Prompt Engineering Patterns

Compact, actionable patterns for writing agent prompts and skill files. See docs/PROMPT-ENGINEERING.md for full rationale and academic citations.

---

Constraint Budget (MOSAIC finding)

Range	Compliance
1–6	Reliable
7–15	Unpredictable
>15	Degraded

Rule: Max 6 hard constraints per section. Hard gates at the END of sections (Claude recency effect).

---

Register Shifting Checklist

Before finalizing any hard gate or enforcement directive, verify the verbs:

• Tier 1 (deep): audit, verify, critically analyse, enumerate — use for HARD GATEs
• Tier 2 (balanced): review, check, describe — use for standard task instructions
• Tier 3 (minimal): look at, mention, note — advisory only; never in enforcement

FORBIDDEN: Tier 3 verbs in REQUIRED or FORBIDDEN lists.

---

Persona Anti-Pattern

FORBIDDEN for analytical agents (reviewers, auditors, analysts, planners):

• ❌ You are a [role]... persona prefixes
• ❌ Act as an expert in... framing
• ❌ Imagine you are... setups

Source: PRISM (arxiv:2603.18507) — expert personas reduce knowledge-retrieval accuracy for analytical tasks.

Correct pattern: Role description + behavioral directive.

# BAD
You are a senior security engineer. Review this code.

# GOOD
Perform a security audit. Enumerate all CWE categories present.

---

HARD GATE Pattern Template

### HARD GATE: [Topic]

**[One-sentence scope statement]**

**REQUIRED** (all must be met):
- Specific action A
- Specific action B

**FORBIDDEN** — You MUST NOT do any of the following:
- ❌ You MUST NOT [prohibited action]
- ❌ You MUST NOT [prohibited action]

**Why**: [One sentence — increases compliance]

---

Model-Tier Compensation Template

<model-tier-compensation tier="[haiku|sonnet|opus]">
## Model-Tier Behavioral Constraints ([Tier Name])

- Do NOT [specific tendency to suppress].
- ALWAYS [specific required behavior].
</model-tier-compensation>

Rules:

• Include tier attribute for automated validation
• No meta-commentary about model weaknesses (model-visible; may reinforce behavior)
• Actionable directives only — positive framing where possible
• Place at TOP of agent prompt before role description

---

Self-Refine Loop (when quality matters)

Use for planning, security scans, and architecture tasks. 2–3 passes optimal (NeurIPS 2023).

## Pass 1: GENERATE
[Initial output]

## Pass 2: FEEDBACK
Critically analyse Pass 1. Enumerate:
- Gaps (missing)
- Errors (wrong)
- Ambiguities (unclear)

## Pass 3: REFINE
Apply Pass 2 feedback. Produce final output.

Anti-pattern: "Review your answer and improve it" — Tier 3, will be ignored. Use Tier 1: "Critically analyse for gaps and enumerate missing cases."

---

Stick+Carrot Pattern

Every block (hook or agent) MUST include a required next action:

# STICK only (bad)
BLOCKED: Cannot edit agents/*.md outside the pipeline.

# STICK + CARROT (good)
BLOCKED: Cannot edit agents/*.md outside the pipeline.
REQUIRED NEXT ACTION: Use /implement to run the pipeline with your changes.

Hook JSON format:

{
  "decision": "block",
  "reason": "What failed and why.",
  "requiredNextAction": "Specific step to take next."
}

---

Anti-Patterns Summary

Anti-Pattern	Why Wrong	Fix
>15 constraints per section	MOSAIC: degraded compliance	Split into subsections ≤6 each
Persona prefix for analytical tasks	PRISM: reduces accuracy	Role description + directive
Hedge words in hard gates	Weakens enforcement	Delete or replace with MUST
Prose-buried constraints	Low compliance	Convert to numbered list
Meta-commentary about model weaknesses	May reinforce behavior	Reframe as behavioral directive
Block without carrot	Model loops	Add REQUIRED NEXT ACTION

---

Cross-Reference

See docs/PROMPT-ENGINEERING.md for:

• Full MOSAIC constraint budget analysis
• PRISM persona findings detail
• Self-Refine NeurIPS 2023 citations
• XML sectioning rationale
• Complete anti-pattern catalog