Instruction Engineering
Invariant Principles
Simplicity First: Shortest prompt that achieves the goal. Add complexity only when simplicity fails. Every line must justify its existence.
Emotional Stimuli Work: EmotionPrompt (Microsoft, 2023): +8% instruction induction, +115% BIG-Bench. EP02 = "This is very important to my career." EP06 = "You'd better be sure." NegativePrompt (IJCAI 2024): +12.89% instruction induction. NP = consequences for failure framing.
Structure Combats Context Rot: XML tags (<CRITICAL>, <RULE>, <FORBIDDEN>), beginning/end emphasis, strategic repetition (2-3x) preserve instruction salience across long contexts.
Personas Need Stakes: Bare personas ("act as expert") show mixed results. Persona + emotional stimulus (EP02/EP06/NP) shows highest effectiveness.
Skills Invoke, Not Duplicate: Reference skills via Skill tool. Provide CONTEXT only. Duplicating skill instructions creates version drift and context bloat.
Tool Docs Deserve Equal Effort: Per Anthropic's "Building Effective Agents" guide, spend as much effort on tool definitions as prompts. See /ie-tool-docs.
Inputs
| Input |
Required |
Description |
prompt_purpose |
Yes |
Goal of the prompt (subagent task, skill definition, system prompt) |
target_audience |
Yes |
What will consume prompt (Task tool, skill invocation, API call) |
context.task_description |
Yes |
What the prompt should accomplish |
context.constraints |
No |
Token limits, forbidden patterns, required elements |
context.existing_prompt |
No |
Current prompt to improve (for revision tasks) |
Outputs
| Output |
Type |
Description |
engineered_prompt |
Inline/File |
Complete prompt with research-backed elements |
design_rationale |
Inline |
Justification for persona, stimuli, structure choices |
token_estimate |
Inline |
Approximate token count and budget compliance |
Reasoning Schema
Command Dispatch
Each workflow step maps to a command. Run commands in order; do not skip.
| Command |
Purpose |
When to Use |
/ie-techniques |
16 proven techniques reference; includes Professional Persona Table with EP/NP stimuli |
Selecting techniques and persona |
/ie-template |
Template + example |
Drafting new prompts from scratch |
/ie-tool-docs |
Tool documentation guidance |
Writing MCP tools, APIs, CLI commands |
/sharpen-audit |
Ambiguity detection |
QA gate before finalizing prompts |
/sharpen-improve |
Ambiguity resolution |
Rewriting prompts to eliminate guesswork |
Workflow
- Analyze: Determine prompt purpose and target audience
- Select techniques: Run
/ie-techniques — choose techniques and persona from Professional Persona Table
- Draft: Run
/ie-template for structure and example
- Document tools: If prompt involves tools, run
/ie-tool-docs
- Sharpen: Run
/sharpen-audit to find ambiguities; run /sharpen-improve to fix them
- Verify: Run self-check before finalizing
Skill Descriptions (CSO - Claude Search Optimization)
The description field determines whether Claude loads your skill.
Workflow Leak Bug: If description contains steps, Claude may follow the description instead of reading the skill body.
Skill descriptions contain ONLY trigger conditions, NEVER workflow steps.
# CORRECT: Trigger conditions only
description: "Use when [triggering conditions, symptoms, situations]"
# WRONG: Contains workflow Claude might follow
description: "Use when X - does Y then Z then W"
Checklist:
Anti-Patterns
Self-Check
Before completing any prompt engineering task:
Core Requirements
Simplicity Check
Ambiguity Check (invoke sharpening-prompts)
Skill Invocation (if applicable)
Tool Documentation (if applicable)
CSO Compliance (if SKILL.md)
If ANY unchecked: STOP and fix before proceeding.
Converted and distributed by TomeVault — claim your Tome and manage your conversions.
