Dev & Data Insights — Libretto

Dev&Data

Getting the Most Out of Claude Code: How to Save Your Mistakes to Memory

Why the moves that matter aren't the commands you ran, but the judgment calls behind them

Every time you open a new Claude Code session, you may find yourself re-explaining your project structure and re-fixing mistakes you already corrected yesterday. The difference between someone who ends up with a sharp, well-trained assistant and someone stuck retraining a new hire every time comes down to one thing: what you choose to save.

Dev&Data

FPV Drone Parts Compatibility: Why Picking the Best of Everything Still Fails

Motors, ESCs, and batteries aren't chosen one by one — they're chosen as a chain

There's a common way first-time builders pick parts: search for a well-reviewed motor, do the same for the ESC, then the battery. Each part is good on its own — but that's not enough.

Dev&Data

Does Making AI Agents Collaborate Actually Help?

If you've ever had multiple AI models review and critique each other's work to get a better result, a new paper suggests that approach may backfire — a single round of cross-talk can erase the very diversity that made using multiple models worthwhile in the first place.

Dev&Data

Why Checking Only the Final Answer Isn't Enough to Verify AI

When AI summarizes a contract or organizes product data, a clean-looking table with a citation attached usually makes people stop double-checking against the original. A recent study found a model that never opened the source datasheet still passed verification — by fabricating the citation itself — exposing why checking only an AI's final answer isn't enough.

Dev&Data

When Models Converge, Two Variables Remain: Harness and Copyright

As model choice matters less, execution architecture and training-data legality are becoming the variables that count.

As open-weight models close the performance gap with closed ones, which model you use matters less than it used to. What remains are the harness — the execution scaffolding that connects a model to real work — and the copyright uncertainty around training data. Here's what small teams should check now.

Dev&Data

Why Pre-Set Permission Rules Make AI Agents Less Safe, Not Safer

Once you start delegating tasks to an AI agent — from sorting email to making payments — you'll hit a "Should I go ahead with this?" prompt several times a day. More often than not, you probably click approve out of habit rather than reading it carefully, just to keep things moving.

Dev&Data

The Hugging Face Buyout Rumor Raises a Pointed Question: Which Side Is the Solo Founder On?

Now that vertical integration in AI infrastructure has reached all the way to model-sharing platforms, the options open to solo operators who lean on the open ecosystem are starting to narrow.

In mid-May 2025, a rumor out of Silicon Valley rippled through the AI developer community: Hugging Face was reportedly in talks to be acquired in a deal worth $13 billion. Hugging Face currently hosts more than 500,000 public models and over 100,000 datasets, and nearly every open-source AI developer passes through the platform at some point to download or share a model.

Dev&Data

What Is Vibe Coding? How to Build an App Without Knowing How to Code

AI conversation is replacing syntax — and reshaping who gets to build software

Vibe coding — building apps and services through nothing but conversation with AI — is the talk of the tech world, with more people asking whether non-developers can really create software, and if so, what made it possible.

Dev&Data

The Docs AI Agents Actually Read — And How to Write Them

If you've ever tried rewriting your API docs or cleaning up a Notion ops manual, you've probably paused at the question of how to write it so AI coding tools like Claude Code will actually reference it well — a question that feels especially urgent for solo founders and product people running a service without a developer on staff.

Dev&Data

How AI Agents Came to Operate Your Computer: What Actually Changed

Tools like Claude Code, Operator, and Copilot now handle coding, file organization, and repetitive tasks almost on their own. Here's how AI moved from rigid, rule-based systems to agents that read a screen, judge the situation, and pick their own next move.

Dev&Data

AI Coding Agents Don't Fail at the Shell — They Fail on the Way There

A new study called QuoteBench shows why the "success" an agent reports may have little to do with what actually happened on your machine.

For solo founders using AI coding agents, a "task complete" message isn't proof that files or settings actually changed as expected. QuoteBench, a new study, shows how commands can get corrupted between generation and shell execution — and why marketing success-rate numbers can be misleading.

Dev&Data

성수동 카페 상권의 매출은 얼마나 될까?

접근 장벽이 낮아진 자리에 검증 책임이 들어섭니다

성수동 카페 평균 매출이 어떻게 되는지를 챗GPT에게 물으면, 특정 데이터를 근거로 삼은 답변이 돌아오는 시대가 됐습니다. 빅밸류가 전국 사업자·상권·카페 시장 등 20개 분야, 515만 페이지 분량의 데이터를 AI가 읽을 수 있는 형태로 개방했습니다. 이용자가 생성형 AI 서비스에 질문하면 이 데이터를 근거로 답을 만드는 방식입니다.

