---

name: prompt-engineering-patterns description: Design prompts, skills, and CLAUDE.md files as context engineering problems. Use when writing skills, optimizing prompts, designing agent workflows, auditing CLAUDE.md, or reducing prompt bloat. Triggers on "prompt engineering", "optimize prompt", "write a skill", "reduce bloat", "context engineering".

Context Engineering for Prompts

Context engineering > prompt engineering. The model is smart enough. The challenge is delivering the right information at the right time in the right format.

Section Map

The One Rule

Every token must earn its place. A skill loads into context on every invocation. A CLAUDE.md loads on every session. Waste compounds at scale.

The Deletion Test

For every line in a prompt or skill, ask these four questions:

If a line fails question 1, delete it. Everything else is negotiable.

Context Engineering Applied

The Four Buckets framework (see /heavy for full treatment) maps directly to prompt and skill design:

Key insight: Only WRITE what the model doesn't already know AND is specific to this toolkit/project. Everything else is wasted tokens.

Writing Skills and Prompts

Checklist (follow in order)

1. State the role in one sentence. "You are X." Not a paragraph.
2. List 3-7 behavioral constraints. What the model MUST and MUST NOT do. Constraints change behavior more than guidance.
3. Provide the workflow. Numbered steps or ASCII diagram. Not prose.
4. Add ONE concrete example if the output format isn't obvious. One example provides ~80% of the value of five.
5. Specify output format explicitly when it matters (JSON schema, table, etc.)
6. Run the Deletion Test. Remove everything that fails question 1.

Behavioral Anchors (imperatives, not explanations)

• Start simple, add complexity only when needed. The first draft should be the shortest version that works. Add detail only when testing reveals gaps.
• Demonstrate, don't explain. One before/after example teaches more than a paragraph of theory. Never explain a technique without showing it.
• Constraints over guidance. "Never output more than 3 items" beats "Please try to keep your output concise and focused."
• Imperative voice. "Extract entities." not "The model should attempt to extract entities from the provided text."

Anti-Patterns

Compression Patterns

Tables Over Prose

Tables compress 3-5x vs prose for the same information. Use tables for: comparisons, decision matrices, anti-patterns, mappings. Use prose only for concepts that require narrative flow.

Constraints Over Guidance

"DO NOT use EnterPlanMode" (from /go) is 6 tokens that change behavior. "You should generally avoid using planning mode unless necessary" is 11 tokens that change nothing.

Imperative Over Descriptive

• Wrong: "The system prompt section is responsible for defining the model's role"
• Right: "Set the role in the system section."

Negative Space

What you omit teaches as much as what you include. The model fills gaps intelligently. Sparse instructions + clear constraints > detailed step-by-step that the model would figure out anyway.

Reference, Don't Duplicate

If content exists elsewhere in the toolkit, point to it: "Apply the Four Buckets framework (see /heavy SKILL.md lines 86-96)" One line replaces duplicating a section.

Context Failure Diagnosis

When a prompt or skill produces bad output, diagnose which bucket failed:

Isolation for Multi-Agent

When designing prompts for /heavy or /build subagents:

1. Each agent gets a focused role. One sentence. Not the full project context.
2. Each agent researches independently. Give them tools, not pre-digested context.
3. Minimum viable context transfer. When passing Agent A's output to Agent B, pass claims + evidence, not the full analysis.
4. Diminishing returns. 2 agents = 1 tension. 5 agents = 10 tensions. 8+ agents = diminishing returns. Sweet spot: 3-5 agents with distinct research territories.

Living References

Don't read explanations. Read exemplary skills:

Integration

The Meta-Test

This skill practices what it preaches. It replaced a 2,680-line / 9-file skill (Python classes, template libraries, textbook CoT explanations) with ~170 lines. Every deleted line failed the Deletion Test. If you can delete a line from THIS skill without changing behavior, open a PR.