DietrichGebert/ponytail

A skill/plugin that pushes an AI coding agent to write the least code that solves the task, rather than over-building.

What it is

Ponytail is an always-on ruleset and set of commands for AI coding agents that biases them toward minimal, necessary code. Before writing anything, the agent walks a decision ladder — does this need to exist, is it already in the codebase, does the standard library or platform or an installed dependency already do it — and stops at the first rung that holds. It is aimed at developers who use agentic coding tools and find that those agents tend to over-engineer (installing a dependency and writing a wrapper component where a native element would do).

Key features

An "over-engineering" ladder the agent runs after it understands the problem: YAGNI, reuse, stdlib, native platform feature, installed dependency, one line, then the minimum that works.
Explicit carve-outs so the agent never drops trust-boundary validation, data-loss handling, security, or accessibility in the name of brevity.
Commands for reviewing and auditing: /ponytail-review (over-engineering review of the current diff), /ponytail-audit (whole-repo), /ponytail-debt (ledger of deferred shortcuts), /ponytail-gain (benchmark scoreboard), and /ponytail-help.
Intensity levels — lite, full (default), ultra, off — set per session with /ponytail, or by default via the PONYTAIL_DEFAULT_MODE env var or ~/.config/ponytail/config.json.
Adapters for roughly twenty agents (Claude Code, Codex, GitHub Copilot CLI, OpenCode, Gemini CLI, Qoder, Hermes, Devin CLI, OpenClaw, and others), plus instruction-only fallbacks via AGENTS.md and per-tool rule files.
The active ruleset is also injected into subagents spawned via the Agent tool, scopeable with the PONYTAIL_SUBAGENT_MATCHER regex.

Tech stack

JavaScript; distributed on npm as @dietrichgebert/ponytail (version 4.8.4 in the manifest), MIT licensed.
Ships as an OpenCode plugin (main points at .opencode/plugins/ponytail.mjs) with keywords opencode-plugin, pi-package, skills, qoder.
Two small Node.js lifecycle hooks for Claude Code and Codex, so node must be on the non-interactive shell PATH for always-on activation; the skills still work without it.
Bundled skills under skills/, a Pi extension, and an MCP server (ponytail-mcp/); tests run via node --test.

When to reach for it

Your agent reaches for a library or writes a bespoke component when a native platform feature or existing code would do.
You want a consistent, portable "write only what the task needs" rule across several different coding agents rather than re-authoring prompts per tool.
You want a review or audit pass that hands back a delete-list of over-engineered code.
You want to track shortcuts the agent deferred so they do not silently accumulate.

When not to reach for it

You want the agent to be more verbose or exploratory — ponytail biases the opposite direction.
The README notes a reasoning model that spends thinking tokens deliberating the ladder can end up slower and costlier (it cites GPT-5.5 going the other way on cost/latency), so the cost/speed benefits are not universal.
Your host has no skill or rule-file support at all; the plugin needs at least an instruction-file mechanism to inject the ruleset.
You are optimizing the agent's prose length rather than the amount of code it produces — that is a different problem (see Alternatives).

Maturity signal

The repository is recent (created mid-June 2026) but heavily developed, with a high star count, frequent pushes as of mid-July 2026, an MIT license, and a versioned npm release at 4.8.4. The breadth of adapters, a test suite, and generated-artifact checks point to active, fast-moving maintenance rather than a static prompt drop. The README is candid about its own benchmarks, revising an earlier headline figure downward after issue feedback, which suggests the numbers are maintained rather than frozen marketing.

Alternatives

caveman — use it when the problem is how much the agent says, not how much it builds; the README frames the two as complementary and non-overlapping.
A hand-maintained "YAGNI + one-liners" prompt or project ruleset — use it when you want full control over the wording and are willing to give up the multi-agent adapters, review commands, and safety carve-outs; the README benchmarks this as a separate, lower-scoring arm.

Notes

The published cost, token, and speed reductions come from a benchmark against a fair agentic baseline (a headless Claude Code session editing a real FastAPI + React repo), and the README explicitly retracts an earlier 80–94% single-shot figure as partly a conversational-baseline artifact.
Uninstalling via the host leaves behind state (mode flag, config file, an optional statusline entry); a separate scripts/uninstall.js is provided and is meant to be run before the host's own remove command, since removing the plugin first deletes the script.

Wiki: DietrichGebert/ponytail