standouts, and how to actually use them.

Every model launch ships with bar charts where the new model wins. Here's a reusable checklist for sanity-checking those numbers — using Grok 4.5's July 2026 launch as the worked example.

Grok 4.5 ranks #4 on the independent Intelligence Index at a price >60% below Opus 4.8 and GPT-5.5. Here's what that price-per-intelligence unlocks for high-volume skill and agent workloads.

xAI reports Grok 4.5 using ~4x fewer output tokens than Opus 4.8 on SWE-Bench Pro. For long agentic loops where output dominates the bill, that can matter more than winning the benchmark.

'On par with GPT-5.5/Codex at lower cost' sounds decisive — until you notice it's a harness score, not a model score. What that means when you pick a coding backend for a skill.

Grok's DeepSearch is an iterative RAG loop with live X integration. Treated as an agent primitive, it changes what a research skill can answer. Here's how it compares to Claude tool-use plus web search.

Stop hardcoding one model. Build a router that sends cheap high-volume work to Grok 4.5, hard long-horizon work to Fable 5, and the balanced middle to Opus 4.8 — with a decision table you can copy.

GPT-Live does live translation, but its Hindi demo landed with a heavy American accent. That gap is the case for purpose-built translation voice skills on gpt-realtime-translate — here's when a demo isn't enough.

SIP telephony turns a marketplace voice skill into a real inbound and outbound calling agent. Here's the architecture, the support use cases it unlocks, and the cautions before you dial out.

The Realtime API's remote MCP, image, and SIP support let a voice agent call the same skills your text agents use. Here's the architecture — and a wiring diagram as a list.

OpenAI's GPT-Live makes always-listening full-duplex voice the default in ChatGPT. Here's why that resets how you design voice-first agents — and what to build on.

GPT-Live shipped without video or screen sharing — you fall back to legacy voice for those. Here's a route-to-X-when-Y decision list for where a custom Realtime-API skill fills the gaps.

OpenAI shipped two voice things in one week. GPT-Live is the ChatGPT app; gpt-realtime is the API you build on. Here's which track to pick and why it matters.

gpt-realtime-2.1 adds configurable reasoning effort and function calling over live voice. Here's how to build an agent that plans a step, calls your function, then answers — out loud.

Grok 4.5's API is OpenAI- and Anthropic-SDK compatible. Adding it to an agent is a base-URL and model-string swap, not a rewrite. Here's the wiring and where it fits next to Claude.

Grok 4.5 is fast and cheap — but the spec sheet has four gotchas that break naive integrations: 500K (not 1M) context, image-in but text-only out, ~17s first token, and voice that isn't the LLM.

xAI's launch page pits Grok 4.5 against Opus 4.8. The honest read is a 2-of-4 split with Claude Fable 5 leading all four. Here's how to separate first-party numbers from independent ones.

A skeptic's field guide to Grok 4.5. What's independently verified, what's only xAI-claimed, and what it means if you build skills and agents on top of frontier models.

Without design context, AI agents produce generic, AI-looking UIs. A DESIGN.md gives them the colors, type, and spacing to build on-brand work instead.

The dev-tool design aesthetic decoded: dark-first, dense, keyboard-driven systems from Raycast, MongoDB, and Linear-like tools — and how to give your agent the look.

Fable 5 requires 30-day data retention and is not available under zero-data-retention. A ZDR org gets a 400 on every request. Here's the compliance decision guide for teams shipping skills.

GPT-Live can say 'mhmm', stay quiet while you think, and let you cut in mid-sentence. Turn that full-duplex behavior into concrete UX rules and a checklist for your own voice-agent skills.

OpenAI's GPT-Live ships predefined voices only, blocks cloning, and builds in teen safety. Translate those safeguards into a concrete publishing checklist for responsible voice skills.

How fintech and crypto design systems engineer trust through restraint, precise typography, and generous spacing — and how to give your agent the same signals.

Fable 5 costs roughly 2x Opus 4.8. That doesn't mean run everything on the cheaper model. Here's a routing heuristic for deciding which turns deserve the premium tier.

gpt-realtime-2.1 cut p95 voice latency by 25%+ via caching. Here's the latency budget of a voice turn, the cached-vs-fresh audio economics, and when to reach for the mini variant.

A practical Claude Code workflow for handing your agent a DESIGN.md so it builds on-brand UIs. Includes the exact prompt and project setup to use today.

