Releases·Tool Release
Dev cuts .cursorrules token tax 73% with pseudocode 'Agent Action Grammar' and open-source okf CLI for rule syncing
The okf project released v0.3.0 of an open-source Go CLI that replaces verbose English prose rules files (.cursorrules, CLAUDE.md, copilot-instructions.md) with structured pseudocode it calls Agent Action Grammar. The authors report cutting a rules file from 420 to 115 tokens (73%) while improving model adherence, and ship a single canonical AGENTS.md symlinked across tools.
29 engagement1 sourceTue, Sep 15, 2026, 07:36 PM
AI Overview
- A Reddit user reports .cursorrules written as English prose (~420 tokens) suffer 'context tax' and behavioral drift in Cursor's Claude/GPT models.
- Reframing rules as EBNF-style pseudocode with RFC 2119 verbs (MUST, NEVER) cut the example file to 115 tokens, a 73% reduction.
- The author released okf v0.3.0, a Go CLI that symlinks .cursorrules, CLAUDE.md, and copilot-instructions.md to one canonical AGENTS.md with static AST linting.
- A token gate under 400 tokens keeps root working memory compact; heavier docs move to a pull-based memory tier accessed via MCP tools.
Related
Social·Community
Users report Cursor burns tokens fast, seek refunds
A user reports difficulty getting a refund for Cursor, saying it is less capable than Claude Code and Codex and burns through millions of tokens in a few prompts. The user regrets the annual subscription purchase and is considering a chargeback.
A user reports difficulty getting a refund for Cursor, saying it is less capable than Claude Code and Codex and burns through millions of tokens in a few prompts. The user regrets the annual subscription purchase and is considering a chargeback.
11 engagement1 sourceredditThu, Aug 6, 2026, 03:02 PM
Developments·Paper
SkillGate paper proposes runtime detection of malicious skill files in coding agents
A new arXiv paper introduces SkillGate, a cost-efficient runtime detector for malicious skill files in AI coding agents like Cursor, Claude Code, and GitHub Copilot. It targets the supply-chain risk of skill files installed via commands like npx skills add, which can exfiltrate credentials or inject backdoors. The approach aims to flag malicious Markdown-based skills without heavy overhead.
A new arXiv paper introduces SkillGate, a cost-efficient runtime detector for malicious skill files in AI coding agents like Cursor, Claude Code, and GitHub Copilot. It targets the supply-chain risk of skill files installed via commands like npx skills add, which can exfiltrate credentials or inject backdoors. The approach aims to flag malicious Markdown-based skills without heavy overhead.
0 engagement1 sourcearxivTue, Jul 28, 2026, 11:50 AM
Social·Community
ZDNET: 80% of developers find AI coding more addictive than helpful
A ZDNET report highlights that 80% of developers find AI coding tools addictive but not necessarily helpful, citing exhaustion and potential burnout. The article notes that while these tools speed up boilerplate generation, fixing AI-written code can delay overall progress.
A ZDNET report highlights that 80% of developers find AI coding tools addictive but not necessarily helpful, citing exhaustion and potential burnout. The article notes that while these tools speed up boilerplate generation, fixing AI-written code can delay overall progress.
0 engagement1 sourcerssFri, Aug 21, 2026, 07:30 PM
Social·Community
Teams split on reviewing single-dev AI features as 10,000-line branches pile up 'AI Comprehension Debt'
Multiple Reddit threads in late August and early September 2026 describe code review breaking down as developers use Cursor, Claude Code, Codex, and Copilot to generate large features fast. One team merged 214 PRs in a month with the poster properly reading about 12, and another calls agent diffs of 20-40 files unreviewable despite green tests. Teams are debating which AI first-pass tools (CodeRabbit, Copilot review, Claude as a second model) add real value versus noise.
Multiple Reddit threads in late August and early September 2026 describe code review breaking down as developers use Cursor, Claude Code, Codex, and Copilot to generate large features fast. One team merged 214 PRs in a month with the poster properly reading about 12, and another calls agent diffs of 20-40 files unreviewable despite green tests. Teams are debating which AI first-pass tools (CodeRabbit, Copilot review, Claude as a second model) add real value versus noise.
Updated Mon, Sep 7, 2026, 05:50 AM — A September 2 thread extends the debate to team workflow: how to peer-review a 10,000+ line AI-built feature across 50+
021114 engagement4 sourcesredditFri, Aug 28, 2026, 06:46 AM
Developments·Paper
Scanning the Harness: An Empirical Study of Supply-Chain Defects in AI Coding-Agent Configurations
A new empirical study examines supply-chain risks in how developers configure AI coding agents like Claude Code, Cursor, GitHub Copilot, and OpenAI Codex. It finds configuration artifacts — instruction files, skills, hooks, MCP server declarations, subagents — form an unmanaged dependency layer installed from public repositories, running with developer privileges and lacking lockfiles, install-time checks, or capability declarations. The work gives practitioners a vocabulary for a widely overlooked attack surface.
A new empirical study examines supply-chain risks in how developers configure AI coding agents like Claude Code, Cursor, GitHub Copilot, and OpenAI Codex. It finds configuration artifacts — instruction files, skills, hooks, MCP server declarations, subagents — form an unmanaged dependency layer installed from public repositories, running with developer privileges and lacking lockfiles, install-time checks, or capability declarations. The work gives practitioners a vocabulary for a widely overlooked attack surface.
0 engagement1 sourcearxivMon, Sep 7, 2026, 11:27 AM