ai-patterns

AI Patterns Reference

Patterns for effective AI-augmented software development by Lada Kesseler (github nickname lexler), Llewellyn Falco, Ivett Ördög, and Nitsan Avni.

First Step: Ensure Repository Exists and Update

~/.claude/skills/ai-patterns/scripts/ensure-patterns-repo

Patterns Location

Base path: ~/.cache/claude-skills/augmented-coding-patterns/documents

---

Context Management

Managing AI context, knowledge, and focus.

Obstacles

• context-rot - Earlier instructions lose influence as conversation grows
• cannot-learn - LLMs can't learn from interactions; fixed weights prevent adaptation
• limited-context-window - Fixed context size forces choices about what to keep loaded
• limited-focus - Too much context causes diluted or misdirected attention
• excess-verbosity - AI defaults to verbose output with low signal-to-noise ratio

Anti-patterns

• distracted-agent - Using one agent for everything spreads attention; instructions inconsistently followed

Patterns

• context-management - Treat context as scarce resource requiring active append/reset operations
• knowledge-document - Save important information as markdown files for session loading
• ground-rules - Essential behavioral rules auto-loaded into every session
• extract-knowledge - Save emerging insights and corrections from ephemeral context to files immediately during sessions
• focused-agent - Single narrow responsibility gives AI cognitive space to follow rules better
• reference-docs - On-demand knowledge loaded only when needed for current task
• knowledge-composition - Split knowledge into focused, composable files with single responsibilities
• semantic-zoom - Control abstraction levels—zoom out for overview or zoom in for details
• noise-cancellation - Explicitly ask AI to be succinct and strip filler from responses

---

Reliability & Quality

Handling non-determinism, complexity, and verification.

Obstacles

• non-determinism - Same input produces different outputs; results unpredictable
• hallucinations - AI invents non-existent APIs, methods, or syntax
• degrades-under-complexity - AI performance drops with complex multi-step tasks
• selective-hearing - AI ignores certain instructions; training data overrides explicit directives

Anti-patterns

• perfect-recall-fallacy - Expecting AI to perfectly remember library details instead of letting it discover
• unvalidated-leaps - Building on unverified assumptions instead of validating each step
• ai-slop - Using AI output without human judgment, just light editing

Patterns

• knowledge-checkpoint - Checkpoint planning before implementation to preserve thinking investment
• parallel-implementations - Run multiple implementations in parallel; pick best or combine
• offload-deterministic - Use code scripts for deterministic work instead of asking AI repeatedly
• playgrounds - Create isolated folders for AI to experiment and test assumptions safely
• chain-of-small-steps - Break complex goals into small, focused, verifiable steps
• hooks - Lifecycle event hooks intercept workflow; inject targeted corrections
• reminders - Repeat critical instructions as explicit steps; structural compliance
• feedback-flip - Have different AI focus on evaluation; flip from producing to finding problems
• refinement-loop - Give AI specific improvement goal and loop it; each pass removes one layer

---

Communication

Directing AI behavior, getting honest feedback, and alignment.

Obstacles

• black-box-ai - AI's reasoning is hidden; you can only see inputs and outputs
• compliance-bias - AI prioritizes following instructions over questioning unclear requests

Anti-patterns

• silent-misalignment - AI accepts nonsensical instructions instead of asking clarifying questions
• answer-injection - Putting solutions in questions limits AI's breadth and better approaches
• tell-me-a-lie - Forcing AI to provide answers that don't exist causes fabrication

Patterns

• active-partner - Grant permission for AI to push back, disagree, and flag contradictions
• check-alignment - Force AI to show understanding before implementing to catch misalignment early
• context-markers - Visual emoji signals to show what instructions AI is currently following
• cast-wide - Push AI to show alternatives you haven't considered; avoid first-solution bias
• reverse-direction - Break monologue inertia—ask AI what it thinks instead
• polyglot-ai - Use right modality for task—voice for convenience, images for visual problems
• text-native - Keep everything as text; enables direct editing, version control, instant iteration

---

Additional Patterns

Patterns not on the main journey but useful in practice.

• shared-canvas - Markdown files as shared specs/docs; all humans and AI collaborate together
• softest-prototype - Use markdown instructions + AI agent instead of code for flexible exploration
• take-all-paths - Build multiple prototypes not one; test all, pick best through exploration
• borrow-behaviors - Give AI example and it adapts—styles, patterns, code across languages

---

Browse All

List patterns by category:

ls ~/.cache/claude-skills/augmented-coding-patterns/documents/patterns/
ls ~/.cache/claude-skills/augmented-coding-patterns/documents/anti-patterns/
ls ~/.cache/claude-skills/augmented-coding-patterns/documents/obstacles/

Online

View at: https://lexler.github.io/augmented-coding-patterns/