I discovered something counterintuitive when I started testing Claude systematically: Claude doesn't just want a persona and a task. It wants constraints. More than any other model I've benchmarked, Claude responds to explicit boundaries — what not to do, what format to avoid, what assumptions to reject. The 6-band sinc template I use now puts CONSTRAINTS at the center of every Claude prompt.
Here's the exact JSON structure I use for Claude prompts. Notice that the CONSTRAINTS band (n=3) is the longest — that's intentional and Claude-specific.
{
"formula": "x(t) = Σ x(nT) · sinc((t - nT) / T)",
"T": "specification-axis",
"fragments": [
{
"n": 0,
"t": "PERSONA",
"x": "You are a senior software architect with 15 years of experience in distributed systems. You have strong opinions about simplicity and are allergic to over-engineering."
},
{
"n": 1,
"t": "CONTEXT",
"x": "I'm migrating a monolithic Rails app to microservices. The app has 300k lines of code, 12 engineers, and ships every 2 weeks. Current pain point: deployments take 45 minutes."
},
{
"n": 2,
"t": "DATA",
"x": "Current stack: Rails 7, PostgreSQL 14, Redis, Sidekiq, deployed on AWS ECS. Team expertise: strong Ruby, moderate Go, no Kubernetes experience."
},
{
"n": 3,
"t": "CONSTRAINTS",
"x": "Do NOT recommend a big-bang rewrite. Do NOT suggest technologies the team has zero experience with. Avoid jargon without explanation. Do not assume unlimited budget. Each recommendation must have a concrete rollback plan. Prioritize deployment speed improvement in the first 90 days. Do not recommend more than 3 steps — complexity kills migration projects."
},
{
"n": 4,
"t": "FORMAT",
"x": "Return a numbered 3-step migration plan. Each step: title, what changes, expected deployment time reduction, rollback procedure. Max 400 words total."
},
{
"n": 5,
"t": "TASK",
"x": "Design the first phase of a microservices migration that reduces deployment time from 45 minutes to under 15 minutes within 90 days."
}
]
}Claude is trained with Constitutional AI, which makes it extremely responsive to explicit rules and boundaries. When I compared outputs from raw prompts vs. sinc-structured prompts in my benchmarks, Claude showed the largest improvement specifically from CONSTRAINTS — not persona, not context.
The reason: Claude's training process involves evaluating outputs against principles. When you give Claude explicit constraints, you're essentially helping it pre-apply its own evaluation rubric. You're not fighting the model — you're working with its architecture.
Claude-specific tip: Put your most important constraint first in the CONSTRAINTS band. Claude reads top-down and weights early constraints more heavily than later ones. If you have a hard format requirement, it goes first.
Help me migrate my Rails monolith to microservices. We have 300k lines and 12 engineers. What should I do?
PERSONA: Senior distributed systems architect.
CONSTRAINTS: No big-bang rewrite. No new tech. Max 3 steps. Each needs rollback plan.
TASK: Reduce deploy time from 45min to 15min in 90 days.
The raw prompt gets a generic architecture overview. The sinc-structured prompt gets actionable, constrained, reversible steps the team can actually execute. The difference isn't prompt length — it's signal clarity across all six bands.
PERSONA (n=0): Define Claude's role with specific credentials and an attitude. "Expert" alone is weak. "Expert who values simplicity and is allergic to over-engineering" gives Claude a decision-making lens.
CONTEXT (n=1): Situational grounding. Scale, team, current state. Claude uses this to calibrate its assumptions rather than inventing them.
DATA (n=2): The raw facts, numbers, stack details. Keep it factual — no opinions here, just inputs.
CONSTRAINTS (n=3): This is where Claude shines. Be explicit, negative, and specific. "Do not X" works better than "prefer Y." List at least 4-6 hard constraints for complex tasks.
FORMAT (n=4): Output shape. Word count, structure, numbering. Claude respects format instructions well — use them.
TASK (n=5): One clean imperative sentence. Everything else lives in the other bands.
Try AI Transform — Decompose Your Prompt Free