ROUTING IMPERATIVE — read this first, every prompt

Default to delegation. Before responding to ANY user request:

1. Check the Cascade hook output (top of context — [CASCADE SKILLS], [CASCADE ROUTING], and [CASCADE MODEL] blocks). If a skill or agent appears with score ≥ 0.20, INVOKE IT before doing inline work. If [CASCADE MODEL] recommends Opus or Haiku, delegate the work via the Task tool with that model: parameter rather than running on the session's default model.
2. Scan the routing table below. If the request matches a row, MUST invoke the listed skill — do not paraphrase or do it yourself.
3. Spawn a specialist subagent for any task spanning > 1 file or requiring domain depth (SF, Python, frontend, security, etc.).

Bias toward skill/agent invocation, not inline action. A 299-skill stack only pays off if it fires. Every prompt is a routing decision first, an action second.

Read-before-edit discipline (CodeBurn-flagged: 6.8M tokens/mo)

Before editing any file, read it first. Before modifying a function, grep for all callers. Research before you edit. Edit-to-read ratio should be 1:4 or better — if you're editing without reading, you're causing retries that compound token waste.

Avoid re-reading the same file. If you need to look at <file> again, point Claude at exact lines or function names in your prompt: In <file> lines <start>-<end>, look at the <function> function. Top historical re-read offenders: sf_attribution_context.json, page.tsx, generate_all_digests.py.

Anti-patterns — NEVER do these inline:

If the work matches an anti-pattern row above, do not start typing — invoke the skill first.

Cascade Model Router (aggressive mode)

The Cascade UserPromptSubmit hook automatically classifies every prompt and emits a [CASCADE MODEL] directive when the work would be better handled by a Claude tier other than Sonnet 4.6 (the recommended session default). The router is silent for standard work — only Opus escalations and Haiku downgrades produce a directive.

The router is in aggressive mode: it leans toward Haiku for ambiguous prompts and demands strong evidence before recommending Opus. Asymmetric thresholds:

• Opus emits at confidence ≥ 0.85 (narrow, conservative)
• Haiku emits at confidence ≥ 0.65 (broad, eager)
• Plus a length-based fallback: prompts under 80 chars / 12 words default to Haiku unless a stronger pattern fires

Rationale: Opus on trivia wastes ~5× credits silently. Haiku failing on hard work wastes one turn the user retries with !opus. Recoverable failures are the cheaper failure mode.

How to act on [CASCADE MODEL] blocks:

Tier intent (aggressive mode):

• Opus 4.7 — GSD planning/debug/research skills, "design a [system|architecture|pipeline|layer|...]", security audits, strategic memos, board updates, mbb-frame, production-grade hard debugging (race conditions, intermittent failures, root-cause analysis)
• Sonnet 4.6 (silent default) — standard build tasks (implement/build/create/write + code noun), code review, bug fixes, test writing, doc updates, refactors, multi-file work, drafting messages
• Haiku 4.5 — confirmations (yes/lgtm/ship-it), typo/lint fixes, commit messages, Cascade/GSD quick commands, gws-* / recipe-* / sf-(soql\|metadata\|deploy\|...) mechanical ops, single-field updates, deletions, run/execute, lookups, and any short prompt without a stronger signal

Manual overrides — include in any prompt:

• !opus / --opus / force opus → escalate
• !sonnet / --sonnet / force sonnet → standard
• !haiku / --haiku / force haiku → downgrade

Kill switch: export CASCADE_MODEL_ROUTER=off disables the router for the session.

Tuning: decisions are logged to ~/Desktop/personal-os-main/.cascade/sessions/model-decisions.jsonl. Inspect with:

• node ~/Desktop/personal-os-main/scripts/hooks/model-router.cjs --stats — distribution of recent decisions
• node ~/Desktop/personal-os-main/scripts/hooks/model-router.cjs --tail 20 — last 20 with confidence + reason
• node ~/Desktop/personal-os-main/scripts/hooks/model-router.cjs "<prompt>" — classify a specific prompt

Thresholds and pattern lists live in ~/Desktop/personal-os-main/scripts/hooks/model-router.cjs — tune EMIT_THRESHOLD_OPUS, EMIT_THRESHOLD_HAIKU, SHORT_PROMPT_CHARS, SHORT_PROMPT_WORDS, or the MODEL_PATTERNS array directly.

Recommended baseline: Sonnet 4.6 is the configured default — set via the "model" key in settings.json, so new sessions start on Sonnet. Opt up to Opus per-session with /model opus (or !opus) for marathon ship-work and hard reasoning; the router still escalates the cases that genuinely need Opus and downgrades trivial work to Haiku. The content-aware guard (hasRealWorkSignal()) rescues short real-work prompts — troubleshooting, scoped questions, review/feedback — so they fall through to Sonnet instead of getting downgraded.

Skill Auto-Routing

When a user request matches an installed skill, MUST invoke it automatically via the Skill tool. Do NOT wait for a slash command — detect the intent and activate the skill.

Routing Table

Routing Rules

• Match on intent, not exact keywords. "Help me write a blog post" triggers content-research-writer. "Make my tweet better" triggers twitter-algorithm-optimizer.
• If multiple skills could apply, pick the most specific one. Ask only if genuinely ambiguous.
• Skills with supporting files store resources in ~/.claude/commands/skill-resources/{skill-name}/. Reference these paths when the skill instructions mention scripts, templates, themes, or examples.
• Some skills require external dependencies (Composio API key for connect/connect-apps, Playwright for webapp-testing, yt-dlp for video-downloader, langsmith-fetch CLI for langsmith-fetch, codeburn Node CLI for token-waste audits, impeccable Node CLI for HTML anti-pattern detection). If a dependency is missing, tell the user what to install and offer to set it up.

@RTK.md