DietrichGebert/ponytail

Ponytail – a “senior dev” prompt for AI coding agents

What it is – Ponytail is a single‑file prompt (plus a tiny npm package) that you load into an LLM‑powered coding assistant such as Claude Code, Codex, Cursor, Copilot, Gemini, etc. Once installed, the agent behaves like a veteran developer who looks at a change request, says nothing, and returns the minimal implementation that satisfies the task while preserving validation, security and accessibility.

Key ideas

• The ladder – after the agent understands the problem it walks a “ladder” of seven questions (does the feature already exist? can the standard library do it? can it be a one‑liner?, …) and stops at the first rung that works. This forces the agent to reuse existing code and only write what is truly needed.
• Review & audit commands – /ponytail‑review inspects the diff you are about to commit, flagging bugs, security holes, missing tests, performance concerns and unnecessary code. /ponytail‑audit runs the same analysis on the whole repository and ranks the findings so you know what to fix first.
• Band & board – in Claude Code a small “band” above the prompt shows live statistics (tokens used, cost, agent time, reuse rate, test result). The optional board is a local dashboard that aggregates those metrics across all sessions.
• Modes – /ponytail lite|full|ultra|off lets you tune how aggressive the prompt is. “Ultra” will keep digging even when the codebase has treated you badly.
• Zero‑config – the tool works out‑of‑the‑box; an optional ~/.config/ponytail/config.json or PONYTAIL_DEFAULT_MODE env var can set a default mode.

How it works

1. Installation – add the plugin to Claude Code (/plugin marketplace add DietrichGebert/ponytail) or to Codex (codex plugin marketplace add …). For other agents you simply copy the AGENTS.md or skills/ponytail/SKILL.md file into your project.
2. Prompt injection – the skill file is loaded as a system‑level prompt that overrides the agent’s normal behaviour. All subsequent messages are interpreted through the Ponytail rules.
3. Command handling – the plugin registers slash‑commands (/ponytail‑review, /ponytail‑audit, etc.) that the host IDE/CLI forwards to the LLM. The LLM returns a structured answer (code diff, test snippet, risk note) which the host displays to you.
4. Metrics – after each turn the agent reports token usage, cost and whether the generated test passed; the band visualises this, and the board stores a local log under ~/.config/ponytail/band/.

Typical workflow

1. Open a new thread in Claude Code (or start a Codex session).
2. Type `/ponytail` to enable the default mode.
3. Ask for a feature, e.g. “Add a date‑picker to the checkout form”.
4. Ponytail reads the repo, finds the existing `Input` component, and returns a 10‑line file that reuses the native HTML date input.
5. Run `/ponytail‑review` – the agent points out any missing validation and adds a tiny test.
6. Commit the change; the band shows you you saved ~45 % of tokens and $0.72 per turn.

What it claims (backed by the README’s benchmark)

• ‑53 % code, ‑41 % time, ‑26 % cost, ‑45 % tokens compared with a baseline “no‑skill” agent.
• 98 % of risky logic shipped with a test (vs 68 % without Ponytail).
• In security‑focused tasks (SQL injection, path traversal, etc.) it passed all 30 runs while using less code.

Who might use it

• Teams that already rely on LLM coding assistants and want a guard‑rail that enforces minimal, safe changes.
• Solo developers looking to reduce token spend and keep their repos tidy.
• Organizations that need an audit‑style safety net (security checks, test coverage) baked directly into the AI workflow.

Limitations / cautions

• Ponytail is only a prompt; its quality depends on the underlying model (Claude, Gemini, etc.). If the model hallucinates, the safety checks may miss it.
• The “band” visualisation works only in Claude Code 2.1.290+; other hosts get the core prompt but not the UI.
• It assumes the host can run the /plugin commands and expose a small filesystem for the local board – not all IDE integrations support that yet.
• The tool does not ship any compiled binaries; a malicious fork that adds executables would be outside the author’s control.

Bottom line – Ponytail is a well‑documented, open‑source prompt‑based extension that turns generic LLM coding agents into a “senior dev” that writes the smallest correct change possible, while automatically reviewing for bugs, security, and test coverage. It stays within the AI‑agent tooling space and provides concrete metrics to justify its cost‑saving claims.