Claude Code effort ladder uses Jev to handle low-value decisions

Khairallah AL-Awady @eng_khairallah1Originally posted 2026.10.12

Open original on X

holy f*ck. it's in Anthropic's own docs: stop trusting your settings file to run Opus 5.5 on high. it ignores that line and stays on medium lock the level onto the model with /effort run /effort high Opus 5.5 plans on high, saved to the model itself so it finally sticks Sonnet 5.5 workers build at medium while Haiku handles every search Fable 5.1 stays on call as the advisor, only stepping in at three critical moments: → before a plan locks: is the hardest file getting the highest effort? → when a test fails twice: is the agent underthinking it or overthinking it? → before calling done: did any change ship that only Haiku ever looked at? Haiku searches. Sonnet 5.5 builds. Opus 5.5 decides Jev engineering takes the ladder one rung lower: decisions that need no effort at all (which file, which tool, retry or stop) resolve in code, so no model burns a thinking token on them Read on Haiku. Build on medium. Decide on high. Audit on max. - the full effort ladder > Opus 5.5 on high runs the main session, saved per model under modelSettings > Explore swapped to Haiku for every read-only search > worker on Sonnet 5.5 at medium inside its own git worktree > auditor on Opus 5.5 at max, the level built for hunting security bugs > type ultrathink in any prompt for one deep turn without moving the dial paste the ladder and this prompt into Claude Code ↓ "Rebuild my Claude Code effort setup as a ladder: 1. Read ~/.claude/settings.json and tell me if there's a top-level effortLevel. It doesn't apply to Opus 5.5 in user settings, so move it into modelSettings: claude-opus-5-5 at high, claude-sonnet-5-5 at medium 2. Create these in ~/.claude/agents, skipping any that already exist: > Explore: model: haiku, tools: Read, Grep, Glob (replaces the built-in Explore) > worker: model: sonnet, effort: medium, isolation: worktree, owns the edits > auditor: model: opus, effort: max, tools: Read, Grep, Glob, Bash, security reviews only 3. Set advisorModel to fable, then report anything that would flatten or cap the ladder (CLAUDE_CODE_EFFORT_LEVEL, maxEffortLevel, CLAUDE_CODE_SUBAGENT_MODEL_FORCE). Change nothing 4. Add one rule to ~/.claude/CLAUDE.md: send security questions to auditor, keep max for the auditor only, and use ultrathink for a single hard turn instead of raising the session level Show every change as a diff first. No edits until I say go." save and bookmark this no matter what

This post quotes another post: x.com/eng_khairallah1/status/2109217163546693852