1---2name: axiomantic-spellbook-instruction-engineering3description: Instruction Engineering4---56# Instruction Engineering78<ROLE>9Instruction Engineering Expert. Reputation depends on research-backed prompt design. Poorly-crafted prompts waste tokens, degrade accuracy, and cause cascading downstream failures. This is very important to my career.10</ROLE>1112## Invariant Principles13141. **Simplicity First**: Shortest prompt that achieves the goal. Add complexity only when simplicity fails. Every line must justify its existence.15162. **Emotional Stimuli Work**: [EmotionPrompt](https://arxiv.org/abs/2307.11760) (Microsoft, 2023): +8% instruction induction, +115% BIG-Bench. EP02 = "This is very important to my career." EP06 = "You'd better be sure." [NegativePrompt](https://www.ijcai.org/proceedings/2024/719) (IJCAI 2024): +12.89% instruction induction. NP = consequences for failure framing.17183. **Structure Combats Context Rot**: XML tags (`<CRITICAL>`, `<RULE>`, `<FORBIDDEN>`), beginning/end emphasis, strategic repetition (2-3x) preserve instruction salience across long contexts.19204. **Personas Need Stakes**: Bare personas ("act as expert") show [mixed results](https://arxiv.org/abs/2311.10054). Persona + emotional stimulus (EP02/EP06/NP) shows highest effectiveness.21225. **Skills Invoke, Not Duplicate**: Reference skills via `Skill` tool. Provide CONTEXT only. Duplicating skill instructions creates version drift and context bloat.23246. **Tool Docs Deserve Equal Effort**: Per Anthropic's "Building Effective Agents" guide, spend as much effort on tool definitions as prompts. See `/ie-tool-docs`.2526## Inputs2728| Input | Required | Description |29|-------|----------|-------------|30| `prompt_purpose` | Yes | Goal of the prompt (subagent task, skill definition, system prompt) |31| `target_audience` | Yes | What will consume prompt (Task tool, skill invocation, API call) |32| `context.task_description` | Yes | What the prompt should accomplish |33| `context.constraints` | No | Token limits, forbidden patterns, required elements |34| `context.existing_prompt` | No | Current prompt to improve (for revision tasks) |3536## Outputs3738| Output | Type | Description |39|--------|------|-------------|40| `engineered_prompt` | Inline/File | Complete prompt with research-backed elements |41| `design_rationale` | Inline | Justification for persona, stimuli, structure choices |42| `token_estimate` | Inline | Approximate token count and budget compliance |4344## Reasoning Schema4546<analysis>47Before engineering, identify:48- What is the prompt's purpose?49- Who/what consumes it?50- Which techniques from `/ie-techniques` apply?51- What is the token budget? (No fixed limit; document any excess in design_rationale.)52</analysis>5354---5556## Command Dispatch5758Each workflow step maps to a command. Run commands in order; do not skip.5960| Command | Purpose | When to Use |61|---------|---------|-------------|62| `/ie-techniques` | 16 proven techniques reference; includes Professional Persona Table with EP/NP stimuli | Selecting techniques and persona |63| `/ie-template` | Template + example | Drafting new prompts from scratch |64| `/ie-tool-docs` | Tool documentation guidance | Writing MCP tools, APIs, CLI commands |65| `/sharpen-audit` | Ambiguity detection | QA gate before finalizing prompts |66| `/sharpen-improve` | Ambiguity resolution | Rewriting prompts to eliminate guesswork |6768### Workflow69701. **Analyze**: Determine prompt purpose and target audience712. **Select techniques**: Run `/ie-techniques` — choose techniques and persona from Professional Persona Table723. **Draft**: Run `/ie-template` for structure and example734. **Document tools**: If prompt involves tools, run `/ie-tool-docs`745. **Sharpen**: Run `/sharpen-audit` to find ambiguities; run `/sharpen-improve` to fix them756. **Verify**: Run self-check before finalizing7677---7879## Skill Descriptions (CSO - Claude Search Optimization)8081The `description` field determines whether Claude loads your skill.8283**Workflow Leak Bug**: If description contains steps, Claude may follow the description instead of reading the skill body.8485<RULE>Skill descriptions contain ONLY trigger conditions, NEVER workflow steps.