Haiku 4.5 Prompting Guide
Prompt engineering patterns that exploit Haiku 4.5's strengths (speed, cost, routine tool use) while guarding against its documented behavioral flaws. Complements generic prompting skills with model-specific mitigations.
Scope boundary: This skill covers how to prompt Haiku effectively. For model selection (when to use Haiku vs Sonnet vs Opus), apply model-selection. For generic prompt structure patterns, apply prompting-guide. For reasoning directive quality (phrasing, timing, chain length), apply reasoning-calibration. For prompt file mechanics (.prompt.md format), apply prompt-file-design. For Sonnet-specific guards, apply sonnet-prompting.
When to Use This Skill
- Writing or reviewing
<constraints>for a Haiku-powered subagent - Designing methodology sections where Haiku will execute routine tasks
- Diagnosing poor output quality from a Haiku-powered subagent (hallucination, missed instructions, truncation)
- Optimizing prompt structure for Haiku's processing characteristics (speed, cost, capacity tradeoffs)
- Deciding which guard sections to include for a given Haiku task type
- Calibrating existing Sonnet guard patterns for Haiku's tighter operating envelope
Haiku 4.5 Flaw Catalog
Nine documented behavioral flaws from benchmark data, practitioner reports, and observed workspace behavior. Each is assigned a severity and primary mitigation. Flaws are categorized by their relationship to the Sonnet 4.5 flaw catalog (F1-F10).
Amplified Sonnet Flaws (worse in Haiku)
| ID | Flaw | Severity | Sonnet Equiv. | Delta | Description | Primary Guard |
|---|---|---|---|---|---|---|
| H1 | Reasoning depth wall | CRITICAL | F6 (3-4 hops) | 50-66% worse | Reliable at 1-2 hops only. Beyond 2, produces wrong conclusions, skipped steps, or premature "done". Any "if X then analyze Y then decide Z" chain exceeds capacity. | Task decomposition to ≤2 hops |
| H2 | Instruction fragility | HIGH | F4 (drift) | Faster onset | Drops IMPORTANT-tier constraints after 3-5 tool calls (vs Sonnet's ~10). CRITICAL constraints hold longer but the safe window is narrower. | Constraint compression — ≤3 CRITICAL |
| H3 | Synthesis incapacity | HIGH | F1+F10 | New failure class | Cannot connect disparate findings. Extracts data point A and B individually but fails at "A implies B because C." Tends to list findings without drawing conclusions. | Restrict to extraction, not analysis |
| H4 | Tool chain breakdown | HIGH | F3 (premature) | Worse | When task requires multi-tool sequences (grep → read → analyze), degrades at step 3+. Produces premature conclusions from step 1-2 results, skipping deeper investigation. | Single-tool-per-invocation design |
| H5 | Verification rubber-stamp | HIGH | F1+F8 | Significantly worse | Near-zero capacity for critical evaluation. If caller says "verify X works," Haiku confirms even with contradictory evidence. Cannot balance expectations vs evidence. | Never assign verification tasks |
| H6 | Output truncation | MEDIUM | F2 (lazy) | Moderately worse | Summarizes prematurely or drops tail content on long outputs. Cuts off earlier than Sonnet. Silent truncation without acknowledgment. | Explicit truncation markers |
| H7 | Hallucination under ambiguity | MEDIUM | F7 (tool halluc.) | Worse | When target not found, invents plausible file paths, function names, or parameter values rather than reporting absence. Less capacity to "know it doesn't know." | Mandatory NOT FOUND protocol |
New Haiku-Specific Flaws (not present in Sonnet)
| ID | Flaw | Severity | Description | Primary Guard |
|---|---|---|---|---|
| H8 | Format compliance erosion | MEDIUM | Follows structured output formats well on short tasks, but compliance degrades faster than Sonnet as task complexity grows. JSON schemas, tables, and markdown break under multi-step pressure. | Format re-anchor before output |
| H9 | Overly literal interpretation | LOW-MED | Interprets instructions literally, missing intent. "Find the test file for this module" returns only exact filename matches rather than understanding project conventions. Lacks contextual inference. | Explicit search strategy instructions |
Haiku Operating Envelope
Understanding Haiku's safe operating parameters prevents assigning tasks that exceed its capacity.
| Parameter | Haiku 4.5 | Sonnet 4.5 | Implication |
|---|---|---|---|
| Reliable reasoning hops | 1-2 | 3-4 | T0 tasks only (reasoning-strategy) |
| CRITICAL constraint cap | ≤3 | ≤5 | Tighter constraint budget |
| Tool calls before drift | 3-5 | ~10 | Shorter agent sessions |
| Structured output reliability | Short tasks only | Sustained | Re-anchor format on multi-step |
| Synthesis / connection | None | Basic | Extract only, no analysis |
| Self-correction | None | Limited | Parent must verify all output |
| Speed | 4-5x Sonnet | Baseline | Real advantage for high-volume work |
| Cost | 0.33x Sonnet | Baseline | 3 Haiku calls ≈ 1 Sonnet call |
Key insight: Haiku's operating envelope is T0 (Direct) on the reasoning-strategy tier model. Any task requiring T1+ reasoning should use Sonnet minimum.
