Prompting Claude Haiku 4.5
How to prompt Claude Haiku 4.5, the fastest, lowest-cost Claude tier. Read the shared Prompting guides principles first; this guide carries only what is distinctive to Haiku 4.5. It is the resolution of the portable fast model type (ki-tokenomics, ADR-KI-HARNESS-009).
When to reach for it
Haiku 4.5 is the tier for mechanical, bulk, and high-volume steps — the fan-out members of a sub-agent orchestration, bulk classification, scaffolding, and conform-style edits — where per-step reasoning depth is not the bottleneck and throughput and cost are. It reaches near-Sonnet coding quality (SWE-bench Verified ~73% vs Sonnet's ~77%) at roughly a third of the cost and over twice the speed, which is why the recommended pattern is a barbell: route the bulk of traffic to Haiku and escalate only the genuinely hard cases to a reasoning/frontier type. It is the first Haiku with extended thinking, computer use, and context awareness (it tracks its remaining token budget).
Prompting principles
| Principle | Steer |
|---|---|
| Bound the thinking | The single biggest lever for this tier. Limit steps, tokens, and scope explicitly ("think in 3–5 steps"; "target 120–180 tokens, never exceed 220"). It respects effort strictly at the low end — drop effort to cut thinking and cost. |
| Extended thinking is opt-in per task | Enable it for the complex cases inside an otherwise-cheap flow; thinking tokens bill as output. Leave it off for the mechanical bulk it is there to handle. |
| Prefer exemplars over prose | Short, targeted few-shots steer it more reliably than long instructions. Keep prompts tight. |
| Separate policy from task | It responds well to labelled, modular prompts ([Context], [Policy], [Task], [Output]) — modular prompts scale better across a fan-out than one monolithic instruction. |
| Be explicit, request "above and beyond" | Like the larger tiers it takes prompts literally and won't infer unrequested work; if you want more than the minimum, say so. |
| Lean on context awareness | It tracks its remaining context window — useful on long tool traces, but size max_tokens with headroom (200k context, up to 64k output) so thinking + response don't truncate. |
Orchestration
- Fan-out member. This is Haiku's home: the many parallel sub-agents in a
pipeline()/parallel()block where the lever is fan-out count, not per-step quality. Keep each member's prompt bounded and self-contained. - Escalation, not replacement. Default a flow to Haiku and escalate the specific steps that demonstrably need more — adversarial verify, hard synthesis — to a stronger type, rather than raising the whole flow's tier.