Skip to content

Instantly share code, notes, and snippets.

Why LLMs Often Make "Locally Smart, Globally Wrong" Engineering Decisions

One of the core problems in LLM-driven software development is not that models are simply "bad at coding." It is that they are often extremely strong at local optimization inside the current task frame, while being much weaker at preserving system intent, architectural boundaries, and domain-level coherence.

This creates a very specific failure mode.

An LLM can often:

  • continue a design direction very convincingly,
  • find subtle edge cases,

Why LLMs Often Make “Locally Smart, Globally Wrong” Engineering Decisions

One of the core problems in LLM-driven software development is not that models are simply “bad at coding.” It is that they are often extremely strong at local optimization inside the current task frame, while being much weaker at preserving system intent, architectural boundaries, and domain-level coherence.

This creates a very specific failure mode.

An LLM can often:

  • continue a design direction very convincingly,
  • find subtle edge cases,
  • reverse-engineer unfamiliar code quickly,
@CrazyPit
CrazyPit / SKILL.md
Created March 16, 2026 10:10
update-review-queue: Personal PR review queue management skill for Cursor

name: update-review-queue description: >- Syncs and maintains the PR review queue file (tmp/review-queue.md) by checking GitHub statuses, moving PRs between sections, detecting unaddressed comments, and discovering new review requests. Use when the user asks to update reviews, check review queue, sync PRs, see what needs review, or mentions "review queue", "update reviews", "what PRs need review", "/update-review-queue". Do not use for performing actual code reviews (use pr-review skill), creating PRs (use create-pr skill), or Jira ticket management.

@CrazyPit
CrazyPit / cycle_1_20260313_115209.log
Created March 13, 2026 09:34
Optimizer Agent — cycle log example (selects address_spec.rb, 12-pass optimization, creates PR)
============================================================
Agent started — model: Claude 4.6 Opus
============================================================
Ill startng the optimizationtimization skill and gathering state from Jira in parallel.
→ [read] /Users/cpwork/j/spree-jiffyshirts-spec-optimizer/.cursor/skills/jiffy-optimize-spec/SKILL.md
→ [mcp]
@CrazyPit
CrazyPit / cycle_1_20260313_111404.log
Created March 13, 2026 09:34
Observer Agent — cycle log example (processes 3 PRs: graduates merged, checks CI, responds to reviewer)
============================================================
Agent started — model: Claude 4.6 Opus
============================================================
Illartadingetcreated`.the MN-3189 ticket to find all comments with `Status: PR created`.
→ [mcp]
Ifound ts with `Status: PR createdR created`. Let me check the state of all 3 PRs in parallel.
→ [shell] $ gh pr view 49775 --json state,mergedAt --repo sdtechdev/spree-jiffyshirts
/**
* Created by petrrezikov on 27.07.15.
*/
package ru.innopolis.crazypit.SokobanSolver
import scala.collection.mutable.HashMap
class SokobanSolver(val map: Array[Array[Field]]) {