Methodology
Phase 1: Verify Task Fitness
Before writing a prompt for a Haiku-powered subagent, confirm the task fits Haiku's envelope:
| Task Characteristic | Haiku Fit | If No → |
|---|---|---|
| Requires 0-1 reasoning hops? | ✅ Proceed | Use Sonnet |
| Single tool type per invocation? | ✅ Proceed | Use Sonnet |
| No evaluation/judgment needed? | ✅ Proceed | Use Sonnet |
| Output is data, not analysis? | ✅ Proceed | Use Sonnet |
| Session ≤5 tool calls? | ✅ Proceed | Add constraint anchor or use Sonnet |
| No synthesis across findings? | ✅ Proceed | Use Sonnet |
If ANY answer is "No" → either redesign the task to fit, or upgrade to Sonnet. No prompt trick compensates for a capability gap.
Phase 2: Assess Risk Profile
If the task fits, identify which Haiku flaws it's exposed to:
| Task Characteristic | Exposed Flaws | Required Guards |
|---|---|---|
| File reading / extraction | H7, H9 | NOT FOUND protocol, search strategy |
| Command execution | H6, H7 | Truncation markers, no-guess directive |
| Search / grep operations | H7, H9 | Literal interpretation guard, widening strategy |
| Multi-file parallel reads | H2, H8 | Constraint compression, format re-anchor |
| Doc navigation | H9, H3 | Explicit path instructions, no synthesis |
| Data transformation | H6, H8 | Output scope lock, format anchor |
Phase 3: Apply Guard Patterns
Inject guards matching the risk profile. Haiku has a tighter constraint budget than Sonnet — every guard must earn its place.
Priority 1 — Always include for Haiku subagents:
-
Anti-hallucination (guards H7): The single most important Haiku guard.
If target not found, respond "NOT FOUND: {what was searched}". NEVER guess, infer, or fabricate file paths, function names, or values. -
Constraint compression: Keep CRITICAL tier to ≤3 items. Haiku's adherence drops sharply beyond 3. Move overflow to IMPORTANT.
Priority 2 — Include based on risk profile:
-
Output scope lock (guards H6): Required when output may be long.
If output would exceed {N} lines, summarize with "[TRUNCATED: {count} additional items]". Never silently drop content. -
Search strategy (guards H9): Required for search/discovery tasks.
Search exact match FIRST, then broaden with pattern/glob. If zero results, try {alternative strategies} before reporting NOT FOUND. -
Format re-anchor (guards H8): Required for multi-step tasks with structured output.
Re-read the output_format section before writing your final response.
Priority 3 — Include for edge cases:
- Constraint anchor (guards H2): If session may reach 4-5 tool calls.
⚠️ Re-read CRITICAL constraints before proceeding.
Phase 4: Calibrate from Sonnet Guards
When adapting a Sonnet prompt for Haiku, apply these calibration rules:
| Sonnet Guard | Haiku Calibration |
|---|---|
<anti-sycophancy> | Remove — assign evaluation tasks to Sonnet instead (H5) |
<completeness> | Replace with output scope lock — Haiku truncates rather than placeholders |
<scope-fence> | Simplify — Haiku rarely makes bold changes (insufficient capacity for F5) |
<constraint-anchor> | Lower threshold — trigger at 3 tool calls, not 10 |
<reasoning_guidance> | Remove — if task needs reasoning, upgrade to Sonnet (H1) |
| CRITICAL ≤5 cap | Tighten to ≤3 (H2) |
| Convention anchoring | Unnecessary — Haiku doesn't generate code (H1 prevents it) |
Phase 5: Validate Prompt
| Check | What to Verify | Fix If Failing |
|---|---|---|
| CRITICAL cap | ≤3 items in CRITICAL tier? | Move overflow to IMPORTANT |
| Task fitness | Task fits Haiku operating envelope (Phase 1)? | Redesign or upgrade model |
| Guard coverage | All high-risk flaws from Phase 2 have guards? | Add missing guards |
| No evaluation | No judgment/verification/analysis expected? | Remove or upgrade to Sonnet |
| Output format | Structured format with concrete example? | Add format with example |
| NOT FOUND protocol | Anti-hallucination guard present? | Always add — non-negotiable |
Quick Reference: Guard Injection Patterns
Minimal guard blocks for Haiku subagent prompts. Each targets a specific flaw.
