Ponytail：让AI代理像懒惰的高级开发人员一样思考

Ponytail He says nothing. He writes one line. It works. Ponytail 5: rebuilt from the ground up. -53% code · -41% time · -26% cost · -45% tokens And yet: 98% of risky logic ships with a test. Without Ponytail: 68%. Benchmarked in Claude Code, the same agent with and without the skill: 39 tasks including a real FastAPI + React repo, Opus 5.5, 5 runs each. Details. Español · 한국어 · 简体中文 · 日本語 --- ## Already built with Ponytail --- You know him. Long ponytail. Oval glasses. Has been at the company longer than the version control. You show him fifty lines; he looks at them, says nothing, and replaces them with one. Ponytail puts him inside your AI agent. ## Numbers Two things the chart does not show: in a blind comparison, Ponytail 5's replies beat the previous Ponytail's 110 to 67. And on the six security tasks (SQL injection, path traversal, forged tokens, rate limiting, malformed CSV rows, caching) it passed all 30 runs: less code, no less safe. Method, per-task tables and limits: [benchmarks/results/2026-10-07-agentic.md](benchmarks/results/2026-10-07-agentic.md). **The rule was never "fewest tokens."** It is: write only what the task needs, and never cut validation, error handling, security, or accessibility. The code ends up small because it is necessary, not golfed. Lower cost and latency are a side effect. ## Before / after You ask for a date picker. Without Ponytail, the agent installs a date picker library or builds a whole calendar by hand: 335 lines. Ponytail 5 first looks at what is already there: the repo has an `Input` component, and every browser has a date picker. It puts the two together. 10 lines. More survivors in [examples/](examples/). ## The review, rebuilt `/ponytail-review` used to look only for code to cut. Now it reviews like the senior dev who gets paged when it breaks: it reads the code your change touches, not just the diff, and checks bugs, security, real load, missing tests, speed, and what to cut. Each finding says what the code does, what goes wrong, how to fix it, and what happens if you don't. ## The audit, rebuilt `/ponytail-audit` runs the same checks on the whole repo. It maps the code first: entry points, how data moves, what load the project expects. Then it ranks what it finds and tells you what to fix first. The old audit only listed what to delete. ## New in 5.3: the band In Claude Code, ponytail now shows a band above the prompt: what the current task costs, how many tokens and how much agent time it used, how often it reused existing code instead of writing new code, and whether the tests passed after the last change. Its buttons switch the level, run `/ponytail-review` or `/ponytail-audit`, copy the task as Markdown, and fold the band into one line. Don't want it? `/ponytail band` hides it, the same command brings it back. Needs Claude Code 2.1.290 or later, in the terminal or the desktop app. It does not show in the VS Code chat panel, cloud sessions or `claude -p`. **Board** opens a local page with every task across your projects: cost, agent time and tokens over time, every session with its tests, and the days you worked. By default, everything stays on your machine, in `~/.config/ponytail/band/`. `/ponytail band import` adds your past Claude Code sessions to it, with their cost, agent time and tokens. ## How it works The ladder runs *after* it understands the problem, not instead of it: it reads the code the change touches and traces the real flow before picking a rung. Lazy about the solution, never about reading. Lazy, not negligent: trust-boundary validation, data-loss handling, security, and accessibility are never on the chopping block. Logic with a branch, a loop, a parser, money or security leaves one small test behind. Every reply ends with what was skipped or not checked and any risk you should know. ## The prompt Ponytail is one prompt: [`skills/ponytail/SKILL.md`](skills/ponytail/SKILL.md). The compact version, for agents that read a rules file, is [`AGENTS.md`](AGENTS.md). Everything else in this repo loads that prompt into different agents. ## Install **Claude Code**, as two separate prompts: ``` /plugin marketplace add DietrichGebert/ponytail ``` ``` /plugin install ponytail@ponytail ``` **Codex:** ```bash codex plugin marketplace add DietrichGebert/ponytail codex plugin add ponytail@ponytail ``` Then open `/hooks` in Codex, trust its two lifecycle hooks, and start a new thread. **Any other agent:** copy [`AGENTS.md`](AGENTS.md) into your project, or ask your agent to install [`skills/ponytail/SKILL.md`](skills/ponytail/SKILL.md) as a skill. Step by step for Copilot, Cursor, OpenCode, Gemini and the rest: **[INSTALL.md](INSTALL.md)**. That was it. He'd be proud. He won't say it. Active every session, with a handful of commands (see [Commands](#commands)). `/ponytail ultra` exists for when the codebase has wronged you personally. Startup and mode-change text shows the current mode. Only install ponytail from `DietrichGebert/ponytail` on GitHub or `@dietrichgebert/ponytail` on npm. It never ships `.exe` or `.dll` files; a copy that does is not mine. ## Commands | Command | What it does | |---------|--------------| | `/ponytail [lite \| full \| ultra \| off]` | Set the intensity, or turn it off. No argument switches ponytail on at the default level if it is off, and otherwise reports the current level. | | `/ponytail band [on \| off]` | Claude Code only: hide or show the band above the prompt. No argument switches it. The level stays. | | `/ponytail band import` | Claude Code only: add your past Claude Code sessions to the board, with their cost, agent time and tokens. | | `/ponytail-review` | Review the current diff like the senior dev who gets paged when it breaks: bugs, security, real load, risky code without a test, slow paths, and what to cut. Each finding says what the code does, what goes wrong, how to fix it, and what happens if you don't. Name a target in plain words to narrow or widen it: `uncommitted`, `staged`, `branch`, or a PR link. | | `/ponytail-audit` | The same check for the whole repo, most important first. | | `/ponytail-debt` | Harvest the `shortcut:` comments you've deferred into a ledger, so "later" doesn't become "never". | | `/ponytail-gain` | Show the measured impact scoreboard (less code, less cost, more speed) from the benchmark. | | `/ponytail-help` | Quick reference for the commands above. | Commands need a skill-capable host (Claude Code, Codex, Devin CLI, OpenCode, Gemini, Antigravity CLI, pi, Hermes Agent, Qoder, Grok Build). In Codex CLI and the IDE extension they're skills under the plugin's namespace; invoke with `$ponytail:ponytail-review`. Cursor with the [hooks](INSTALL.md#cursor) gets `/ponytail` level switching only, typed as a plain message. The instruction-only adapters (Cursor's rule file, Windsurf, Cline, Copilot, Kiro, Antigravity's `AGENTS.md` fallback) load the always-on ruleset without the commands. ## FAQ **Does it need a config file?** No. An optional `~/.config/ponytail/config.json` or `PONYTAIL_DEFAULT_MODE` env var can set the default level, but nothing is required. **Why does it write `shortcut:` comments?** They mark a deliberate shortcut and when to revisit it, and `/ponytail-debt` collects them into a ledger. Want another word, or none? Say so in your project's `CLAUDE.md` or `AGENTS.md`, then run `/ponytail-debt `. **What if I really need the 120-line cache class?** You don't. Insist anyway and he'll build it. Slowly. Correctly. While looking at you. **Does it scale?** The code you never wrote scales infinitely. Zero bugs, zero CVEs, 100% uptime since forever. **Why "ponytail"?** You know exactly why. ## Sponsors ## License [MIT](LICENSE). The shortest license that works. ## Star History