Releases·Tool Release
Dev releases llm-cure, MIT-licensed CLI that installs MD principles steering coding agents away from regex overuse and 'cheating'
A developer frustrated with coding agents' mechanical problem-solving — regex where judgment is needed, excessive self-asserting tests, mocking everything — distilled corrective principles into two markdown files and packaged them as llm-cure, an MIT-licensed npm CLI. The tool installs the principles across supported coding agents and offers an optional Agent Skill for Claude Code and Codex. It targets common LLM shortcuts and encourages wider, evidence-driven solution search.
4 engagement1 sourceWed, Sep 23, 2026, 10:33 AM
AI Overview
- llm-cure ships two MD files: LLM_BOOST sets habits (evidence-based checks, clear uncertainty) and LLM_CURE adds four diagnosis methods for recurring problems
- Author built it after policing agents that used regex where LLM judgment was needed and mocked everything in excessive unit tests
- Available on npm and GitHub under MIT license, installs principles across supported coding agents, with optional Agent Skill for Claude Code and Codex
- Author refined the approach through repeated agent runs in one codebase and reports improved coding-agent performance
Entities
Related
Social·Community
9 prompt rules cut coding agent wasted thinking up to 70% (GLM 5.3 & GLM 5.3 Flash)
A developer posted a 9-rule 'thinking discipline' instruction block that reduced wasted reasoning tokens in GLM 5.3 and GLM 5.3 Flash coding-agent sessions by up to 70% across 360 A/B runs. The rules target premise-checking, approach-switching, and re-verification loops. The author released the exam harness so others can replicate the test.
A developer posted a 9-rule 'thinking discipline' instruction block that reduced wasted reasoning tokens in GLM 5.3 and GLM 5.3 Flash coding-agent sessions by up to 70% across 360 A/B runs. The rules target premise-checking, approach-switching, and re-verification loops. The author released the exam harness so others can replicate the test.
37 engagement1 sourceredditTue, Sep 29, 2026, 06:44 AM
Releases·Tool Release
Developer shares Empirical Diagnostics skill for Claude Code debugging
A developer built a skill for Claude Code that forces structured diagnosis before code edits, addressing the agent's tendency to jump to speculative fixes. The skill is shared on Reddit for community use.
A developer built a skill for Claude Code that forces structured diagnosis before code edits, addressing the agent's tendency to jump to speculative fixes. The skill is shared on Reddit for community use.
2 engagement1 sourceredditThu, Aug 20, 2026, 06:45 AM
Developments·Paper
Study identifies agent-reactive bugs at model-harness boundary in LLM agents
A new empirical study analyzes bug reports from LLM agent systems like Codex and LangChain, revealing that many failures arise from interactions between model outputs and harness code rather than from either component alone. The research categorizes these 'agent-reactive bugs' and provides a taxonomy to help developers debug and design more robust agents.
A new empirical study analyzes bug reports from LLM agent systems like Codex and LangChain, revealing that many failures arise from interactions between model outputs and harness code rather than from either component alone. The research categorizes these 'agent-reactive bugs' and provides a taxonomy to help developers debug and design more robust agents.
0 engagement2 sourcesarxivFri, Jul 17, 2026, 06:53 AM
Social·Community
Developer argues agent doom loops stem from failed attempts dominating the transcript, not model capability
A Reddit post offers a mechanical explanation for agent failure loops: because models are stateless and re-read the full session transcript each turn, repeated failed attempts become the dominant text pattern, biasing the model toward reproducing them. The author cites an agent that opened the same file eleven times while apologizing. The takeaway for practitioners is that prompt-level pleas like 'try a DIFFERENT approach' don't reliably break the loop.
A Reddit post offers a mechanical explanation for agent failure loops: because models are stateless and re-read the full session transcript each turn, repeated failed attempts become the dominant text pattern, biasing the model toward reproducing them. The author cites an agent that opened the same file eleven times while apologizing. The takeaway for practitioners is that prompt-level pleas like 'try a DIFFERENT approach' don't reliably break the loop.
9 engagement1 sourceredditTue, Sep 1, 2026, 03:00 AM
Social·Community
Practitioners track AI output failures to improve prompts
A practitioner reports that logging failed AI outputs is more useful for prompt improvement than saving successful ones. Categorizing failures reveals recurring patterns, shifting the focus from optimizing prompts to addressing specific failure modes.
A practitioner reports that logging failed AI outputs is more useful for prompt improvement than saving successful ones. Categorizing failures reveals recurring patterns, shifting the focus from optimizing prompts to addressing specific failure modes.
17 engagement1 sourceredditFri, Aug 14, 2026, 06:20 AM