Anti-Hallucination (H7) — MANDATORY, 2 lines
If target not found, respond "NOT FOUND: {what was searched}".
NEVER guess, infer, or fabricate file paths, function names, or values.
Output Scope Lock (H6) — 2 lines
If output would exceed {N} lines, summarize with "[TRUNCATED: {count} items omitted]".
Never silently drop content — acknowledge what was omitted.
Constraint Compression (H2) — design rule
Keep CRITICAL constraints to ≤3 items.
Merge related constraints. Move all overflow to IMPORTANT tier.
Search Strategy (H9) — 3 lines
Search exact match FIRST, then broaden with glob/regex patterns.
If zero results after broadening, try alternative names or paths.
Report all search attempts before concluding NOT FOUND.
Format Re-Anchor (H8) — 1 line
Re-read the <output_format> section before writing your final response.
Constraint Anchor (H2) — 1 line, place at tool call 3
⚠️ Re-read CRITICAL constraints before proceeding.
Synthesis Fence (H3) — 2 lines
Report findings as extracted data only.
Do NOT draw conclusions, infer causality, or recommend actions.
Haiku vs Sonnet: Guard Selection Matrix
Side-by-side comparison for deciding which guards apply to which model:
| Guard Section | Sonnet | Haiku | Why Different |
|---|---|---|---|
<anti-sycophancy> | Required for eval tasks | Never — don't assign eval to Haiku | H5 makes it pointless |
<completeness> | Required for code gen | Replace with output scope lock | Haiku truncates, doesn't placeholder |
<scope-fence> | Required for write tasks | Simplify or remove | Haiku lacks capacity for bold changes |
<constraint-anchor> | At 10+ tool calls | At 3+ tool calls | H2 faster drift onset |
<reasoning_guidance> | For structured thinking | Never — upgrade model instead | H1 makes reasoning guards futile |
<anti-hallucination> | Optional (verify-first) | Always mandatory | H7 is Haiku's #1 risk |
<output-scope-lock> | Rarely needed | Often needed | H6 earlier truncation |
<search-strategy> | Rarely needed | For search tasks | H9 literal interpretation |
<synthesis-fence> | Not applicable | For multi-finding tasks | H3 prevents analysis |
<format-re-anchor> | Not needed | For multi-step tasks | H8 format erosion |
Anti-Patterns
- ❌ Assigning evaluation tasks — "Review this code" or "verify the fix" on Haiku. H5 (rubber-stamp) means it will confirm anything. Use Sonnet for all evaluation.
- ❌ Expecting synthesis — "Analyze these 5 files and explain the pattern." Haiku can extract from each file individually but cannot connect findings (H3). Use Sonnet for synthesis.
- ❌ Long tool chains — Tasks requiring 4+ sequential tool calls where each depends on the previous. Haiku degrades at step 3+ (H4). Redesign as independent parallel reads or upgrade to Sonnet.
- ❌ Overloading CRITICAL — More than 3 CRITICAL items. Haiku's constraint adherence drops sharply beyond 3, faster than Sonnet's 5-item cliff. Keep CRITICAL tight.
- ❌ Reasoning guards without model upgrade — Adding
<reasoning_guidance>to a Haiku prompt. No prompt trick fixes a 1-2 hop ceiling (H1). If the task needs reasoning, upgrade to Sonnet. - ❌ Trusting Haiku output without parent verification — Parent agents (Sonnet/Opus) must always verify Haiku subagent outputs. Haiku has near-zero self-correction capacity.
- ❌ Silent failure assumption — Expecting Haiku to report "I can't do this." Haiku will attempt any task and produce plausible-looking but potentially wrong output (H7). Design for failure at the architecture level.
- ❌ Using Sonnet guard patterns unmodified — Sonnet guards assume 3-4 hop reasoning, 5-item CRITICAL cap, and 10+ call sessions. All three are wrong for Haiku. Always calibrate per Phase 4.
Resources
- template.md — Haiku-specific calibration rules and flaw-to-section mapping
- examples.md — Concrete guard examples per flaw with before/after comparisons