Cross-cutting frontend topics, tooling, and DX Compiled for immediate developer deployment.
calendar_todaysummarizeWeek 16-2026
article
Making Your Site Visible to LLMs: 6 Techniques That Work, 8 That Don't
TAG: AI-VISIBILITY
Evil Martians outlines six concrete techniques for making website content legible to LLMs and AI agents, ranked by impact and effort. The core approach centers on shipping clean Markdown at every URL via .md routes, advertising those routes through both HTML link rel="alternate" tags and HTTP Link headers, and maintaining an /llms.txt index file at the site root. Content negotiation using the Accept: text/markdown header is identified as the most standards-compliant and durable technique, since tools like Claude Code and Cursor already send this header. The post also debunks eight popular AI SEO tactics that lack evidence, including meta tags, HTML comments, User-Agent sniffing, and JSON-LD structured data. Critically, a Princeton/IIT Delhi study is cited showing that only enriching visible text (direct quotes, statistics, authoritative citations) measurably improves AI visibility.
Addy Osmani synthesizes the emerging discipline of agent harness engineering, arguing that a coding agent equals model plus harness and that the harness is where most of the real engineering leverage sits. The post details the anatomy of a harness: system prompts, AGENTS.md, tools, MCP servers, sandboxes, orchestration logic, hooks, and observability. A key insight is the ratchet principle — every agent mistake becomes a permanent rule in AGENTS.md or a hook, making the harness progressively tighter. Techniques for combating context rot include compaction, tool-call offloading, progressive skill disclosure, and full context resets for long-horizon tasks. The post also covers planner/generator/evaluator splits, sprint contracts, and the Harness-as-a-Service (HaaS) trend where SDKs like the Claude Agent SDK provide the loop, tools, and context management out of the box. Benchmarks show that the same Claude Opus model jumped from Top 30 to Top 5 on Terminal Bench 2.0 by changing only the harness.
Vercel announces the general availability of Vercel Workflows, a durable execution SDK that eliminates the need for a separate orchestration service by embedding workflow logic directly in application code using "use workflow" and "use step" directives in TypeScript or Python. Each step gets automatic retries, persistence, observability, and durable continuation, all backed by an event log, Fluid compute, and Vercel Queues. Deep integration with AI SDK enables infinitely long-running durable agents, while the getWritable() API provides durable streams that clients can disconnect from and reconnect to without losing state. Since launching in beta, the platform has processed over 100 million runs across 1,500+ customers. The post details real-world usage at Mux (video AI pipelines), Durable (parallel AI steps for website creation in under 30 seconds for 3 million SMBs), and Flora (50+ image model orchestration). Workflows 5 is in development with native concurrency locks and a snapshot-based runtime to reduce replay overhead.
Iasmim Cristina provides a practical introduction to Jujutsu (JJ), a version control system created by Google's Martin von Zweigbergk that uses Git as its storage backend while rethinking the workflow model. Key differences from Git include the absence of a staging area (every file change is automatically tracked in the working copy), stable change IDs that persist across commit rewrites, and commits that can record conflicts rather than blocking progress. JJ's automatic rebase ensures that modifying any commit instantly rebases all its descendants, and the revset functional query language lets users select revision ranges with expressions like root() and all(). The post walks through practical workflows — the Squash Workflow (emulating Git's staging), the Edit Workflow (inserting new changes before the current working copy), and bookmark creation — using jj git init, jj describe, jj new, jj squash, and jj edit commands with terminal output. The VisualJJ VSCode extension is recommended for visualization while learning.
How to Keep Human Experts Visible in Your AI-Assisted Codebase
TAG: AI-TOOLS
Daniel Nwaneri presents proof-of-contribution, a Claude Code skill and local Python CLI that tracks the provenance of AI-generated code back to the specific human sources that inspired it. The tool stores attribution records in a local SQLite database and exposes commands like poc.py trace (full provenance chain for any file), poc.py verify (static AST-based gap detection with zero API calls), and poc.py import-spec (seeding knowledge gaps from a spec-writer assumptions list before the agent writes code). The detection mechanism compares code structure against pre-seeded claims rather than relying on AI self-reporting, which the author argues is unreliable due to hallucination. Knowledge Gaps — parts of generated code with no traceable human citation — surface as deterministic findings that can be wired into CI via exit codes. The post frames the problem through the lens of Stack Overflow's 78% question volume collapse since 2023 and the breakdown of the human-knowledge feedback loop in AI-assisted development.
