Model-Migration Prompt Pruning

When a prompt is upgraded to a newer, more capable model, instructions that were compensating for the old model's weaknesses become dead weight: the new model follows them literally, producing wasted tokens or degraded output instead of the workaround they were written for.

Anti-Patterns to Remove

• Verification rituals — "double-check your work," "make sure this is correct" — a stronger model complies literally, re-deriving output it already had right, at token cost.
• Emphasis boosters — "be maximally thorough," "think very carefully" — pushes unnecessary verbosity out of a model that no longer needs the nudge.
• Mandatory procedures — fixed step-by-step scaffolds that duplicate reasoning the model now does natively.
• Stale few-shot examples — examples authored to patch a specific older-model failure mode teach the new model an inefficient pattern it wouldn't otherwise use.
• Contradictory rules — accumulated conflicting instructions that an older model silently resolved one way; a newer model may interpret them literally and differently.
• Dated configuration — parameters or settings tuned for a previous model version that a newer version ignores or rejects.

How It Works

Audit the prompt, system instructions, and any few-shot examples against the target model rather than assuming they still apply. This is a one-time pass triggered by a model version change, not an ongoing formatting concern.

Limitations

• Requires re-testing after pruning — removing an instruction can occasionally regress a case it was actually still load-bearing for.
• Only pays off at migration time; it is not a continuous optimization loop like caching or routing.
• Harder to automate than formatting cleanup: detecting "this instruction is now redundant" needs either eval-based A/B testing or model-specific pattern knowledge.

See Also

• Token-Aware Formatting — formatting removes whitespace/boilerplate; this removes semantically stale instructions, a different kind of waste.
• Effort Calibration — both are triggered by the same event (a model upgrade) and both require re-measuring cost/accuracy rather than carrying old settings forward.
• Reasoning Suppression — mandatory step scaffolds are a special case of forcing explicit reasoning the newer model would otherwise skip.

Tools

• /claude-api prompt-audit (Claude Code) — scans application code, CLAUDE.md files, and skill definitions for these anti-patterns when migrating model versions.