Techegic · ISSUE #3 · May 22, 2026
THE BRIEF
The engineers who matter stopped writing code weeks ago. They're writing specifications in HTML, generating throwaway interfaces to think through problems, and treating AI tokens like a resource to be allocated—not a tool to be used. If you're a tech professional who still measures productivity by lines shipped, this issue is the wakeup call.
THIS WEEK'S SIGNAL...
...came through Lenny's Newsletter Lenny's Newsletter delivers tactical product management and growth insights through deep-dive interviews with operators at companies like Airbnb, Stripe, and Anthropic. If you're building products or leading teams, this is one of the few newsletters that consistently surfaces how top practitioners actually work—not just what they ship. → https://lennysnewsletter.com
THE SIGNAL
Anthropic Engineers Are Using HTML Instead of Markdown—And It's Redefining What "Coding" Means
Thariq Shihipar, an engineer on Anthropic's Claude Code team, just revealed how the company's builders actually work with AI coding tools. The headline shift: they've replaced Markdown with HTML for planning and implementation workflows. The reasoning is simple and devastating to traditional workflows—richer visual formats get better human engagement, which produces better products. HTML lets you brainstorm with interactive mockups instead of bullet lists. It turns specs into living, editable design systems that travel with your codebase.
But the real signal isn't the file format. It's the new mental model Shihipar articulated: the "compute allocator" mindset. His claim—backed by how Anthropic's own team operates—is that 99% of your AI-generated tokens should go to planning, interfaces, and communication, not production code. Let that sink in. The engineers building the AI coding tools themselves spend almost all their AI compute on thinking, not typing.
The math says this makes sense. When Claude can write implementation code in seconds, the bottleneck isn't generation—it's specification. The engineer who spends three hours prompting for perfect code is optimizing the wrong variable. The engineer who spends those three hours building a visual spec that Claude interprets correctly in one shot is playing a different game entirely.
What nobody's talking about is what this does to the career ladder. The technical skills that got rewarded for decades—syntax fluency, debugging speed, implementation patterns—are becoming commodities. The skills that now compound are specification writing, visual communication, and what Shihipar calls knowing when "complexity has to earn its keep." Translation: knowing what to build matters more than knowing how to build it.
BY THE NUMBERS
The data behind this week's Signal.
|
45 Planning & Specs |
30 Interfaces |
24 Communication |
Source: Lenny's Newsletter, May 2026
THE CAREER PLAY
What this means for your positioning
What this means for your positioning
1. Build one HTML-based spec this week instead of your usual Markdown doc. 2. Track what percentage of your AI tokens go to planning versus implementation. 3. Reframe your value pitch from "I write code" to "I allocate compute effectively."
Q: What skills do software engineers need in 2026 to stay competitive with AI coding tools?
A: The highest-value skill is specification quality—the ability to describe what you want built in enough detail that AI executes correctly on the first pass. Engineers who master HTML-based visual planning, system design thinking, and compute allocation strategy will outperform those who only know how to write implementation code.
THE 2-MINUTE MOVE
Do this today.
Open your last technical spec or planning document. Ask yourself: is this a wall of text, or could someone understand the architecture just by glancing at it? If it's the former, you're still operating in the old paradigm.
Take your current project's main workflow and describe it—not in Markdown bullets, but as a visual HTML mockup. Use Claude or any AI tool to generate a single-page interactive version of your spec. Don't polish it. Don't make it pretty. Just see how different it feels to plan in a visual medium where you can actually click things. The whole point is that it's throwaway—Anthropic calls this approach "just-in-time documentation."
QUICK HITS
Three more moves on your radar
Enterprise AI Agents Keep Failing Because They Forget What They Learned
RAG-based systems are good at surfacing relevant documents—and that's exactly where they stop (VentureBeat). Startup Rippletide is building "decision context graphs" that give AI agents structured memory, time-aware reasoning, and non-regressive action sequences. For tech professionals: if you're implementing AI agents at work, the architecture choice between simple RAG and memory-enabled systems will determine whether your project actually delivers value or becomes another demo that falls apart in production. The play is learning to spec memory requirements before you build—not after it fails.
Google I/O Put AI Everywhere—For Better and Worse
Stratechery's analysis of Google I/O 2026 highlights the tension between DeepMind's research ambitions and Google's business objectives. Demis Hassabis claimed we're "standing in the foothills of the singularity," while the actual product announcements were a scattered mix of incremental features. The career read: Google's AI strategy looks like internal empire-building, not coherent product vision. If you're evaluating Google tools for your stack, wait for the dust to settle before betting your workflow on any single announcement.
Waymo Pauses Service in Four Cities After Robotaxis Drive Into Floods
Waymo has now suspended operations in Atlanta, San Antonio, and paused freeway rides after its vehicles kept driving into flooded roads and struggling with construction zones (TechCrunch). The robotaxis passed standard tests but failed edge cases that any human driver would handle. This is the gap between demo-ready and production-ready—a distinction that matters enormously if you're building or evaluating autonomous systems. Shipping AI that works 95% of the time is shipping AI that fails catastrophically 5% of the time.
THIS WEEK'S QUESTION
One question. One reply. Takes 2 seconds. What percentage of your AI tool usage goes to planning versus implementation?
Hit reply — I read every response.
Techegic is published every Friday.
You're receiving this because you're building your career like a business.
Disclaimer: This is career intelligence for educational purposes. We provide the signal; you provide the judgment. Techegic and its parent, Egic Holdings, are not a law firm, medical board, or financial advisory. Use your brain.
Techegic PREMIUM — WHAT'S BELOW THIS LINE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
→ THE DEEP DRILL: Anthropic Engineers Are Using HTML Instead of Markdown — The Execution Layer
→ THE CAREER ASSET: COMPUTE ALLOCATOR SCORECARD
→ THE STACK PICK: Claude Code (claude.ai/code or via API) — Claude Code (claude.ai/code or via API)
→ THE IRREPLACEABILITY CHECK: One question. Answer it honestly....
The execution layer is below this line.
The Deep Drill, Career Asset, Stack Pick, and Irreplaceability Check are Premium-only. The free Signal tells you what's happening. Premium tells you what to do about it.
Unlock Premium — $9/month →What's waiting below this line::
- The Deep Drill — the specific execution layer on this week's Signal
- The Career Asset — one plug-and-play resource ready to deploy this week
- The Stack Pick — one tool, step-by-step, for your role
- The Irreplaceability Check — one question that stress-tests your edge