Michael Marsiglia, who runs a 100-person software consultancy, argues from direct observation that AI has widened the gap between expert engineers and novices rather than closing it, using the power-tool analogy: a skilled carpenter with a circular saw outperforms a novice who risks injury. Despite Anthropic research suggesting AI could theoretically automate 94% of tasks in computer and math occupations, observed usage covers only 33%, because the tools get harder to use as they get more powerful. Marsiglia contends that writing code was never the hard part of software engineering — understanding what to build, what to avoid, and what the consequences of failure are remains the irreducible human judgment. He also addresses the business-model implications via a discussion with a fellow consultancy leader, examining the Jevons Paradox applied to software delivery, fixed-price versus time-and-materials contracts, and how AI introduces a new lever — pushing more design and implementation to AI — when budget or scope pressures arise.
System Design Course: APIs, Databases, Caching, CDNs, Load Balancing & Production Infra
This comprehensive freeCodeCamp system design course, authored by Hayek, teaches mid-level developers the architectural thinking required to reach senior roles. Starting from a single-server setup, it progresses through vertical versus horizontal scaling, seven load balancing algorithms (round robin, least connections, least response time, IP hash, weighted, geographical, consistent hashing), and database selection criteria contrasting SQL (ACID transactions, complex joins) against NoSQL options (document stores, wide-column, graph, key-value). The course covers three primary API styles — REST, GraphQL, and gRPC — with a detailed treatment of HTTP/HTTPS, WebSockets, AMQP message queuing, and TCP vs UDP. Production infrastructure topics include caching strategies, CDN usage, big data processing pipelines, and avoiding single points of failure. The instructor frames each concept around real interview and job scenarios, emphasizing that companies pay senior rates for architectural decision-making, performance optimization, and judgment under ambiguous requirements.
Making Your Site Visible to LLMs: 6 Techniques That Work, 8 That Don't
Evil Martians outlines six concrete techniques for making website content legible to LLMs and AI agents, ranked by impact and effort. The core approach centers on shipping clean Markdown at every URL via .md routes, advertising those routes through both HTML link rel="alternate" tags and HTTP Link headers, and maintaining an /llms.txt index file at the site root. Content negotiation using the Accept: text/markdown header is identified as the most standards-compliant and durable technique, since tools like Claude Code and Cursor already send this header. The post also debunks eight popular AI SEO tactics that lack evidence, including meta tags, HTML comments, User-Agent sniffing, and JSON-LD structured data. Critically, a Princeton/IIT Delhi study is cited showing that only enriching visible text (direct quotes, statistics, authoritative citations) measurably improves AI visibility.
System Design Course: APIs, Databases, Caching, CDNs, Load Balancing & Production Infra
This comprehensive freeCodeCamp system design course, authored by Hayek, teaches mid-level developers the architectural thinking required to reach senior roles. Starting from a single-server setup, it progresses through vertical versus horizontal scaling, seven load balancing algorithms (round robin, least connections, least response time, IP hash, weighted, geographical, consistent hashing), and database selection criteria contrasting SQL (ACID transactions, complex joins) against NoSQL options (document stores, wide-column, graph, key-value). The course covers three primary API styles — REST, GraphQL, and gRPC — with a detailed treatment of HTTP/HTTPS, WebSockets, AMQP message queuing, and TCP vs UDP. Production infrastructure topics include caching strategies, CDN usage, big data processing pipelines, and avoiding single points of failure. The instructor frames each concept around real interview and job scenarios, emphasizing that companies pay senior rates for architectural decision-making, performance optimization, and judgment under ambiguous requirements.
Addy Osmani synthesizes the emerging discipline of agent harness engineering, arguing that a coding agent equals model plus harness and that the harness is where most of the real engineering leverage sits. The post details the anatomy of a harness: system prompts, AGENTS.md, tools, MCP servers, sandboxes, orchestration logic, hooks, and observability. A key insight is the ratchet principle — every agent mistake becomes a permanent rule in AGENTS.md or a hook, making the harness progressively tighter. Techniques for combating context rot include compaction, tool-call offloading, progressive skill disclosure, and full context resets for long-horizon tasks. The post also covers planner/generator/evaluator splits, sprint contracts, and the Harness-as-a-Service (HaaS) trend where SDKs like the Claude Agent SDK provide the loop, tools, and context management out of the box. Benchmarks show that the same Claude Opus model jumped from Top 30 to Top 5 on Terminal Bench 2.0 by changing only the harness.
Vercel announces the general availability of Vercel Workflows, a durable execution SDK that eliminates the need for a separate orchestration service by embedding workflow logic directly in application code using "use workflow" and "use step" directives in TypeScript or Python. Each step gets automatic retries, persistence, observability, and durable continuation, all backed by an event log, Fluid compute, and Vercel Queues. Deep integration with AI SDK enables infinitely long-running durable agents, while the getWritable() API provides durable streams that clients can disconnect from and reconnect to without losing state. Since launching in beta, the platform has processed over 100 million runs across 1,500+ customers. The post details real-world usage at Mux (video AI pipelines), Durable (parallel AI steps for website creation in under 30 seconds for 3 million SMBs), and Flora (50+ image model orchestration). Workflows 5 is in development with native concurrency locks and a snapshot-based runtime to reduce replay overhead.
