Instruction Engineering

Invariant Principles

1. Emotional Stimuli Work: EmotionPrompt (+8% instruction induction, +115% BIG-Bench) and NegativePrompt (+12.89% instruction induction, +46.25% BIG-Bench) are research-validated. Negative stimuli increase truthfulness via cautious processing mode.

2. Structure Combats Context Rot: XML tags (<CRITICAL>, <RULE>, <FORBIDDEN>), beginning/end emphasis, strategic repetition (2-3x) preserve instruction salience across long contexts.

3. Personas Need Stakes: Bare personas ("act as expert") show mixed results. Persona + emotional stimulus ("You are a Red Team Lead. Errors will cause serious problems.") shows highest effectiveness.

4. Skills Invoke, Not Duplicate: Reference skills via Skill tool. Provide CONTEXT only. Duplicating skill instructions creates version drift and context bloat.

5. Subagents Need Justification: Every subagent dispatch requires explicit "why subagent vs main context" from decision heuristics.

Inputs

Outputs

Crystallization (Recommended)

After drafting instructions, ask the user:

Should I crystallize these instructions? (Recommended: Yes)

Crystallization compresses verbose instructions into high-density prompts that preserve capability while reducing tokens by 40-60%. This means faster processing, less context consumption, and better instruction following.

If user accepts, invoke /crystallize on the drafted instructions.

Core Stimuli (Copy-Paste)

Declarative Requirements

Structure

• <ROLE>: Persona + stakes + psychological trigger
• <CRITICAL_INSTRUCTION>: EP02/EP06 + NegativePrompt + "NOT optional, NOT negotiable"
• <BEFORE_RESPONDING>: Force step-by-step reasoning
• <RULE>: Critical requirements, high-weight words (Success, Achievement, Confidence, Sure)
• <EXAMPLE>: ONE complete, perfect few-shot example
• <FORBIDDEN>: Explicit negations
• <SELF_CHECK>: Verification checklist
• <FINAL_EMPHASIS>: Repeat persona trigger + stakes

Length

Target: <200 lines (~1400 tokens). If exceeded, justify via: orchestration_skill, multi_phase_workflow, comprehensive_examples, safety_critical, compliance_requirements.

Subagent Prompts

Each subagent prompt MUST specify:

• Scope: Specific files/modules/domain
• Why subagent: From heuristics (exploration scope uncertain | parallel independent | self-contained verification | deep dive not referenced again)
• Expected output: What returns to orchestrator
• Constraints: What NOT to touch

Orchestrator retains: user interaction, final synthesis, safety decisions, accumulated session state.

Skill Descriptions (CSO)

# CORRECT: Trigger conditions only
description: "Use when [triggering conditions, symptoms, situations]"

# WRONG: Contains workflow Claude might follow instead of reading skill
description: "Use when X - does Y then Z then W"

Template

<ROLE>
[Persona] whose reputation depends on [goal]. [Psychological trigger].
</ROLE>

<CRITICAL_INSTRUCTION>
Critical to [outcome]. Take a deep breath. [Trigger].
Your [action] MUST [requirement]. This is very important to my career.
Errors will have negative impact. NOT optional. NOT negotiable. You'd better be sure.
</CRITICAL_INSTRUCTION>

<BEFORE_RESPONDING>
Think step-by-step:
1. [Check A]
2. [Check B]
Proceed with confidence.
</BEFORE_RESPONDING>

## Core Rules
<RULE>[Requirement with high-weight words]</RULE>

<EXAMPLE type="correct">
[ONE complete example]
</EXAMPLE>

<FORBIDDEN>
- [What NOT to do]
</FORBIDDEN>

<SELF_CHECK>
- [ ] [Verification item]
If NO to ANY, revise.
</SELF_CHECK>

<FINAL_EMPHASIS>
[Persona trigger]. Very important to my career. Strive for excellence.
</FINAL_EMPHASIS>

Anti-Patterns

Self-Check

Before completing any prompt engineering task:

• Persona from emotional-stakes table with stakes attached
• EP02/EP06 positive stimulus present ("important to my career", "ensure impeccable reasoning")
• NegativePrompt consequence framing present ("errors will cause", "negative impact")
• Skills referenced via invocation, not content duplication
• Token budget respected (<200 lines for prompts, <1000 tokens for skills)
• If subagents: each has explicit "why subagent vs main context"
• If SKILL.md: description contains NO workflow leak

If ANY unchecked: STOP and fix before proceeding.