claude-wannabe-gpt

claude-wannabe-gpt

English | Tiếng Việt

Moved. This skill now lives in duu261/agent-skills. This repo is archived and no longer updated.

For operators who let coding agents work on live servers.

Fast agents fail on production in a predictable way: they chain a read, an edit, and a restart in one command, peek at .env to find a setting, and treat their own "want me to...?" as a yes. If you came from GPT models, you notice what is missing: GPT asks before it changes things, keeps secrets out of its output, makes one change at a time, and gets a second opinion before a deploy.

This repo packages that strictness as an Agent Skill: a mindset plus safe defaults. Strict about changes, fast about reads. It also lists the strict-model habits to avoid, so you get the discipline without the ceremony.

The skill is named production-ops-discipline. The Agent Skills spec reserves "claude" in skill names, so only the repo carries the joke.

What it teaches

• Classify the latest message before acting. A question is not approval. Your own offer is not a go.
• Write a change line before each mutation: target, exact change, impact, rollback.
• One change per call, with set -euo pipefail, a backup, and a verify step.
• Secret hygiene: never read .env or credential files; allowlist output instead of filtering it.
• Independent review on fixed triggers (prod config, releases, migrations, deletions), not on every message.
• A block is an answer: stop and report, never route around a guardrail.
• Do not over-correct: no permission questions for reads, no asking twice, no SLA prose in a small README.

Built and tuned on Claude Opus 5.5. The mechanical defaults (backups, one change per call, secret hygiene, fixed review triggers) are simple patterns that should carry over to other models, but that is untested. The judgment sections assume a strong model; pair weaker models with harness deny rules or hooks.

Install

Easiest: let your agent do it

Paste this into Claude Code, Codex, Hermes Agent, or any coding agent that supports Agent Skills:

Install the Agent Skill "production-ops-discipline" from https://github.com/duu261/agent-skills.
1. Clone the repo into a temporary directory.
2. Copy the folder skills/production-ops-discipline into the user-level skills directory of the agent you are running in (Claude Code: ~/.claude/skills/, Codex: ~/.agents/skills/, Hermes Agent: ~/.hermes/skills/). If you are not sure which one applies, ask me.
3. If a folder with that name already exists, stop and ask me before replacing it.
4. Do not install the optional delegation hook unless I ask for it.
5. Confirm the installed path and that SKILL.md is there, then delete the temporary clone.

Manual

Copy skills/production-ops-discipline/ into your harness's skills directory:

git clone https://github.com/duu261/agent-skills
cp -r agent-skills/skills/production-ops-discipline ~/.claude/skills/

The skill loads when a task touches production. Invoke it directly with /production-ops-discipline in Claude Code.

For the Opus 5.5 delegation hook (why the model needs a per-turn user-side nudge to use subagents) and harness-specific tool mapping, see references/harnesses.md.

Evidence

Field notes from one operator running Claude Opus 5.5 on a production server, about 12,000 tool calls. This is observational, not a controlled test.

After the skill, Opus matches or beats both GPT models on working directory, cd, and recursive grep. Shell file reads are the clear gap that stays, and in-place edits are close but still above both. The GPT columns are from the same server before the skill, with different periods and task mix, so read them as directional.

"Opus after" means the skill plus matching standing rules that landed the same evening, so this table cannot credit the skill alone. Review delegation barely moved until a one-line per-turn prompt hook was added. Confounds and the full numbers are in references/evidence.md. A controlled A/B in Claude Code is planned.

License

MIT