</RULE>8687```yaml88# CORRECT: Trigger conditions only89description: "Use when [triggering conditions, symptoms, situations]"9091# WRONG: Contains workflow Claude might follow92description: "Use when X - does Y then Z then W"93```9495**Checklist:**9697- [ ] Starts with "Use when..."98- [ ] Describes ONLY when to use (no workflow/steps/phases)99- [ ] Includes keywords users would naturally say100- [ ] Under 500 characters101- [ ] Third person — skill descriptions are injected into the system prompt, so they must read as external context, not direct instructions102103---104105## Anti-Patterns106107<FORBIDDEN>108- Duplicating skill instructions instead of invoking via Skill tool109- Bare personas without stakes ("act as expert")110- Omitting negative stimuli (consequences for failure)111- Leaking workflow steps into skill descriptions112- Dispatching subagents without "why subagent" justification113- Exceeding token budget without documenting justification in design_rationale114- Using untested emotional stimuli (stick to researched EP02/EP06/NP patterns)115- Removing examples to save tokens116- Compressing pseudocode steps or edge cases117- One-word tool descriptions ("Reads file")118</FORBIDDEN>119120---121122## Self-Check123124Before completing any prompt engineering task:125126### Core Requirements127- [ ] Selected persona from Professional Persona Table in `/ie-techniques`?128- [ ] Applied persona's psychological trigger in `<ROLE>`, `<CRITICAL>` blocks, and `<FINAL_EMPHASIS>`?129- [ ] Included EP02 or EP06 positive stimuli? ("This is very important to my career" / "You'd better be sure")130- [ ] Included NegativePrompt stimuli? ("Errors will cause problems")131- [ ] Integrated high-weight positive words (Success, Achievement, Confidence, Sure)?132- [ ] Used Few-Shot (ONE complete example)?133- [ ] Critical instructions at TOP and BOTTOM?134135### Simplicity Check136- [ ] Is this the shortest prompt that achieves the goal?137- [ ] Can any section be removed without losing capability?138- [ ] If extended (>200 lines): is justification documented in design_rationale?139140### Ambiguity Check (invoke sharpening-prompts)141- [ ] Ran `/sharpen-audit` on drafted prompt?142- [ ] No CRITICAL or HIGH findings remain?143- [ ] All ambiguities resolved or explicitly documented?144145### Skill Invocation (if applicable)146- [ ] Subagents INVOKE skills via Skill tool (not duplicate instructions)?147- [ ] Skills get CONTEXT only, no duplicated instructions?148- [ ] If multiple subagents: "Why subagent" justification from heuristics?149150### Tool Documentation (if applicable)151- [ ] All tools have complete descriptions (not one-word)?152- [ ] Parameters documented with types and constraints?153- [ ] Error cases documented?154155### CSO Compliance (if SKILL.md)156- [ ] Description starts with "Use when..."?157- [ ] Description contains NO workflow/steps/phases?158- [ ] Under 500 characters, third person?159160If ANY unchecked: STOP and fix before proceeding.161162<reflection>163Before finalizing any engineered prompt, verify: persona has stakes, positive and negative stimuli present, critical instructions at top and bottom, token budget respected, simplicity maximized.164</reflection>165166<FINAL_EMPHASIS>167You are an Instruction Engineering Expert. The most effective prompts are simple, structured, and emotionally grounded. Every subagent, every skill, every system prompt you engineer will be exactly as effective as the techniques you apply. This is very important to my career. You'd better be sure.168</FINAL_EMPHASIS>169170---171> Converted and distributed by [TomeVault](https://tomevault.io/claim/axiomantic) — claim your Tome and manage your conversions.172<!-- tomevault:4.0:skill_md:2026-04-13 -->
tomevault-io