tomevault-io Prompt Engineering
Reference for writing and reviewing LLM system prompts. Based on Harada et al. (2024) — Curse of Instructions, EARS syntax, and ARQ pattern.
1. Minimize the Rule Count (Curse of Instructions)
The probability of the model following all rules simultaneously is approximately P = p^N, where N is the number of independently verifiable instructions and p is the per-rule adherence rate.
| N rules | p = 0.95 | p = 0.90 |
|---|---|---|
| 5 | 77% | 59% |
| 10 | 60% | 35% |
| 20 | 36% | 12% |
| 30 | 21% | 4% |
Target: 10 or fewer distinct rules in the system prompt.
The math is unforgiving — every additional rule multiplicatively reduces the chance the model follows all of them. This is why minimizing rule count matters more than perfecting any single rule.
- Move procedural/structural rules (e.g., parameter formatting, count constraints) into tool descriptions, not the system prompt.
- Merge rules that protect the same invariant into one EARS statement.
- Delete rules already enforced by the tool schema or the agent framework.
2. Write Rules in EARS Syntax
Easy Approach to Requirements Syntax produces unambiguous, testable rules. Free-form instructions invite misinterpretation because the model has to guess the scope and trigger condition. EARS removes that ambiguity.
Four forms:
| Form | Template | Use when |
|---|---|---|
| Ubiquitous | The system shall [action]. | Rule always applies |
| Event-driven | When [trigger], the system shall [action]. | Rule fires on a condition |
| State-driven | While [condition], the system shall [action]. | Rule applies during a state |
| Unwanted behavior | If [situation], the system shall [response]. | Error / boundary case |
Before (free-form): "NEVER output a URL that contains {product_name} as a literal string. Always make sure the product name is properly substituted before including any link."
After (EARS): "If a URL contains an unresolved template placeholder, the system shall omit the link and respond without it."
The EARS version is shorter, has one clear trigger, and one clear action — no room for partial compliance.
3. Five Failure Patterns to Detect
When reviewing a prompt, check for structural signals of each pattern. These patterns emerge from how prompts grow organically over time — patches on patches — and each one has a telltale signature in the prompt text.
| Pattern | What happens | Signal in the prompt |
|---|---|---|
| Focus loss | Model satisfies some constraints but silently drops others | Post-response self-check blocks; mandatory verification sections |
| Context misinterpretation | Query routed to wrong product or tool | Same routing rule restated in 3+ sections with slightly different wording |
| Hallucination | Model invents values for parameters it cannot resolve | Same constraint (e.g., "never output unresolved URLs") in both system prompt AND tool descriptions |
| Business protocol bypass | Access control, upsell, or boundary responses skipped | Multiple alternative response templates for the same situation (Option A / B / C) |
| Instruction ignored | Specific operational rules (count, format, single call) repeatedly violated | Rule exists only because the model kept breaking it — belongs in tool description, not system prompt |
4. ARQ Reasoning Scaffold (VERIFY Step)
Attentive Reasoning Query pattern: include a numbered VERIFY checklist as the last step of the workflow. Each item maps 1:1 to one EARS rule, forcing instruction-level chain-of-thought before output.
VERIFY before responding:
1. [EARS rule 1 — single sentence]
2. [EARS rule 2 — single sentence]
3. [EARS rule 3 — single sentence]
...
This makes multi-instruction compliance explicit and testable. Per Harada et al. (2024), instruction-level chain-of-thought improves all-rules compliance by 2-3x on GPT-4o and Claude 3.5 Sonnet.
Keep the VERIFY list to 8 items or fewer. If you need more, consolidate rules first — a long VERIFY list defeats the purpose.
5. Structural Rules
- One truth per rule. If the same invariant is stated in two places, pick the most authoritative location and delete the duplicate. Duplication is how hallucination patterns start (Section 3).
- Urgency markers devalue each other. Use CRITICAL/MANDATORY for at most 2 rules per prompt. If everything is critical, nothing is.
- Static before dynamic. Put the unchanging role/rules at the top. Inject dynamic context (product lists, user data) at the bottom. This enables prompt caching and keeps the model's attention on rules first.
- Examples cost tokens. Include examples only for non-obvious rules. One positive + one negative example per rule is enough.
- Patch-on-patch is a signal. If you find rule -> exception -> exception-to-exception chains, consolidate. Organic growth via successive fixes is the most common source of rule-count bloat.
Source: alfonsograziano/auto-agent — distributed by TomeVault.