Claude Code Daily Briefing - 2026-03-28

Release Summary

Full release notes

New Features & Practical Usage

Claude Mythos — “Most Capable” Next-Gen Model Confirmed (3/26)

A CMS misconfiguration at Anthropic exposed unpublished blog drafts, revealing the existence of Claude Mythos. An Anthropic spokesperson officially confirmed the model represents “a step change” and is “the most capable we’ve built to date.”

• New tier: Capybara tier above Opus — Mythos is the first model in this tier
• Performance: “Dramatically higher scores” on coding, academic reasoning, and cybersecurity benchmarks vs. Opus 4.6
• Cybersecurity: Described as “far ahead of any other AI model” in cybersecurity — Anthropic itself acknowledges this as a serious risk factor
• Status: Limited testing with early-access customers only — general release delayed due to cost and capability concerns

Fortune exclusive | The Decoder | Seeking Alpha

Developer Workflow Tips

Claude Code Creator Boris Cherny’s 100-Line CLAUDE.md Workflow

Boris Cherny, Staff Engineer at Anthropic and the creator of Claude Code, publicly shared his workflow. While the community often writes 500–1,000+ line CLAUDE.md files, Boris keeps his at ~100 lines (~2,500 tokens) — every line earns its place by solving a real problem.

Key principles:

• Self-improvement loop: Every time Claude makes a mistake, add a rule to lessons.md to prevent recurrence
• Plan Mode for non-trivial tasks: Always enter Plan Mode (Shift+Tab) before complex work
• Liberal subagent use: Keep context clean by delegating to subagents
• 10–15 parallel sessions: 5 terminal tabs + 5–10 web + mobile, each with its own git worktree
• “Claude Code is now 100% written by Claude Code”

MindWired AI | VentureBeat

Anatomy of the .claude/ Folder

A detailed breakdown of the .claude/ directory structure is gaining attention for developers looking to optimize their Claude Code setup.

• CLAUDE.md: Core guidelines (recommended under 200 lines)
• CLAUDE.local.md: Personal overrides not committed to Git
• rules/: Modular rules organized by topic
• commands/: Custom slash commands
• skills/: Auto-triggered workflows based on conversation context
• agents/: Sub-agents with independent system prompts

The key takeaway: start minimal and expand based on actual needs. Over-configuration often degrades results rather than improving them.

Daily Dose of DS

Security & Limitations

Pentagon Ruling: Judge Grants Preliminary Injunction — “First Amendment Retaliation,” “Orwellian” (3/26)

Judge Rita Lin issued a 43-page ruling granting Anthropic’s preliminary injunction request. The order bars the Trump administration from implementing, applying, or enforcing the president’s directive blacklisting Anthropic.

Key quotes from the ruling:

• “Punishing Anthropic for bringing public scrutiny to the government’s contracting position is classic illegal First Amendment retaliation”
• “Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government”

What happens next:

• One-week delay granted for government to appeal
• Pentagon CTO claims “ban still stands” — legal battle continues to final verdict
• Final verdict could be months away

CNBC | TechCrunch | NPR

Anthropic CMS Data Leak — ~3,000 Unpublished Assets Exposed (3/26)

The Mythos model leak originated from a CMS misconfiguration (“human error”). Researchers from LayerX Security and Cambridge University discovered the exposure and reported it to Fortune.

• ~3,000 unpublished assets including blog drafts, product plans, and internal documents were publicly searchable
• Details of a private CEO event were also exposed
• Anthropic removed public access immediately after being notified

Fortune | Fortune — Cybersecurity risk

Copyright Settlement Claim Deadline D-2 (3/30)

The $1.5 billion copyright settlement claim deadline is in 2 days. Approximately 100,000 authors are eligible.

Android Headlines

Ecosystem & Plugins

2x Usage Promotion Ends Today (3/28)

The Claude 2x usage promotion expires today at 11:59 PM PT.

• When: Weekday off-peak hours, weekends 24h
• Who: Free, Pro, Max, and Team plans

Claude Help Center

Community News

• Mythos leak shakes the industry: Following Fortune’s exclusive, active discussions across Hacker News and tech communities. Excitement about Capybara tier capabilities is tempered by concerns over cybersecurity implications. Fortune

• Tech industry welcomes Pentagon ruling: Judge Lin’s “Orwellian” language is being hailed as a landmark for AI companies’ right to refuse government contracts. However, the Pentagon CTO insists the ban remains in effect pending final resolution. Breaking Defense

• Anthropic-linked fund drops post-IPO: A fund holding Anthropic stakes extended its decline in what Bloomberg calls a “stunning IPO reversal,” attributed to uncertainty from the Mythos leak and ongoing legal disputes. Bloomberg

Minor Changes Worth Knowing

• API Session ID header: X-Claude-Code-Session-Id added to API requests for proxy-level session aggregation (v2.1.86)
• Jujutsu/Sapling VCS support: .jj and .sl directories added to VCS exclusion lists — reduces unnecessary file scanning for non-Git VCS users (v2.1.86)
• Read tool compact format: Line numbers now use a compact format and unchanged re-reads are deduplicated, reducing token consumption (v2.1.86)
• Marketplace plugin permissions fix: Official marketplace plugin scripts no longer fail with “Permission denied” on macOS/Linux (v2.1.86)
• macOS startup improvement: Keychain cache interval changed from 5s to 30s, reducing startup event-loop stalls (v2.1.86)
• VSCode “Not responding” fix: The VSCode extension no longer shows “Not responding” during long operations (v2.1.86)

Recommended Reads

• “Shell Tricks That Make Life Easier”: A practical compilation of terminal productivity techniques — line editing with CTRL+W/U/K/Y, reverse history search, brace expansion, process substitution. The key philosophy: master fundamentals incrementally rather than memorizing everything at once. Hofstede Blog

• “Why Daily Coding Is Limited to 4 Hours”: Cognitive psychology research shows deep work capacity maxes out at 3–4 hours daily. Analysis of 250,000+ developers found the actual median coding time is just 52 minutes per day. A single interruption takes 23 minutes to recover from; full context rebuilding takes 30–45 minutes. Tech World with Milan

• “Only Two Paths Left in Software”: An analysis arguing that AI coding tools are bifurcating software development into two lanes — highly specialized systems programming requiring deep expertise, and an application layer where AI generates most of the code. GeekNews discussion

Interesting Projects & Tools

• Ghostmeet: A self-hosted AI meeting assistant. Records meetings, generates automatic summaries, and extracts action items — all processed locally with no external service dependencies. Ideal for teams where data privacy is paramount. GeekNews

• OpenDocuments: An open-source RAG platform for searching organizational documents with natural language. Aggregates scattered docs across systems into a unified searchable interface. GeekNews