The‐5‐Pillars.md

The 5 Pillars of Token Optimization

These are the five core techniques the skill teaches Claude to apply. Each pillar addresses a different source of token waste.

---

Pillar 1 — Prompt Compression

Problem: Most prompts waste 50–80% of their tokens on filler phrases that add no information.

Solution: Shorthand syntax that compresses tasks to their essential meaning.

Examples:

Verbose	Compressed	Savings
"Please analyze this code and tell me what's wrong"	debug:	85%
"Write a detailed blog post about"	blog-post:	78%
"Give me a step by step guide on how to"	steps:	82%
"Compare X and Y in detail"	compare: X vs Y	70%
"Summarize the key points of"	summarize:	76%

Full pattern library: See Compression Cheatsheet

---

Pillar 2 — Output Format Control

Problem: Claude adds preamble, restates your question, writes filler conclusions, and pads responses — wasting 30–50% of output tokens.

Solution: Explicit format and constraint instructions eliminate all padding.

Universal Anti-Filler Block:

No preamble. No "Great question!". No restating my prompt. No conclusion paragraph.
Start directly with the answer.

Format instruction examples:

Format: bullet points | max: 6 bullets | each bullet max 15 words
Format: numbered steps | action verb starts each step
Format: JSON only | no markdown fences | parseable output only
Format: table | columns: X | Y | Z | no intro or outro

Result: 30–50% shorter responses with the same information density.

---

Pillar 3 — XML Tag Structuring

Problem: Unstructured prose prompts are ambiguous. Claude asks clarifying questions, misinterprets intent, and requires multiple rounds of correction.

Solution: XML tags create semantic clarity that Claude parses precisely.

Basic structure:

<context>What Claude needs to know</context>
<task>What Claude must do</task>
<constraints>What Claude must NOT do</constraints>
<format>How the output should look</format>

Result: Tasks completed correctly on first attempt 90% of the time vs 40% for unstructured prompts. Eliminates 2–3 clarification rounds per session.

---

Pillar 4 — Instruction Density

Problem: Users send 4–6 separate messages for what could be one structured prompt, wasting tokens on conversation overhead and context repetition.

Solution: Bundle multiple related tasks into one dense, structured prompt.

Example:

For [topic]:
1. explain (2 sentences max)
2. one concrete example
3. ELI5 version
4. one-line summary
Deliver all 4 in order. No filler between sections.

Result: 60–80% fewer messages needed for the same output.

---

Pillar 5 — Negative Space Instructions

Problem: Without explicit restrictions, Claude defaults to adding introductions, conclusions, qualifiers, and filler that users never asked for.

Solution: Explicit "do NOT" instructions eliminate default padding instantly.

Standard negative block:

Do NOT:
- Repeat or rephrase my question
- Add an introduction or conclusion paragraph
- Use phrases like "Certainly!", "Great!", "Of course!"
- Ask clarifying questions unless the task is impossible without them
- Add disclaimers unless specifically relevant

Result: 20–35% shorter responses immediately on first application.

---

Combined Effect

Applying all 5 pillars together:

Metric	Unoptimized Session	Optimized Session
Input tokens per message	~200 avg	~45 avg
Output tokens per response	~500 avg	~280 avg
Clarification rounds	2–3 per task	0–1 per task
Messages needed per task	4–6	1–2
Total session cost	Baseline	3–8x cheaper