Gather — 29 July 2026

Boris Cherny at YC Startup School: Opus 5 Needs Less Prompting, Not More #

• Boris Cherny: Building Claude Code (Y Combinator Startup Library, 2026-07-27) / Boris Cherny: We Deleted 80% of Claude Code’s System Prompts for Opus 5. The Model Got Smarter. (BigGo Finance) — At YC Startup School, Cherny confirmed and extended the “80% system prompt cut” detail from the July 21 Cat/Thariq fireside chat with a stated team practice: “delete all of the codebase, delete all of the prompt, and start from scratch every time” a new frontier model ships, rather than incrementally patching the old harness onto it. Concrete claim: Opus 5 under Claude Code’s auto mode “can go for days, weeks, months at a time. It just won’t stop.” Cherny separately posted on X that Opus 5 is Anthropic’s “least prompt injectable model yet,” with the strongest defence coming from layering model alignment, prompt-injection probes, and Auto Mode together rather than any single safeguard.
• Prompting Claude Opus 5 (Anthropic, Claude Platform Docs) — First model-specific prompting guide seen in this journal, distinct from the general “Claude prompting best practices” page. Core reversal from prior-model guidance: give the complete task specification upfront rather than drip-feeding instructions; stop telling Opus 5 to “double-check” or “re-verify” since it already does this unprompted and the instruction just adds cost; explicitly ask for shorter output if brevity is wanted — effort level controls how much the model thinks, not how much it writes; default to low/medium effort and raise only when needed.
• Anthropic Removed 80% of Claude Code’s System Prompt. Here Is What They Learned. (Developers Digest) — Synthesis of the Hacker News discussion (197 points, 133 comments) around the system-prompt-cut story: developer reception is that Opus 5 is markedly better at error recovery and multi-step task completion but can over-check and over-engineer simple requests when prompted with old, heavily-scaffolded instructions written for weaker models.

MCP Protocol 2026-07-28 — Stateless Core, Enterprise Auth, Task Extensions #

• MCP 2026-07-28 spec: stateless core, coming to Claude (Anthropic) / The 2026-07-28 Specification (Model Context Protocol Blog) — The largest MCP protocol revision since launch: the initialize handshake and Mcp-Session-Id are removed, moving MCP from a bidirectional stateful protocol to a self-describing request/response model that can run on serverless and edge infrastructure. Authorization now aligns with production OAuth 2.0/OIDC deployments, so MCP servers can connect directly to enterprise identity providers (Entra, Okta) without workarounds. Three standardised extensions ship alongside the core: MCP Apps (server-rendered UI), Tasks (async/long-running operations), and Enterprise Managed Auth (IdP-based org-wide provisioning). A new Multi Round-Trip Requests (MRTR) mechanism allows mid-tool user confirmations (e.g., “delete this row?”) without holding a persistent connection open. Anthropic cites 400M+ monthly MCP SDK downloads (4x this year) and 950+ servers in Claude’s connectors directory as the adoption backdrop for the rewrite.

Claude Code Changelog — Another Silent-Failure Bug in the Hooks System #

• Claude Code changelog (Anthropic, ~2026-07-28) — Two fixes worth flagging for anyone relying on hooks or memory: hooks returning exit code 2 were silently failing to block the tool call as documented whenever the hook’s stdout JSON failed schema validation — the hook would appear to run correctly but its block instruction was dropped without any error surfaced. Separately, memory frontmatter values were being silently truncated at an inline # character when memory files are saved. Both continue the “shipped quietly, fixed quietly” pattern named as a recurring theme in the 2026-07-23 gather — this time the silent failure is in the enforcement mechanism (hooks) itself, which is a more consequential place for it to occur than a cosmetic bug.

Security — “ClaudeBleed” Named, Still Unpatched #

• ClaudeBleed Reopened: Browser Extensions Can Still Push Claude for Chrome to Read Your Gmail (Manifold Security) — Update to the Gap flagged in the 2026-07-23 gather: the Claude for Chrome flaw (malicious extension impersonates Claude’s own site to trigger Gmail/Docs/Calendar reads) now has a name — “ClaudeBleed” — and remains unpatched as of July 20, roughly two months after Manifold’s May 21 disclosure and across nine released versions (through v1.0.80+). Manifold’s technical detail: the root cause is a content-script click handler that forwards prompts to Claude’s side panel without verifying the click was genuinely user-initiated, exploitable in six lines of JavaScript.

Cross-links #

• Claude Integrations MCP 2026-07-28’s OAuth/OIDC alignment to enterprise identity providers and the new Enterprise Managed Auth extension is a direct integration-layer capability upgrade for anyone building or deploying MCP connectors at org scale.
• AI Code Architecture Cherny’s “delete the codebase, delete the prompt, start from scratch every model release” practice, paired with the official Opus 5 prompting doc’s reversal of prior scaffolding advice, reframes harness/prompt design as something rebuilt per model generation rather than incrementally tuned — directly relevant to how agentic coding architectures should treat model upgrades.
• Data, IP & Training Rights ClaudeBleed is a second consecutive gather where a Claude browser-extension flaw remains unpatched weeks after disclosure — the same “AI browsing tool as personal-data exfiltration vector” pattern tracked for web_fetch’s “Memory Heist” two gathers ago, now recurring in a different Claude surface (Chrome extension vs. claude.ai browsing tool).

Meta-observations #

• Emerging pattern: Four independent sources (Cherny’s YC talk, the official Anthropic prompting guide, a Developers Digest HN-analysis piece, and community commentary) converged within days of Opus 5’s July 24 launch on the same non-obvious claim: for this model, less prompt engineering produces better results than the scaffolding practices that worked for prior models. This is a stronger, more corroborated signal than the usual single-source launch coverage.
• Quality signal: platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5 is the first model-specific prompting guide tracked in this journal, distinct from the general “Claude prompting best practices” page — worth checking whether Anthropic makes per-model prompting docs a standard practice for future releases.
• Gap: ClaudeBleed (Claude for Chrome) remains unpatched at next check — now over two months since disclosure. Worth continuing to track status explicitly rather than letting it drop, per the pattern flagged in the 2026-07-23 gather.
• Method note: MCP protocol-level spec changes (2026-07-28) sit right at the boundary between this journal (developer-facing Claude Code/API practice) and claude-integrations (platform/enterprise integration). This gather placed it here because the practical implications (serverless MCP servers, OAuth alignment) are immediately relevant to anyone building Claude tool integrations day-to-day; future cycles should keep splitting protocol-level MCP news from Managed-Agents/enterprise-integration news based on which audience the story primarily serves.