{"slug": "how-to-stop-correcting-ai-coding-agents-on-the-same-conventions", "title": "How to stop correcting AI coding agents on the same conventions", "summary": "Harbor, a tool launched by an unnamed vendor behind the `npx harborloop init` command, learns team coding conventions from a single mention in a session, review thread, Slack or doc and injects at most three ranked rules matched to the task and files before each prompt across Claude Code, Codex, Cursor and GitHub Copilot. In a controlled test with one saved rule, the agent got 10 of 10 right on the first try with 0 corrections typed and 38% less agent time, which the company says is a test rather than a customer result. Harbor reviews each pull request against the team's rules and is free for one user and one teammate with no card and no end date.", "body_md": "Blog\n\n# How to stop correcting AI coding agents on the same conventions\n\nPublished\n\nYou tell Claude Code to write a migration, not reset the database. Next session, a teammate's agent does the reset. You correct it again.\n\n## Why the usual fix falls short\n\nEvery agent has an instruction file: `CLAUDE.md`, `AGENTS.md`, Cursor rules, `copilot-instructions.md`. They help. They also have limits:\n\n- Someone has to write and update them.\n- They go stale as the code moves ([stale rules](https://gethrbr.com/blog/context-rot-stale-rules) ).\n- They are per repo, and per agent ([keeping them in sync](https://gethrbr.com/blog/keep-claude-md-agents-md-cursorrules-in-sync) ).\n- A long file is loaded whole, so important rules get lost ([why a growing CLAUDE.md stops working](https://gethrbr.com/blog/why-growing-claude-md-stops-working) ).\n\nLinters like ESLint and Prettier handle mechanical rules well. Keep using them. Conventions that need judgment still land on you.\n\n## What Harbor does\n\n- **It learns the rule.** When someone says it once, in a coding session, a review thread, Slack or a doc, Harbor keeps it as a rule. You choose whether every rule waits for a person to approve it.\n- **It hands over only what fits.** Before each prompt, the agent gets at most three ranked rules matched to the task and the files it touches.\n- **It reaches everyone.** Every teammate gets the same rules, in Claude Code, Codex, Cursor and GitHub Copilot, on any machine and repo the Space covers.\n- **It checks the result.** Each pull request is reviewed against your rules ([how to review pull requests against your own team's rules](https://gethrbr.com/blog/review-pull-requests-against-team-rules) ).\n- **It keeps rules true.** Conflicts, disputes and stale rules are handled.\n\n## A controlled test\n\nIn a controlled test with one saved rule, the agent got 10 of 10 right on the first try, with 0 corrections typed and 38% less agent time. This is a test, not a customer result. The setup and what it leaves out are on the [Harbor home page](https://gethrbr.com/).\n\n## Start\n\nRun `npx harborloop init`.\n\nRelated: [how to stop Claude Code and Cursor from making the same mistakes on your codebase](https://gethrbr.com/blog/stop-ai-coding-agents-repeating-mistakes), and [how to make Claude Code remember your corrections](https://gethrbr.com/blog/ai-agent-keeps-making-the-same-mistake).\n\n## Questions\n\n### How do I stop correcting AI coding agents on the same team conventions every session?\n\nWrite the convention into the file each agent reads, such as CLAUDE.md, AGENTS.md, Cursor rules or copilot-instructions.md, and keep linters for mechanical rules. Files need someone to write and update them, go stale, and are per repo and per agent. Harbor learns the convention when someone says it once and gives every teammate's agent the rules that fit each task.\n\n### Why do CLAUDE.md and AGENTS.md not stop the repeated corrections on their own?\n\nSomeone has to write and update them, they go stale as the code moves, they are per repo and per agent, and a long file is loaded whole, so important rules get lost.\n\n### How many rules does Harbor give an agent?\n\nAt most three ranked rules before each prompt, matched to the task and the files it touches, not the whole rulebook.\n\n### Which coding agents does Harbor work with?\n\nClaude Code, Codex, Cursor and GitHub Copilot, connected with npx harborloop init. It is free for you and one teammate, with no card and no end date.", "url": "https://wpnews.pro/news/how-to-stop-correcting-ai-coding-agents-on-the-same-conventions", "canonical_source": "https://gethrbr.com/blog/stop-correcting-ai-coding-agents-same-conventions", "published_at": "2026-11-22 00:00:00+00:00", "updated_at": "2026-11-22 08:13:43.578499+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools"], "entities": ["Harbor", "Claude Code", "Codex", "Cursor", "GitHub Copilot", "ESLint"], "also_reported_by": [], "takeaway": "Harbor injects at most three ranked rules per prompt instead of loading a whole CLAUDE.md or AGENTS.md, and its controlled single-rule test produced 10 of 10 first-try correct answers with 0 corrections and 38% less agent time.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/how-to-stop-correcting-ai-coding-agents-on-the-same-conventions", "markdown": "https://wpnews.pro/news/how-to-stop-correcting-ai-coding-agents-on-the-same-conventions.md", "text": "https://wpnews.pro/news/how-to-stop-correcting-ai-coding-agents-on-the-same-conventions.txt", "jsonld": "https://wpnews.pro/news/how-to-stop-correcting-ai-coding-agents-on-the-same-conventions.jsonld"}}