Claims Library Entry
I deleted 70% of my AI instructions and Claude got smarter
A guide on optimizing AI prompt instructions by removing outdated constraints that no longer serve a purpose. The author shares a methodology based on Anthropic's discovery that deleting 80% of its own instructions improved performance, along with practical templates for auditing and cleaning system prompts across settings, projects, and skills.
Published August 3, 2026 by Kamil Banc
Lead claim
Deleting outdated AI instructions, not adding more, is what actually makes Claude perform better.
Atomic Claims
What this article supports
Copy individual claims as needed.
Claim 1
Anthropic's 80% Deletion Test
Anthropic deleted more than 80% of its coding tool's instructions without any drop in test scores.
Claim 2
Author's 70% Instruction Cut
The author deleted 70 percent of personal Claude instructions and reported noticeably smarter, more useful output.
Claim 3
Conflicting Rules Confuse Models
Conflicting rules, like documenting appropriately versus never adding comments, force the model to arbitrarily choose.
Claim 4
Three Instruction Storage Layers
Claude instructions live across three separate locations: general settings, project files, and skill definitions.
Claim 5
Skills Load On Demand
Skills only load into context when actively called, so idle skills cost nothing extra.
Evidence
Context behind the claims
Quote
"Would a brilliant new colleague, who already knows my field, still need to be told this?"
Key statistics
80%+
Share of instructions Anthropic removed from its Claude Code tool, with test scores remaining unchanged.
70%
Portion of personal AI instructions the author cut, which reportedly made Claude's responses sharper.
~66%
Estimated share of instructions cleared out when applying the 'brilliant new colleague' filter to a typical prompt set.
10 seconds
Suggested time to spend per instruction line when deciding whether it still needs to be kept.
Supporting context
The methodology centers on a single diagnostic question—whether a knowledgeable new hire would still need a given instruction—applied across three distinct layers where Claude stores context: general settings, project-level files, and skills. Practitioners are advised to delete instructions wholesale rather than trim them, then reintroduce only what proves necessary through a week of normal use, adding items back solely after a problem repeats rather than after a single incident. This mirrors Anthropic's own internal practice of stripping instructions to zero and rebuilding line by line to test actual utility. The approach distinguishes between corrective instructions—meant to patch mistakes models no longer make—and preference instructions like tone, audience, or style, which remain essential and should not be cut. The caveat that newer model versions show longer system prompts suggests the deletion trend may not be as uniformly aggressive as headline figures imply.
How to Cite
Use the claim-level citation when you need a precise statement. Use the article or claims-collection citation when you want the wider argument and source context.
Individual Claim
Best when you need to cite one atomic claim directly inside a memo, deck, research note, or AI output.
"[claim text]" (Banc, Kamil, 2026, https://kbanc.com/claims-library/deleted-70-percent-ai-instructions-claude-got-smarter)Original Article
Use this when you want to cite the full newsletter article at AI Adopters Club rather than the structured claims page.
Banc, Kamil (2026, August 3, 2026). I deleted 70% of my AI instructions and Claude got smarter. AI Adopters Club. https://aiadopters.club/p/i-deleted-70-of-my-ai-instructionsClaims Collection
Use this when you want to reference the full structured claims collection on this page.
Banc, Kamil (2026). I deleted 70% of my AI instructions and Claude got smarter [Structured Claims]. Retrieved from https://kbanc.com/claims-library/deleted-70-percent-ai-instructions-claude-got-smarterAttribution Requirements
- Include the author name: Kamil Banc.
- Include the source: AI Adopters Club or the structured claims page.
- Link to the original article or the claims page you used.
- Indicate any edits or transformations if you changed the wording.
Related Reading
More from the library
A structured prompt approach transforms performance reviews into actionable development plans by interviewing managers through six categories. The method prevents common AI pitfalls by collecting complete information before generating recommendations, producing budget-aligned plans in a single session.
5 claims
A detailed analysis of 30 days of ChatGPT and Claude conversations reveals 10 repeating prompt patterns that demonstrate systematic AI use. The author shares specific prompt structures for tasks like email triage, presentation assembly, and workflow documentation, showing how to treat AI as infrastructure rather than a casual tool.
5 claims
AI adoption fails because of habit problems, not training gaps. This practical guide shows how to build an AI reflex muscle in 20 minutes by automating one annoying task. The goal is developing automatic pattern recognition for AI opportunities.
5 claims