backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
Strip ~80% of your static guidance — Anthropic did, with no coding-eval loss on Opus/Fable 5 — Conflicting rules across CLAUDE.md/skills/system prompt force judgment models to thrash; run /doctor, keep repo gotchas only, delete “obvious” and always-on constraints. — x.com/trq212/status/2080710971228918066 · claude.com/blog/the-new-rules-of-context-engineering-for-cla
EMERGING: progressive disclosure (skill trees + deferred ToolSearch tools) is the new default harness shape — Move verification/review out of the always-loaded prompt into on-demand skills/tools; stop one mega-CLAUDE.md of every practice you might need. — x.com/trq212/status/2080710971228918066
Opus 5 is live: daily-driver executor; keep Fable for plan/brainstorm/hardest bugs — Official seat split from Claude Code staff; also flagged strong for long-running autonomous work, at ~half Fable price positioning. — x.com/trq212/status/2080703339306913985 · x.com/_catwu/status/2080707593115516985
Opus 5 + PI probes + Claude Code Auto Mode ≈ ~0 successful prompt-injection (Boris) — Highest-leverage security change for unattended loops / tool use on untrusted content; prefer this stack over raw YOLO for fleet runs. — x.com/bcherny/status/2080713091688583312
Prefer rich references over prose specs: HTML artifacts, test suites, port-from code, rubrics → verifier agents — Plan/spec quality jumps when agents judge against executable or high-fidelity references, not markdown essays. — claude.com/blog/the-new-rules-of-context-engineering-for-cla
EMERGING: “skill pack ≈ IaC?” — skill packs framed as deployable environment/config, not just prompts — If skill packs start owning flash/setup/repro of agent environments, your harness inventory should version skill packs like infra. — x.com/GeoffreyHuntley/status/2080472690516017456
Autonomous multi-CLI handoffs: agent writes a minimal handoff.md, then invokes codex/claude CLI (or a harness subagent) in a fresh context — Matches your conductor→executor pattern; steipete: don’t re-chat—spawn the other seat with only the failed-experiment residue. — x.com/steipete/status/2080692876665827485
Multi-model inside Claude Code is trivial if CLIProxyAPI already bridges Claude+Codex auth (gpt-5.6-sol in T3 Code) — Useful only if you want Sol/Codex as an in-harness peer without leaving Claude Code UI; low ops cost if proxy already up. — x.com/theo/status/2080535363207233575
Highest-leverage action today: run /doctor, prune CLAUDE.md + skills to progressive-disclosure gotchas, and A/B Opus 5 as default executor with Fable reserved for conductor/hard bugs.