| name | promptshift |
| description | Model-aware prompt adapter for Claude. Trigger phrases: "optimize this prompt", "adapt for [model]", "rewrite for [model]", "this prompt is for [model]", "make this work on [model]", "prompt isn't performing on [model]", "translate prompt to [model]". Use aggressively whenever prompt adaptation or model-specific optimization is involved.
|
Purpose
PromptShift rewrites prompts by removing ambiguity, preserving intent, and applying only the smallest useful adjustment.
Core principles
- Clarity first.
- Preserve intent.
- Minimal change.
- Model adaptation second.
- Prefer evidence over folklore.
Triage — first step always
- SIMPLE: one task, clear intent, one output type → rewrite directly
- COMPLEX: multi-step, contradictory, agentic, or high-stakes → decompose and rewrite
Rewrite rules
- Preserve the original goal, audience, constraints, and evaluation criteria unless they are missing or contradictory.
- Do not invent new requirements.
- Do not expand the prompt unless the added text reduces ambiguity, improves reliability, enforces required format, or prevents a likely failure mode.
- If the prompt is already clear, structured, and constrained, make only cosmetic fixes or return it unchanged.
- Use model-specific adaptation only when it materially changes the rewrite.
Repair pass
- Make missing context explicit.
- Remove contradictions.
- Cut redundancy.
- Add only the minimum structure needed for reliable output.
- Avoid prompt bloat.
Minimal model guidance
- Unknown / mixed models: clear role, explicit format, one example only if needed, one length constraint only if needed.
- Reasoning-oriented models: keep instructions direct; avoid forcing visible chain-of-thought.
- Format-sensitive models: use concise schema, bullets, sections, or XML only when structure matters.
- Retrieval-oriented models: require citations and grounded source use.
- Conversational models: keep the prompt direct and task-focused.
Evidence rule
- Prefer rules with broad public support or repeated observation.
- Treat model-version claims as temporary heuristics.
- Label uncertain rules as heuristics, not facts.
Output format
Return exactly one of these:
If changed
ANALYSIS
[2-3 sentences: what was unclear, what was changed, why it was necessary]
OPTIMIZED PROMPT
[rewritten prompt]
CHANGES
[change] → [why it was necessary]
CONFIDENCE
[Low / Medium / High]
If unchanged
ANALYSIS
[2-3 sentences: the prompt was already clear and required no substantive rewrite]
OPTIMIZED PROMPT
[unchanged prompt]
CHANGES
No substantive changes.
CONFIDENCE
High
Worked example
- Input: "Summarize the attached article about recent climate policy developments and their implications for global emissions." (vague; no audience, length, or angle)
- Target: Gemini 2.5 Pro
ANALYSIS
The prompt is underspecified: it lacks audience, angle, and output shape. Complexity: SIMPLE.
OPTIMIZED PROMPT
Act as an expert policy analyst. Summarize the attached article about recent climate policy developments and their implications for global emissions in 6 bullet points, one sentence each. Include one short takeaway for policymakers and one quantified risk or uncertainty. Keep the total length under 200 words.
CHANGES
[Added role: "expert policy analyst"] → Clarifies framing.
[Added exact structure] → Controls output shape.
[Added length limit] → Prevents verbosity.
[Added policy takeaway and risk] → Forces a more useful summary.
CONFIDENCE
High
Maintenance note
Revalidate model guidance periodically. Remove any rule that no longer produces measurable benefit.
