What Two New Papers Tell Us About Prompting Tactics
April 4, 2026
D.A.D. today covers 12 stories from 3 sources. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
D.A.D. Joke of the Day: My AI assistant is great at generating ideas. Unfortunately, so is my four-year-old, and neither of them understand why we can't just "make the house a spaceship.
What's New
AI developments from the last 24 hours
Anthropic Cuts Off Subscription Access for Third-Party Coding Tools
Anthropic announced that starting April 4, Claude subscriptions can no longer be used with third-party tools like OpenClaw—users must switch to pay-as-you-go billing for continued access. The company claims these tools 'put an outsized strain on our systems' but provided no specifics. Anthropic is offering a one-time credit equal to one month's subscription price (redeemable by April 17) and up to 30% discounts on prepaid usage bundles. Community reaction has been skeptical, with users on Hacker News questioning why using paid token capacity constitutes unusual strain.
Why it matters: If you use Claude through third-party coding tools rather than Anthropic's native interfaces, your costs are about to become less predictable—and this signals Anthropic is tightening control over how subscribers access its models.
Discuss on Hacker News · Source: news.ycombinator.com
Artemis II Crew Photographs Earth From Deep Space—First Time in 54 Years
NASA released the first high-resolution images of Earth taken by humans traveling beyond Earth's orbit since 1972. The Artemis II crew captured the photos after completing their trans-lunar injection burn Friday, now more than 200,000 miles from Earth. The four-person crew is on track to loop around the Moon's far side on April 6 and return to Earth on April 10. NASA published a side-by-side comparison with an Apollo 17 image, underscoring the 54-year gap since humans last traveled this far.
Why it matters: This is a genuine milestone—the first crewed deep-space mission in over five decades, signaling NASA's return to lunar exploration and validating the Artemis architecture meant to eventually put humans on the Moon and Mars.
Discuss on Hacker News · Source: bbc.com
Claude Subscribers Can Claim Up to $200 in Bonus Credits—With a Catch
Anthropic is offering one-time bonus credits to paid Claude subscribers—$20 for Pro users, up to $200 for Max and Team plans—to mark the launch of its new usage bundles feature. The catch: you must enable extra usage billing to claim, and credits expire after 90 days. The promotion runs April 3-17, 2026, and excludes Enterprise and Console accounts. Community reaction has been mixed, with some users reporting technical errors when claiming and others viewing the offer skeptically as a way to get subscribers to turn on overage billing.
Why it matters: Free credits are nice, but the real play here is getting paid users to opt into usage-based billing—a signal that Anthropic is pushing toward consumption pricing as competition with OpenAI and Google intensifies.
Discuss on Hacker News · Source: support.claude.com
What's Innovative
Clever new use cases for AI
Personal Blog Aggregator Launches for Discovering Individual Writers
A developer launched a frontpage aggregator specifically for personal blogs—essentially a Hacker News-style discovery feed, but for individual writers rather than tech news links. The site displays recent posts from curated personal blogs with titles, sources, and timestamps. At launch, it featured 39 posts from blogs including Simon Willison's Weblog and others, all from the previous seven hours.
Why it matters: For professionals using AI writing tools or maintaining thought-leadership blogs, this signals continued interest in discovering human-written content outside algorithmic social feeds—a potential distribution channel worth watching.
Discuss on Hacker News · Source: text.blogosphere.app
What's in the Lab
New announcements from major AI labs
OpenAI Adds Pay-as-You-Go Option for Codex
OpenAI announced pay-as-you-go pricing for Codex, its AI coding agent, now available to ChatGPT Business and Enterprise customers. Previously, teams had to commit to subscription tiers; the new model lets them pay based on actual usage. OpenAI says this gives teams a lower-friction entry point to test the tool before scaling up. Specific pricing details were not released.
Why it matters: Lowers the barrier for teams evaluating AI coding assistants—you can now pilot Codex without committing to a fixed seat count, making it easier to compare against alternatives like GitHub Copilot or Cursor.
What's in Academe
New papers on AI and its effects from researchers
AI Agents Fail Safety Tests When Conversations Get Longer
Researchers released ATBench, a safety benchmark designed to test AI agents across multi-step tasks rather than single prompts. The benchmark includes 1,000 conversation trajectories averaging 9 turns each, with nearly 2,000 tool invocations, specifically constructed to reveal safety failures that emerge over longer interactions—the kind enterprises worry about when deploying agents that book travel, manage calendars, or handle sensitive workflows. Testing showed even frontier models and specialized safety systems struggle with the benchmark, suggesting current guardrails may miss risks that only surface across extended agent sessions.
Why it matters: As companies deploy AI agents for real work, this research suggests single-turn safety testing may dramatically underestimate the risks of letting agents operate autonomously over longer task sequences.
Robots Learn to Combine Sight and Touch Like Humans Do
Researchers developed a framework called Cross-Modal Latent Filter (CMLF) that helps robots combine what they see and what they touch to understand object properties—mimicking how humans integrate multiple senses. The system uses Bayesian inference to build a structured understanding of physical properties like weight or texture from both visual and tactile input. Interestingly, the robot showed perceptual quirks similar to humans, including susceptibility to cross-modal illusions where one sense misleads another. Real-world experiments showed improved performance over baseline methods.
Why it matters: This is robotics research, not a product—but robots that can feel and see like humans do would matter for manufacturing, logistics, and any setting where machines handle unfamiliar objects.
Emotional Prompts Don't Help Much, Study Finds
New research tested whether emotional framing in prompts—phrases like 'I'm really stressed about this' or 'This is important to me'—actually improves LLM performance. The finding: it's a weak signal, not a game-changer. Static emotional prefixes produced only small accuracy shifts across math, medical, and reasoning tasks, with effects most variable in socially-grounded queries. The researchers also built EmotionRL, an adaptive system that selects emotional framing per-query, which outperformed fixed emotional prompting—but gains remained modest overall.
Why it matters: This challenges viral prompting advice suggesting emotional appeals reliably boost AI output—the effect exists but is too inconsistent to build workflows around.
Rephrasing Prompts in Common Language May Boost AI Performance
New research proposes that LLMs perform better when given more commonly phrased inputs—and that deliberately rewriting prompts into more frequent expressions can improve results. The 'Textual Frequency Law' framework suggests the models essentially learned language patterns proportional to how often they appeared in training data. The researchers tested this across math reasoning, translation, commonsense tasks, and tool use, finding that paraphrasing unusual phrasings into common ones boosted performance.
Why it matters: If validated, this offers a practical prompting technique: when an AI struggles with a query, try rephrasing it in more conventional language rather than assuming the task itself is too hard.
Video Model Controls Seven Game Characters With Independent Commands
Researchers developed ActionParty, a video generation model designed for AI-driven games that can control up to seven characters simultaneously—each responding to independent commands. Current video models struggle with "action binding," confusing which character should perform which action when multiple subjects are on screen. ActionParty addresses this by separating how the scene renders from how individual characters move, tested across 46 game environments.
Why it matters: This is research-stage work, but it points toward AI-generated games where multiple players or NPCs act independently—a prerequisite for procedurally generated multiplayer experiences.