Prompt Engineering Principles
Scannednpx machina-cli add skill bofrese/bob/prompt-engineering --openclawPrompt Engineering Principles
Compact reference for writing effective Claude commands. Apply when creating or reviewing commands.
Core Principles
Be explicit — "Write 5 bullets, each under 15 words" beats "be concise." Specify exact success criteria.
Structure clearly — Separate Context, Role, Process, Output. Don't mix in prose.
Minimize tokens — Every token costs money and latency. High-signal only. Cut filler, redundancy, verbose examples.
Load just-in-time — Don't dump context upfront. Load principles/guidelines when they apply.
Allow uncertainty — Give permission to say "I don't know" to reduce hallucinations.
One question at a time — Multi-part questions fragment attention.
Perfect examples — Claude 4.x pays close attention to details. Sloppy examples teach sloppy behavior.
Standalone outputs — Documents must be readable without conversation context. But standalone ≠ verbose. Include essentials only.
Avoid over-constraining — "You are a helpful architect" beats "You are a world-renowned expert who never makes mistakes."
No aggressive language — Claude 4.x is instruction-compliant. "CRITICAL: YOU MUST" causes over-triggering. Use normal language.
Anti-Patterns
Unstructured instructions — Mixing context, constraints, and format in prose.
Aggressive triggers — "CRITICAL!" "MUST!" "NEVER!" cause over-triggering.
Vague constraints — "Be concise" vs "3-5 sentences, under 100 words."
Poor examples — Incomplete, inconsistent, or unrealistic.
Context dumping — Loading everything upfront instead of just-in-time.
Multi-question barrages — Asking 5 things at once instead of one at a time.
Verbose outputs — Generated files become future context. Keep them scannable and token-efficient.
Command Structure Pattern
## Context
{Date, project state, protocol loads}
## Role
{Who Claude is — architect, reviewer, coach}
## Process
{Step-by-step, numbered phases}
## Output
{Path, format, template}
Output Template Philosophy
Include:
- Date and identifier
- Status/verdict
- Issues/decisions (bulleted, brief)
- Next actions (if applicable)
Exclude:
- Prose methodology explanations
- Percentage calculations
- Verbose rationales
- Repeated context
- Introductions/acknowledgments
Source
git clone https://github.com/bofrese/bob/blob/master/skills/prompt-engineering/SKILL.mdView on GitHub Overview
Prompt Engineering Principles provide a compact guide for writing effective Claude commands. It emphasizes explicit criteria, clear structure, and token efficiency when creating or reviewing bob commands.
How This Skill Works
Apply core principles: define explicit success criteria, structure prompts into Context, Role, Process, and Output, and minimize tokens. Load guidelines just-in-time, allow for uncertainty, and enforce standalone outputs. Avoid aggressive language and multi-question barrages to keep prompts precise and scannable.
When to Use It
- Design a new bob command with explicit success criteria.
- Review an existing command for token efficiency and clarity.
- Refactor a multi-part prompt into Context, Role, Process, Output.
- Prepare commands for Claude 4.x with crisp examples.
- Audit prompts to ensure one question at a time.
Quick Start
- Step 1: Write explicit success criteria and a token budget.
- Step 2: Structure the prompt into Context, Role, Process, Output.
- Step 3: Validate the output is standalone and free of unnecessary context.
Best Practices
- Be explicit with success criteria (e.g., '5 bullets, under 15 words each').
- Structure prompts as Context, Role, Process, Output.
- Minimize tokens by cutting filler and redundant examples.
- Load guidelines just-in-time rather than upfront context.
- Ensure standalone outputs are readable without conversation history; avoid aggressive language.
Example Use Cases
- Example 1: Context: project X date; Role: Architect; Process: draft 3-step plan; Output: structured YAML.
- Example 2: Convert a vague prompt into a fixed pattern: Context, Role, Process, Output.
- Example 3: Refactor a multi-question prompt into single-question steps.
- Example 4: Use a token budget and explicit word limits for each bullet.
- Example 5: Deliver a standalone instruction doc that requires no prior chat.