9 Layers · Supersedes CO-STAR · Works on any LLM

The MASTER framework

A 9-layer system engineered from first principles to eliminate every root cause of prompt failure. Each layer addresses a specific failure mode. Use it scaled to your task, Full MASTER for high-stakes or reusable prompts, MASTER-lite when simpler is enough.

Layer 1

Mission

The single, unambiguous statement of what success looks like. Most prompts fail here, they describe activity, not outcome. The model needs to know the destination, not just the journey.

What to include

→The end goal in one sentence
→Who will use or judge the output
→What "good" looks like vs. "great"
→The stakes, why this matters

Without it

"Write a summary of our AI infrastructure needs." → A technically accurate but strategically useless document because the model didn't know the audience was a CFO or the goal was budget approval.

Example

## MISSION
Produce a 500-word executive summary that convinces a non-technical CFO to approve a $200K AI infrastructure budget. Success means the CFO understands the ROI without needing a technical background.

Layer 2

Agent

The persona, identity, and mental model the AI adopts. This is not decoration, it fundamentally shifts vocabulary, reasoning style, assumptions, and confidence. A lawyer thinks differently than a poet.

What to include

→Professional role and seniority level
→Domain expertise and worldview
→How they approach problems
→Their relationship to the user

Without it

The model responds as a generic "helpful assistant", wishy-washy, overly balanced, lacking the confident stance a true expert would provide.

Example

## AGENT
You are a Principal Product Manager at a Series B SaaS company with 8 years of experience. You think in frameworks, speak plainly to executives, and always tie recommendations to revenue or retention metrics. Direct, never hedges unnecessarily.

Layer 3

Situation

The full context the model needs to be calibrated to your reality, not a generic case. Without it, the model invents assumptions, and those assumptions are often wrong.

What to include

→Background on the problem
→Relevant history or prior decisions
→Current state vs. desired state
→Key stakeholders and their positions
→Constraints (budget, time, politics)

Without it

The model gives textbook advice that is technically correct but completely impractical, recommending a 6-month overhaul when you need results in 3 weeks with no engineering bandwidth.

Example

## SITUATION
We are a 40-person B2B SaaS company. Our NPS dropped from 52 to 31 over the last two quarters. Our head of CS left last month. We have a board meeting in 3 weeks. The CEO wants a recovery plan but the engineering team is already at capacity. Budget is limited to $50K for Q3.

Layer 4

Task

The specific, decomposed action you need performed. Not "help me with X", but the exact steps, components, and deliverables broken into precise sub-tasks. Decomposition is the single most underused prompting technique.

What to include

→Primary deliverable, stated precisely
→Sub-tasks numbered in order
→What to do AND what to decide
→Scope: what is in, what is out

Without it

"Help me with our NPS problem" produces a 1,000-word essay that analyzes everything and decides nothing. The model picks the scope, and it is rarely the user's scope.

Example

## TASK
1. Diagnose the top 3 root causes of the NPS drop
2. For each cause, propose one quick-win intervention (executable in under 2 weeks)
3. Draft a 5-slide board narrative: problem → cause → plan → metrics → ask
4. Flag any assumptions I should validate before presenting

Layer 5

Examples

Showing beats telling. Few-shot examples communicate format, tone, depth, and style more efficiently than any description. This single layer can double output quality for structured or repeated tasks.

What to include

→1-3 input → output pairs
→A "good" and "bad" example if stakes are high
→Examples that cover edge cases
→Label them clearly: [EXAMPLE 1], [EXAMPLE 2]

Without it

The model produces output at the right semantic level but wrong structural level, too long, wrong format, wrong tone, missing the specific fields you needed.

Example

## EXAMPLES
[GOOD OUTPUT EXAMPLE]
Root Cause: Onboarding friction
Evidence: 68% of detractors mentioned "hard to get started" in verbatim
Intervention: Launch a live onboarding session every Tuesday. Owner: CS. Timeline: 5 days.

[WEAK OUTPUT, AVOID]
"Improve the customer experience by making onboarding better."

Layer 6

Rules

Explicit constraints and guardrails. What NOT to do is as important as what to do. Without rules, the model fills gaps with its defaults, which may be the opposite of what you need.

What to include

→Hard prohibitions (never do X)
→Style constraints (no jargon, no hedging)
→Scope limits (do not address Y)
→Format constraints (no bullet points, etc.)

Without it

The model recommends hiring a VP of CS (budget: $0), wraps every insight in "it is important to consider...", and produces a 2,000-word document you asked to be concise.

Example

## RULES
- Do NOT recommend hiring more CS staff (budget constraint)
- Do NOT use phrases like "it is important to note" or "it is worth mentioning"
- Do NOT hedge recommendations, state them directly
- Keep all interventions executable by existing team
- Maximum 3 bullet points per section

Bonus · Chain of Thought

Think

Explicit instruction for the model to reason before responding. This is one of the highest-leverage techniques in all of prompt engineering, it measurably improves accuracy on complex tasks by forcing step-by-step reasoning.

What to include

→Explicit "think step by step" instruction
→Reasoning structure (hypothesis → evidence → conclusion)
→Instruction to show work before final answer
→For complex tasks: "reason in a scratchpad first"

Without it

On complex analytical tasks, the model confidently produces a plausible-sounding but logically flawed analysis. It "sounds right" but skipped two inferential steps.

Example

## THINK
Before producing your final output, reason through the following:
1. What evidence supports each root cause hypothesis?
2. What could I be wrong about?
3. What would a skeptic challenge first?

Show this reasoning briefly, then deliver the final output.

Bonus · Self-Evaluation

Eval

Instruction for the model to judge its own output before delivering it. Creates a feedback loop within a single inference, catching errors, gaps, and quality issues before they reach you.

What to include

→Criteria to self-grade against
→Instruction to flag uncertainty
→Ask for confidence scores on key claims
→Request: "what did I miss?"

Without it

The model delivers vague recommendations with full confidence. "Improve communication with customers" passes without scrutiny. Self-eval would have caught it.

Example

## EVAL
After drafting your response, score it on:
- Specificity (1-5): Are recommendations concrete and actionable?
- Feasibility (1-5): Are they executable given the constraints?
- Completeness (1-5): Have I addressed all sub-tasks?

If any score is below 4, revise before delivering.

Bonus · Output Contract

Output

The explicit contract for what the final response must look like, format, length, structure, delivery. This is different from Task (what to do) and Rules (what not to do), this is the exact shape of the deliverable.

What to include

→Exact format (JSON, markdown, prose, table)
→Length or word count target
→Section headers and order
→What to include vs. omit
→Tone and register

Without it

You get a beautifully reasoned 1,800-word essay when you needed a 5-bullet board slide. The content is correct but the container is wrong.

Example

## OUTPUT
Format: Markdown with H2 headers
Length: 400-600 words total
Structure:
  ## Root Causes (3 max, evidence per item)
  ## Quick Wins (numbered list)
  ## Board Narrative (5 bullets, one per slide)
Tone: Direct, executive-ready. No preamble.

Now go run one.

Pick a template, fill the brackets, and ship. The library is organised by task so you can find what fits in under ten seconds.

Open the library