Dev&Data

How to Turn Claude Code Into an AI That's Actually Yours

Everyone reads the same tips. Only some people get a Claude Code that keeps getting smarter.

Another week, another "must-save Claude Code tips" listicle — yet applying the same tips to the same tool, some people get exactly what they want in a few sentences while others explain themselves from scratch every time. This piece traces where that gap comes from and lays out how to grow Claude Code into an AI that's genuinely yours, using CLAUDE.md and its memory feature.

Dev&Data

Sizing Delegation in Claude Code: How Much to Hand Over at Once

A complete guide to designing your unit of delegation

Handing Claude Code an entire project often ends in a mess — not because the AI can't handle the work, but because of how the work was chopped up. This piece walks through a three-step method for sizing delegation correctly so the results actually hold up.

Dev&Data

What You Lose by Feeding AI More Customer Data: The Complaints Not Yet Voiced

Generative AI can't surface what customers haven't said yet

A consumer goods company built its next product roadmap on a RAG system's analysis of years of customer feedback — then watched a competitor build a whole new category from just 15 interviews the AI never saw. MIT Sloan Management Review's latest research maps where generative AI genuinely helps understand customers, and where its reach quietly runs out.

Dev&Data

Coding No Object: How to Get Work Done With Claude Code

As "vibe coding" spreads, non-developers are wondering what skill actually matters now — and it isn't syntax.

Planners and marketers with no coding background are already building their own automation tools by simply talking to Claude Code. This piece explains where to start if you've never written a line of code, and what capability matters more than tool fluency once you get going.

Dev&Data

When Fable 5 Won't Run in Claude Code, Just Turn Off the Advisor

Turn off the advisor. That's it.

Launch Claude Code in the Visual Studio terminal, pick the newly released Fable 5 model, and a red error appears out of nowhere: a 400 error saying claude-opus-4-8 can't be used as an advisor for the request model. Here's why, and how to fix it in under five minutes.

Dev&Data

Inside Prompt Injection: How It Works and How to Defend Against It in Layers

You Gave the AI Data. It Read a Command.

Anyone running an AI system in production eventually hits the same question: is user-submitted text data, or is it a command? Prompt injection — the most common and hardest-to-solve threat in AI security — is what happens when a system gets that question wrong.

Dev&Data

The Claude Code Slash Command Cheat Sheet Worth Bookmarking

A complete reference to Claude Code's slash commands — as of May 2026

Even people who use Claude Code every day rarely know its full set of slash commands. As of May 2026 there are more than 50, each with a distinct purpose — here's the complete, categorized reference.

Dev&Data

These Claude Code Shortcuts Will Double Your Speed

Getting 100 percent out of Claude Code

Plenty of developers use Claude Code, but surprisingly few ever learn its keyboard shortcuts. A practical tour of the shortcuts and slash commands — from the Esc+Esc rewind to /btw — that unlock the rest of the tool.

Dev&Data

The AI Workspace: What to Do When Agents Start Holding the Context

What to do as AI agents move into the place where your context already lives

Notion's May 2026 developer platform lets AI agents run inside the workspace itself, ending the era of hauling context between tabs. For solo operators, the real shift is that human work moves from connecting tools to making judgment calls.

Dev&Data

A Storefront in One Day Without Engineers: Treating AI Adoption as a Product

One CEO's playbook for designing AI adoption as a product—not a culture project

When a Sendbird marketer launched a live swag store in a single day without any engineering help, it revealed how the company had turned AI adoption into an internal product—with quests, tokens, and a leaderboard—and what solo founders can borrow from that playbook.

Dev&Data

The Limits of Vibe Coding: Notion's Product Lead on More vs. Better

Notion's product lead on what agency really means — and the widening gap between making more and making better

Max Schoening, Notion's product and design lead, argues that vibe coding has produced more software but not better software — and that as AI tools commoditize skills, the ability to define the right problem and judge whether the output is good enough is becoming the human's core job, especially for solo founders.

Dev&Data

Getting Ready for AI Transformation: Why Your Company Needs a Single Source of Truth

The infrastructure that matters more than the model itself: SSoT, a single source of truth

Companies rolling out enterprise AI keep discovering the same problem: the AI gives different answers depending on which department asks, because the underlying documents disagree. Before investing in models or GPUs, organizations need a single source of truth (SSoT) — one agreed-upon home for every critical piece of data.