Claude Design, Skill Stacks & Opus 4.7 Aftermath

A day after the Opus 4.7 / Codex Desktop double-launch, the conversation splinters: Anthropic Labs ships Claude Design; Matt Pocock publishes a full Claude Code skill lineup; LLMJunky drops a Codex Linux app; and the community starts asking "okay, so what workflow actually works with 4.7?"

Agentic & Code AI

Claude Design Launches — And Eats Your Files

Claude Design, a new Anthropic Labs product, lets you make prototypes, slides, and one-pagers by talking to Claude. It's powered by Opus 4.7 (Anthropic's "most capable vision model") and rolls out to Pro, Max, Team, and Enterprise.

Early user reports were mixed:

The design tool shipping one day after Opus 4.7 — paired with Theo's "files are gone" bug — reinforces the undercurrent complaint ("Claude desktop team does not use Claude desktop") from yesterday's launch cycle.

Matt Pocock's Full Skill Lineup

Matt published his evolving Claude Code skill roster:

  • /domain-model — replaces /grill-me, integrates DDD concepts, adds docs & ADR generation during the discussion
  • /to-prd — turn a discussion into a PRD
  • /to-issues — convert a PRD into issues with blocking relationships
  • /github-triage — state machine-based labelling system for existing issues
  • /tdd — structured test-first loop

This is the culmination of his "DDD is the answer to AI-age code" thesis from yesterday. Key framing shift: "'Improve the way you prompt the agent' has become a red flag. Commanding the agent isn't the right approach — collaborate with it, talk to it, reach consensus together." /grill-me (and now /domain-model) is the artifact of that philosophy; his aihero post describes the skill as asking "question after question until you reach a shared understanding," typical session ~45 min.

DDD scope limit per Matt: GOOD: Ubiquitous Language, Bounded Contexts, ADRs. BAD: Entities, Value Objects, Aggregates, Domain Events. i.e., use DDD to document the app without prescribing its shape.

Matt Pocock Announces "Slopwatch" — Coding Agent Observability

In the same livestream, Matt announced a new project: slopwatch, a coding-agent observability platform. The pitch: "Install a plugin which uploads all your session data to a dashboard so you can compare trends over time."

The meta-constraint: he'll only work on it while recording YouTube, explicitly to produce "watch me work" longform content. Complements the independent pi-trace visualization project ClementDelangue announced with HuggingFace — agent trace sharing is clearly crystallizing as a category.

LLMJunky Ships Codex App for Linux, Champions Multi-Agent Workflows

Codex App v26.415.20818 for Linux is live on GitHub — community port of the OpenAI desktop client. Computer Use plugin not available (Wayland limits accessibility; LLMJunky confirmed the same constraint for his GPT-5.4 demo).

On workflow: LLMJunky POV-posted "engineering my new saas with a team of 16 codex subagents." He's been advocating for multi-agent workflows since Codex got subagent support — same core argument as Matt's skill system but aimed at the orchestration layer rather than the ceremony layer.

Adjacent: Ahmad Osman's retweeted take: "If you buy a GPU you can basically have unlimited tokens while all your data stays private." The Codex-first partisan camp is hardening — someone in LLMJunky's thread: "You don't need to learn 1000 skills or download clunky dashboards to get started."

Simon Willison: LLMs Are Now Good For Legacy Code Too

Simon pushed back on a common belief: "Is there still a widespread belief that LLMs and coding agents are good for greenfield development but don't help for maintaining large existing codebases? I don't think that idea holds up any more."

Evidence from his own blog: Datasette 1.0a28 release notes include the note "Most of the changes in this release were implemented using Claude Code and the newly released Claude Opus 4.7." — this is not greenfield; Datasette is ~9 years old. Good counter-narrative to the /domain-model camp's concern that AI agents choke on unmodelled codebases.

Endor Labs: Cursor Is The Most Secure Harness (And Opus 4.7 Widened The Lead)

Endor Labs ran an independent evaluation of coding harnesses on functional and secure code output. Result: Cursor wins across the board, and Opus 4.7 widened the gap this week. Lauren (potetotes, Cursor) promoted it with "cursor agents help you go further and be more ambitious!" Cursor also launched Opus 4.7 with 50% off for a limited time.

LlamaIndex ParseBench: Content Faithfulness Deep-Dive

Jerry Liu followed up yesterday's Opus 4.7 OCR numbers with ParseBench's content-faithfulness methodology: "Did the parser capture all the text, in order, without making things up?" Three grading failure modes for every parse. Useful companion to the headline accuracy numbers — enterprise document pipelines fail on hallucinated content before they fail on accuracy.

Opus 4.7 Tuning Tips Keep Coming

Two new ones since yesterday:

  • Missing thinking summaries? Numman Ali flagged that Claude Code doesn't show thinking summaries for Opus 4.7 by default — the fix is claude --thinking-display summarized.
  • Default effort bumped silently: Matt PSA'd that Anthropic raised default Claude Code effort from medium to xhigh, which is going to burn more tokens than people expect. "Switch off the defaults and experiment to see what effort level suits your work best."

Anthropic's Safety Layer Over-Triggering Bans

Theo flagged: "Looks like the Anthropic 'safety layers' aren't just blocking prompts anymore, they're erroneously banning entire orgs." No firsthand screenshots in-thread, but worth monitoring — there's a steady pattern of Opus 4.7's aggressive injection-detection complaining about its own harness (e.g. Peter Szilagyi's observation carried over from yesterday).

Side Note: Server-Side Templates Are Back

Armin Ronacher, on the slop-induced simplification trend: "Now with clankers the right way to build a website once again turns out to be server side templates." Consistent with his ongoing anti-frontend-complexity arc and the /domain-model school's preference for rigid structure over over-engineered flexibility when agents are the main consumer.

AI Engineer Summit: The Slop Cannons vs. The Grown-Ups

Recap of yesterday's now-posted AI.Engineer talks:

Non-AI / Other

  • Armin on Rust: "I love Rust, but not unconditionally. The language has some obvious warts (compilation model, lack of self borrows, compile times) and I wish we could be all more adult about it." A rare in-community criticism.
  • PyCon US 2026: Long Beach, May 15–17. First West Coast PyCon US since Portland 2017; includes new AI and security tracks. Simon encouraging late Californian RSVPs.
  • steipete's TED talk on OpenClaw happened this week in Vancouver — prepared in a week while working on OpenClaw and OpenAI simultaneously (context).
  • T3 Code crossed 50k users: Theo reported the milestone. Stats: 53% arm Macs, ~25% Windows, Linux is 5× more popular than Intel Macs.