Thinking Lenses

Thinking Lenses

Optimize a task prompt without turning prompt polish into a mandatory heavy workflow.

Quick Start

• Skip this skill if the request is already simple, clear, and low-stakes.
• Use inline-lite for short one-shot tasks that only need a little more structure.
• Use inline-structured for recurring work, medium complexity, or handoff-friendly prompts.
• Use pre-optimizer only for high-risk, high-cost, long-chain, or ambiguity-heavy tasks.
• Read prompt-templates.md when choosing a lens pair or assembling the final prompt.

Inputs

• task
  • the raw prompt, request, or task description
• mode
  • one of: clarify, design, decide, teach, verify
• constraints
  • hard limits, style limits, banned moves, budget, timing, or safety boundaries
• output_contract
  • the required deliverable shape, sections, format, or acceptance criteria
• rigor
  • light by default, strict when ambiguity or failure cost is high
• delivery_mode
  • optional override: inline-lite, inline-structured, or pre-optimizer

If mode is missing, infer the closest fit. If constraints or output_contract are missing, infer only the minimum needed and name the biggest assumption.

Workflow

1. Restate the real task in one or two sentences.
2. Capture the goal, audience, constraints, and what a good output must do.
3. Route the request into the lightest delivery mode that still protects quality.
4. Choose exactly one primary lens and one supporting lens.
5. Produce a compact optimized prompt rather than a long explanation about frameworks.
6. Add ambiguity and risk notes only when rigor = strict or delivery_mode = pre-optimizer.

Default output sections:

1. Task restatement
2. Method choice
3. Optimized prompt

Add these only for stricter runs:

4. Ambiguities
5. Risks
6. Edit rationale

Guardrails

• Use exactly one primary lens and one supporting lens.
• Prefer the lowest-friction delivery mode unless escalation is justified.
• Preserve user intent and wording when nuance matters.
• Do not expand scope while "improving" the prompt.
• Make the output contract verifiable when another agent or model will execute the result.
• Keep the optimized prompt shorter than the explanation about it.
• Avoid forcing a second agent when the task is simple enough to handle inline.

Failure Modes

• Over-framework a trivial request.
• Stack many methods into one bloated prompt.
• Explain the methods instead of producing an operational prompt.
• Quietly change the scope while polishing the task.
• Default to a separate optimizer pass even when the task is low-stakes.
• Omit evaluation criteria, making the prompt sound smart but hard to verify.

Reference Use

• Read prompt-templates.md when:
  • mode is ambiguous
  • choosing the primary/supporting pair
  • converting the result into inline-lite, inline-structured, or pre-optimizer
  • tightening the output contract for another agent