WisyLink Blog
Field notes — Page 4
Practical guides and field notes on building, launching and running websites, apps and tools with AI.
Page 4
Continuing the archive.
Earlier dispatches from the field, in publish order.
Spec-Driven Development: When the Spec Becomes the Product
What spec-driven development is, how spec-first, spec-anchored and spec-as-source differ, and when writing the spec before the code pays off with AI coding tools.
9 min read
Phantom Packages: The Supply Chain Risk in Generated Code
When AI-generated code names a package that doesn't exist, anyone can register that name. How often it happens, why the names repeat, and how to check a dependency before you install it.
9 min read
The 70% Problem: Why Generated Software Needs a Human Last Mile
AI tools get a project most of the way quickly; the last stretch of edge cases, integrations and hardening still needs human judgment. Here is why, and how to cross it.
10 min read
Agent Memory: How Context Is Kept Across Sessions
How AI agents carry context from one session to the next: the memory types, why long context isn't enough, how retrieval works and why forgetting is a design decision.
10 min read
Evaluating What Is Generated: Eval Discipline and Measured Trust
How to evaluate AI-generated output with a repeatable eval set: which cases to collect, how to score answers with no single right one, and how to catch regressions.
10 min read
Generative UI: Interfaces Composed at Runtime From Intent
What generative UI is, where it already ships, and how teams keep screens that are composed per request consistent, trustworthy and testable.
10 min read
Intent Fidelity: The Gap Between What You Describe and What Is Produced
Why AI output drifts from what you meant, how to tell description loss from transfer loss, and a short feedback loop that closes the gap between your request and the result.
10 min read
Intent Debt: When Generated Code Outpaces Comprehension
When software is generated faster than anyone can explain it, a new kind of debt accrues: deferred understanding. What intent debt is, why clean code hides it, and how to pay it down.
10 min read
Multi-Agent Orchestration: When One Loop Becomes Many
Splitting one AI agent into many buys specialization and charges a coordination tax. What the 2025 evidence shows, the common failure modes, and when one agent is the better choice.
10 min read
Vibe Coding: The Discipline and Limits of Building Software by Intuition
Vibe coding, named by Andrej Karpathy in 2025, trades reading code for describing it. Where that speed pays off, where it quietly builds debt, and the discipline that keeps it useful.
10 min read
Taste Is the New Bottleneck When Cheap Execution Floods Work
AI can produce a hundred competent options in an afternoon, so choosing the right one is now the hard part. What taste is, how abundance erodes standards, and how to make judgment teachable.
10 min read
Disposable Software: Apps Built to Be Thrown Away
AI made working apps cheap enough to throw away. When a single-use app is the right call, how to give it an expiry date, and how to stop it turning into permanent, unowned software.
11 min read