Pick Model

Pick Model

Take the user's intended prompt as input. Do NOT execute it. Read the live session state, classify the prompt, emit verdict + delta + strategy: is the current model right, and if not, is switching worth the cache cost.

CLAUDE.md contract: this skill is the single source of truth for the model/effort routing call. CLAUDE.md "Execution defaults" carries no model table — it defers here by capability (auto-discovery on the description). Don't re-derive or duplicate the tier table elsewhere; if routing changes, it changes here.

Two levers, different cost (the core call):

• 🎚️ Effort change, same model → cache SURVIVES (cheap). Recommend freely.
• 🔀 Model switch → cache BREAKS, context re-read uncached (costly). Must beat the switch penalty.

A third lever — parallelism (linear vs fan-out, sub-agents vs Workflow) — is out of scope here. To design a skill/agent's per-step execution topology (and have it call this skill per step), use /pick-workflow.

Recognize-then-route: hit the right tier directly; reserve top-tier (Opus high / Fable) for ambiguous/big/can't-classify. Effort = output-spend, not input. See reference.md for principles, routing detail, escalators, examples.

Workflow

1. Read live state

pick-model-state

Emits model= ctx_tokens= ctx_pct= cache_hit_pct= cost_usd= duration_min= (jq + path resolution live in dstoic/scripts/pick-model-state, on PATH). Bare call — no shell expansion, so it's allowlistable and won't trigger a permission prompt. Don't compute these by hand. Fallback (prints state=missing, or command not found pre-sync): read model + ctx% from the statusline bar or ask user; note ⚠️ estimated state, never guess silently.

2. Classify + pick ideal model + effort

Match prompt to tier. Apply escalators (cap +1 tier). See cheatsheet + routing table in reference.md.

Effort sets: Sonnet 5 low|med|high|xhigh|max · Opus low|high|xhigh|max · Haiku none · Fable adaptive-only. ⚠️ Sonnet 5 default = medium (≈ Sonnet 4.6 @ high). Avoid xhigh/max on Sonnet 5: new tokenizer ≈ +30% tokens → can cost MORE than Opus for similar quality. Need that depth → switch to Opus, don't crank Sonnet effort.

3. Quantify switch cost (only if ideal model ≠ current)

awk -v ctx=CTX_TOKENS -v hit=CACHE_HIT_PCT -v rate=NEW_RATE \
  'BEGIN{printf "switch_penalty ~ $%.3f\n", ctx*(rate/1000000)*(1-0.9*hit/100)}'

Rates $/M in: Sonnet 5 = 3 ($2 intro thru 2026-08-31) · Opus 5 · Fable 10. Higher ctx% + cache hit = costlier switch; late-session switch for marginal gain = waste → prefer effort tweak.

4. Emit verdict

🎯 Verdict: [✅ Stay | 🎚️ Change effort | 🔀 Switch | ⬆️ Upgrade | ⬇️ Downgrade]
Current:  [🤖 model] [effort] · 🧠 [ctx%] · ♻️ [cache%]
Ideal:    [model] [effort]
Why: [task shape vs current capability, 1-2 lines]
[if switch] 💾 ~$[penalty] re-cached → [worth it because… | NOT worth it, tweak effort instead]
🧭 Strategy: [e.g. "start Sonnet 5 med, escalate to Opus high only if scope >3 files or security-sensitive"]

Logic: same model+effort ok → ✅ Stay · same model, effort off → 🎚️ tweak (cheap) · diff model, gain > penalty → 🔀/⬆️/⬇️ · diff model but penalty high + gain marginal → stay/tweak, state trade-off · chore/plumbing on Opus → ⬇️.

Anti-patterns

• ❌ Opus high as default — wastes tokens on chores/plumbing.
• ❌ Late high-context switch for marginal gain — re-cache > savings.
• ❌ Effort-as-input cost — it's output spend; on loops higher can be cheaper.
• ❌ Guessing current model — read state file or ask.
• ❌ Loading /claude-api for a model fact — use this skill's table.