Iasmim Cristina provides a practical introduction to Jujutsu (JJ), a version control system created by Google's Martin von Zweigbergk that uses Git as its storage backend while rethinking the workflow model. Key differences from Git include the absence of a staging area (every file change is automatically tracked in the working copy), stable change IDs that persist across commit rewrites, and commits that can record conflicts rather than blocking progress. JJ's automatic rebase ensures that modifying any commit instantly rebases all its descendants, and the revset functional query language lets users select revision ranges with expressions like root() and all(). The post walks through practical workflows — the Squash Workflow (emulating Git's staging), the Edit Workflow (inserting new changes before the current working copy), and bookmark creation — using jj git init, jj describe, jj new, jj squash, and jj edit commands with terminal output. The VisualJJ VSCode extension is recommended for visualization while learning.
How to Keep Human Experts Visible in Your AI-Assisted Codebase
Daniel Nwaneri presents proof-of-contribution, a Claude Code skill and local Python CLI that tracks the provenance of AI-generated code back to the specific human sources that inspired it. The tool stores attribution records in a local SQLite database and exposes commands like poc.py trace (full provenance chain for any file), poc.py verify (static AST-based gap detection with zero API calls), and poc.py import-spec (seeding knowledge gaps from a spec-writer assumptions list before the agent writes code). The detection mechanism compares code structure against pre-seeded claims rather than relying on AI self-reporting, which the author argues is unreliable due to hallucination. Knowledge Gaps — parts of generated code with no traceable human citation — surface as deterministic findings that can be wired into CI via exit codes. The post frames the problem through the lens of Stack Overflow's 78% question volume collapse since 2023 and the breakdown of the human-knowledge feedback loop in AI-assisted development.
Michael Marsiglia, who runs a 100-person software consultancy, argues from direct observation that AI has widened the gap between expert engineers and novices rather than closing it, using the power-tool analogy: a skilled carpenter with a circular saw outperforms a novice who risks injury. Despite Anthropic research suggesting AI could theoretically automate 94% of tasks in computer and math occupations, observed usage covers only 33%, because the tools get harder to use as they get more powerful. Marsiglia contends that writing code was never the hard part of software engineering — understanding what to build, what to avoid, and what the consequences of failure are remains the irreducible human judgment. He also addresses the business-model implications via a discussion with a fellow consultancy leader, examining the Jevons Paradox applied to software delivery, fixed-price versus time-and-materials contracts, and how AI introduces a new lever — pushing more design and implementation to AI — when budget or scope pressures arise.
Two major themes dominated general web development coverage this week: AI-era tooling and workflows, and foundational engineering practices. Evil Martians published the clearest LLM visibility guide to date — six ranked techniques centered on .md routes, HTML link rel="alternate" tags, HTTP Link headers, /llms.txt, and Accept: text/markdown content negotiation (already sent by Claude Code and Cursor), debunking eight popular AI SEO tactics including meta tags, HTML comments, and JSON-LD with citation from a Princeton/IIT Delhi study. Addy Osmani formalized agent harness engineering, arguing that the harness (system prompts, AGENTS.md, MCP servers, hooks, observability) is where most real engineering leverage lives — the same Claude Opus model jumped from Top 30 to Top 5 on Terminal Bench 2.0 by changing only the harness.
Vercel launched Vercel Workflows GA, embedding durable execution logic directly in TypeScript or Python via "use workflow" and "use step" directives — with 100 million runs across 1,500+ customers already processed in beta. The platform integrates with AI SDK for infinitely long-running agents and provides durable streams via getWritable(). Daniel Nwaneri introduced proof-of-contribution, a Claude Code skill and Python CLI that tracks AI-generated code provenance back to human sources using a local SQLite store, static AST analysis, and CI-wirable exit codes.
On the fundamentals side, TkDodo argued for vertical codebases (domain-first directory structure over technical-type directories) with eslint-plugin-boundaries enforcement. Jujutsu (JJ) got a practical walkthrough as a Git alternative with automatic rebasing and conflict-recording commits. A comprehensive freeCodeCamp system design course covered load balancing, SQL vs NoSQL selection, REST/GraphQL/gRPC, and CDN strategies for developers aiming for senior roles.
Key Takeaways
The most durable technique to make your site visible to LLMs is serving clean Markdown at .md routes and advertising them via HTTP Link headers and Accept: text/markdown content negotiation — Claude Code and Cursor already send this header.
Agent harness engineering is now a measurable discipline: the same model ranked Top 30 jumped to Top 5 on Terminal Bench 2.0 solely by improving the harness — invest in AGENTS.md, hooks, and observability before reaching for a bigger model.
Vercel Workflows GA embeds durable execution directly in application code via TypeScript/Python directives, eliminating the need for a separate orchestration service and enabling infinitely long-running AI agents with reconnectable streams.
