The Signal — April 21, 2026

Two stories today: one about the instructions hidden behind the curtain, and one about the silicon shortage that's making everything more expensive.

What Anthropic's System Prompt Changes Tell Us About Where Claude Is Going

Anthropic remains the only major AI lab that publishes the system prompts for its consumer-facing models. When Opus 4.7 shipped on April 16, the prompt got a meaningful update, and Simon Willison did the diff.

The changes are worth reading closely because system prompts are where a lab's actual priorities show up. Not the blog post version of priorities, but the real ones, written in instructions the model has to follow.

A few standout shifts:

Do it, don't ask. The new prompt tells Claude to make reasonable attempts at a task rather than interviewing the user first. If a tool could resolve ambiguity (looking up the user's location, checking a calendar) Claude should call the tool instead of asking the human to do the lookup. This is a clear push toward agentic behavior: act first, clarify only when genuinely stuck.

Talk less. New language tells Claude to keep responses "focused and concise" to avoid overwhelming users. Separately, a line from 4.6 banning the words "genuinely," "honestly," and "straightforward" has been removed, presumably because 4.7 no longer defaults to those verbal tics.

Respect the exit. If a user signals they want to end a conversation, Claude should stop. No more nudging for another turn. A small change that addresses one of the more annoying chatbot behaviors.

Expanded safety guardrails. The child safety section has been significantly expanded. New rules around disordered eating prevent Claude from giving specific nutrition targets or step-by-step diet plans. And a new defense against the "give me a yes or no" screenshot attack, where users force a controversial binary answer, now lets Claude decline the format and explain why.

Tool search before giving up. Claude now has access to a tool search mechanism. Before saying "I don't have access to X," it's supposed to check whether a matching tool exists but was deferred. This fits with Anthropic's broader tool-search architecture and their November 2025 engineering post on advanced tool use.

The pattern across all these changes: Anthropic is training Claude to be more autonomous, less chatty, and harder to exploit. The model that talks less but does more is the model that actually functions as an agent.

Sources: Anthropic System Prompts, Simon Willison, Claude 4.7 Release Notes

The RAM Shortage Could Last Until 2027, at Least

If you've noticed phones, laptops, VR headsets, and gaming handhelds getting more expensive this year, here's the structural reason: the world doesn't have enough DRAM, and it won't for a while.

According to Nikkei Asia, manufacturers are projected to meet only 60 percent of DRAM demand by the end of 2027. SK Group's chairman has said shortages could persist until 2030.

The math doesn't work. Nikkei reports that production would need to increase by 12 percent annually in 2026 and 2027 to meet demand. Counterpoint Research says only a 7.5 percent increase is actually planned. That's a gap that compounds every quarter.

The three companies that dominate global memory production (Samsung, SK Hynix, and Micron) are all building new fabrication capacity, but almost none of it comes online before 2027 or 2028. The one exception: SK Hynix opened a fab in Cheongju in February. That's it for 2026.

The new fabs are primarily focused on high-bandwidth memory (HBM) for AI data centers, not the general-purpose DRAM that goes into consumer devices. So even when production does ramp up, it's not clear how much relief phones, laptops, and gaming devices will see. AI's appetite for specialized memory is cannibalizing the supply chain for everything else.

As The Verge notes, the price increases have already hit across the board: Samsung phones, Microsoft Surface devices, Meta Quest headsets, and gaming handhelds have all gone up. This isn't a blip. It's a structural reallocation of a finite resource toward AI infrastructure, and consumers are footing the bill.

Sources: Nikkei Asia, The Verge, Gizmodo, Counterpoint Research

On the Editor's Desk

Five stories came in through research today. We ran with the two freshest: the Claude system prompt analysis (rare primary-source material on how a lab actually steers its model) and the RAM shortage (structural impact that touches everyone). Cerebras re-filing for IPO, Tesla's robotaxi expansion to Dallas and Houston, and the Anthropic/White House meeting all passed validation but were 3-4 days old. Covering them now with stale framing would have served nobody well. If any develop further this week, they'll get the treatment they deserve.