Fable 5 performs strongly with a markdown memory system. Here's the exact format Anthropic recommends, how to package it as a reusable skill, and why one lesson per file beats a growing scratch log.

A style guide is a human PDF; a DESIGN.md is an agent-parseable spec. Here's the head-to-head on format, audience, and which one your AI agent can actually use.

Fable 5 dispatches parallel subagents more readily than older models. Spawn-and-block leaves that capability on the table. Here are the orchestration patterns that actually pay off.

Encode a type scale, font family, weights, line-height and letter-spacing as DESIGN.md tokens so your AI agent ships consistent, readable text every time.

Fable 5 turns can run for minutes and autonomous runs for hours. If your harness assumes sub-30-second replies, it will time out mid-thought. Here's how to redesign for long turns.

Define semantic color tokens and pass WCAG contrast for an agent-ready palette. A step-by-step guide to building an accessible DESIGN.md color system.

Fable 5 replaces the token budget with five effort levels. Here's how to route skills and subagents — routine work at low-medium, long-horizon at xhigh-max — plus what Claude Code's ultracode really is.

Convert Figma variables and design tokens into a DESIGN.md file your AI agent can read. A step-by-step guide to tokens-as-code for on-brand output.

On Fable 5, asking the model to explain its reasoning or show its work can be refused with category reasoning_extraction and elevates fallbacks. Read the summarized thinking blocks instead.

Nine standout product design systems — Stripe, Discord, Spotify, Raycast, Humane, MongoDB and more — and the one thing to steal from each for your AI agent.

Fable 5 refusals arrive as HTTP 200 with stop_reason refusal — not an error. Check it before reading content, and wire a server-side or client-side fallback to Opus 4.8. Code and a checklist inside.

How Humane's design system imagines an interface with no screen — projection and voice, black canvas, white foreground, laser-red accent — captured in an agent-readable DESIGN.md.

How Raycast's design system nails the keyboard-first developer aesthetic — sharp dark UI, tight spacing, and monospace accents, captured in an agent-readable DESIGN.md.

Anthropic says skill prompts written for older models are often too prescriptive and degrade Fable 5's output. Here's how to audit yours and swap step-by-step scripts for goal-and-constraints.

How Spotify's design system goes bold and dark — its brand-green accent, expressive large-scale typography, and media-forward layouts, captured in an agent-readable DESIGN.md.

How Discord's design system is built for always-on use — its Blurple brand, dark-first surfaces, rounded friendly components, and high chat readability, captured in a DESIGN.md.

Anthropic's Claude Fable 5 changes six things at once for anyone building skills and agents: always-on thinking, an effort dial, 1M context, new pricing, a refusal stop reason, and 30-day retention.

How Stripe's design system balances technical precision and luxury — its restrained palette, crisp typography, and generous spacing, captured in an agent-readable DESIGN.md.

Stitch-style design systems package design tokens and rationale into one agent-readable file. Here's Google's Stitch pattern and why a standard format matters.

How AI agents turn a DESIGN.md file into on-brand components: the workflow that replaces generic AI output with UI that actually matches your product.

Design tokens are named, reusable values for color, type, spacing, and radii. Here's what they are and how AI agents consume them to build on-brand UIs.

Trust is a design material: confidence, citations, transparency, rubrics, guardrails, and handoffs. The series closer — skills set the floor, taste is yours.

A DESIGN.md file is an agent-readable design spec: YAML design tokens plus markdown rationale that teaches AI agents to build on-brand UIs, not generic output.

A system prompt is an interface. Architect it like one — labeled anatomy, important instructions first and last, parameterized templates, and testable constraints.

A survey of the 2026 loop engineering ecosystem: elorm Loops, explainx, awesome-agent-loops, continuous-claude, Gas Town, and the aiskill.market Loops channel.

Top 10 free AI tools (skills + agents) for solopreneurs in 2026 — covering ops, marketing, support, content, and finance. Install in 60 seconds.

The next design surface isn't the layout — it's how the product behaves. Context budgeting, turn-taking, generative UI, and frustration repair, in one skill.

A mental-model shift for 2026: moving from one-shot prompting to closed-loop systems built on feedback, verification, and bounded autonomy.

Direct 3D and motion graphics by prompt. Claude writes the Blender bpy Python and After Effects ExtendScript over MCP while you art-direct the result.