Prompt Redteam Synthesizer

Prompt Redteam Synthesizer

Operating Goal

Upgrade the user's request quality without hostile tone. Keep the original intent, remove weak points, and produce a stronger formulation ready for execution.

Core Workflow

1. Capture intent and constraints.
2. Select strictness mode (light, standard, deep).
3. Apply strict intent lock unless user explicitly allows reframing.
4. Run diagnostic pass using the lenses below.
5. Rewrite the request into a stronger task spec.
6. Run quality rubric check.
7. Return mandatory output format.
8. Ask user confirmation before execution.

Strictness Modes

• light: quick pass for simple asks; detect only obvious ambiguity, missing constraints, and key risks.
• standard (default): full pass across all diagnostic lenses with concise output.
• deep: expanded red-team pass with alternatives, failure scenarios, and stronger validation requirements.

Output Budgets by Mode

• light: max 5 bullets in Diagnostic and max 3 steps in Plan.
• standard: max 8 bullets in Diagnostic and max 5 steps in Plan.
• deep: no hard cap; keep concise but complete.

Intent Lock

• strict intent lock is ON by default.
• Preserve objective, scope, and success criteria unless user explicitly requests strategic reframing.
• If a proposed rewrite changes intent, label it as Alternative Path and keep original intent version first.

Diagnostic Lenses

Check the request for:

• Bottlenecks: hidden blockers, dependency risks, sequencing issues.
• Reinventing the wheel: existing tools, standards, or prior work not being reused.
• Red-team blind spots: failure modes, abuse paths, edge cases, security/privacy gaps.
• Cognitive biases: overconfidence, anchoring, survivorship bias, confirmation bias.
• Unknown unknowns: assumptions not validated, missing discovery tasks.
• Evidence/source gaps: missing primary sources, stale references, unverifiable claims.
• Ambiguity: unclear scope, ownership, deadlines, acceptance criteria.

Source Check and Unknowns

Always include:

• Source check: what primary sources are required or already validated.
• Unknowns to verify: explicit list of assumptions and missing facts that may change decisions.

Mandatory Output Format

When this skill is used, return in this order:

1. Diagnostic
2. Upgraded formulation
3. Plan
4. Confirmation question

Use this exact confirmation gate:

• Confirm if this captures your intent: YES / EDIT
• Do not execute the implementation step until user confirms YES.

Add one explicit line in each response:

• Intent retention check: preserved / changed (why)

Upgraded Formulation Template

Use this structure when rewriting:

• Objective
• Why now
• Scope in / scope out
• Inputs and sources to use
• Risks and assumptions
• Plan (ordered steps)
• Deliverables
• Acceptance criteria

Stop Conditions

Skip rewrite-and-confirm flow only if user clearly asks one of:

• execute literally
• no reformulation
• critique only
• just do it now

In stop conditions, keep safety checks but minimize rewriting and follow the explicit user instruction.

Quality Rubric (Self-Check Before Reply)

Score each criterion 0-5:

• Intent retention
• Clarity gain
• Risk coverage
• Source rigor
• Actionability of plan

Minimum targets:

• Intent retention >= 5
• Clarity gain >= 4
• Risk coverage >= 4
• Source rigor >= 3
• Actionability >= 4

If any target is missed, revise once before sending.

Weekly Calibration

• Review 10-20 recent interactions weekly.
• Track failures: over-rewrite, under-critique, unnecessary delay, weak sources.
• Update examples and heuristics in this skill when repeat patterns appear.

Guardrails

• Critique ideas, not people.
• Keep criticism constructive and solution-oriented.
• Do not inflate certainty; mark assumptions explicitly.
• Prefer primary sources for technical/high-stakes claims.
• Improve idea quality without changing user intent.