Ponytail skill: less code from your coding agent, gains to verify

Ponytail skill: MIT licence, installation from the README, author-declared gains (-53% code) and a method to measure them with OpenTelemetry metrics.
Ponytail skill agent : moins de code, gains à vérifier

The Ponytail skill is an instruction file that asks a coding agent to write less code: the project presents it as a “lazy senior dev” who replaces fifty lines with one. The DietrichGebert/ponytail repository shows around 172,300 stars at the end of October 2026 for only 363 commits, which justifies a critical look at what this skill really brings.

A skill is an instruction file the agent loads into its context. It does not change the model: it changes its behaviour, and therefore the amount of code produced, the token cost and the working time. This article presents the project according to its README, shows the visuals the repository provides, explains installation, then proposes a method to verify the announced gains with OpenTelemetry. A note on honesty: span.observer has not installed this skill; the figures quoted are the author’s, and the images come from the repository.

Ponytail skill: what the repository says

  • Licence: MIT, 2026 copyright in the name of the repository’s maintainer (DietrichGebert).
  • Figures read on 22 October 2026: about 172,300 stars, about 9,400 forks, 363 commits; the version shown in package.json is 5.3.2.
  • Nature of the project: “Ponytail is one prompt”: a skills/ponytail/SKILL.md file and a compact AGENTS.md version, which adapters load into different agents (Claude Code, Codex, OpenCode, Gemini, Cursor, Copilot and others).
  • Core rule: write only what the task needs, never cutting validation, error handling, security or accessibility. The README insists that the rule is never “fewest tokens”.

The skill’s method is a “ladder”: stop at the first rung that holds (does it need to exist? already in the repository? in the standard library? a native platform feature? an installed dependency?), then write the minimum and one small test if the logic warrants it. Every reply ends with what was skipped or not checked. That is consistent with the safeguards described in our guardrails for agents in production.

Tour: the band and the README figures

Since version 5.3, the project adds a band above the prompt in Claude Code. According to the README, it shows the current task’s cost, the tokens and agent time used, the share of existing code reused, and whether the tests pass after the last change. It needs Claude Code 2.1.290 or later, and does not appear in the VS Code chat panel or with claude -p.

Ponytail band in Claude Code, in detail then compact mode: cost 23.80 dollars, 56.7 million tokens, agent time 52 minutes, reuse 71 percent, tests passed after the last change
Source : github.com

The screenshot comes from the repository; the values are the project’s examples. A “Board” also opens a local page that gathers the tasks of all your projects: cost, time and tokens over time. The README states that everything stays on your machine, in ~/.config/ponytail/band/.

The README’s central chart compares, according to the author, three set-ups: the agent without a skill, Ponytail v4.13 and Ponytail 5.

Chart from the Ponytail README: lines of code at 47 percent of the no-skill baseline, output tokens 55 percent, cost 74 percent, time 59 percent, for 39 tasks repeated 5 times
Source : github.com

Installation step by step

Step 1 — Claude Code

The README asks for two separate prompts:

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

Step 2 — Codex or another agent

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. For any other agent, the README suggests copying AGENTS.md into the project or installing skills/ponytail/SKILL.md as a skill; per-agent details are in INSTALL.md.

Step 3 — Choose the intensity

The /ponytail command accepts lite, full, ultra or off. The README adds /ponytail-review and /ponytail-audit (review of the diff, then of the whole repository), /ponytail-debt (a list of deferred shortcuts) and /ponytail-gain (an impact scoreboard taken from the benchmark).

Excerpt from the Ponytail README: list of /ponytail commands with lite, full, ultra and off, /ponytail band, /ponytail band import, /ponytail-review, /ponytail-audit, /ponytail-debt, /ponytail-gain and /ponytail-help
Source : github.com

The project warns to install only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm, and says it never ships .exe or .dll files.

Limits and open points

  • Figures declared by the author: according to the repository’s benchmark file (7 October 2026), 39 tasks, 5 runs each, so 585 Claude Code sessions without a shell, measured with Opus 5.5. The README announces -53% lines of code, -41% time, -26% cost, -45% output tokens, and 98% of risky logic shipping with a test against 68% without the skill. These are geometric means of per-task medians, with confidence intervals published in the file. It is not an independent measurement.
  • A skill also costs tokens: its instructions are loaded every session. The net result depends on your repository and your tasks.
  • Less code is not always better: the skill may pick a native function instead of a library your team has standardised. Review the result.
  • Hooks and supply chain: the project installs hooks and plugins. Treat it like any code your agent runs (see our article on a security audit skill for the same caution) and pin a version.

How to verify the gains with OpenTelemetry

The band is a handy local indicator, but for a figure of your own, compare sessions with and without the skill (/ponytail off). Claude Code exports the claude_code.token.usage metric through OpenTelemetry, split by token type; the comparison protocol is detailed in our context-mode article and applies identically here: same repository, same model, same prompt, several repetitions, and the sum of all token types. The export set-up for Dynatrace is in supervising a coding agent in the enterprise and the per-task cost logic in the token cost of an AI agent. This holds for Dynatrace SaaS and for Dynatrace Managed receiving OTLP.

Add quality measures: passing tests, review round-trips, bugs found after merge. Shorter code that fails more often is not a saving.

How to check that it works

  1. After installation, /ponytail-help shows the command reference.
  2. Startup or a mode change displays the current mode (the README says so).
  3. On a reference task, the diff’s line count and the tokens go down, without more failing tests.

Best practices

  • Start with lite on a small repository, with a fixed reference task.
  • Keep human review: the skill proposes shortcuts, it does not validate them for you.
  • Record the skill, model and agent versions next to each measurement.
  • Read the shortcut: comments and treat them as debt.

Frequently asked questions

Does Ponytail change the model?

No. It is a prompt (a skill file and a compact rule) that changes the agent’s behaviour. The effects come from those instructions.

Can the -53% code figure be trusted?

It is a result declared by the author, with a method published in the repository. It holds for its 39 tasks and its agent; measure on yours before generalising.

Is an account or an external service needed?

The README mentions none: installation goes through the repository, and the board stays local by default.

Conclusion

The Ponytail skill illustrates a simple, measurable idea: ask the agent to reuse before writing. Its announced gains are the author’s; the right way to adopt them is to measure them on your own tasks, with OpenTelemetry metrics. Source: DietrichGebert/ponytail repository.

span.observer is an independent community blog, not affiliated with Dynatrace LLC or the projects mentioned. Trademarks belong to their owners. Tools change quickly: check commands and options in their official documentation.

Total
0
Shares
Leave a Reply

Your email address will not be published. Required fields are marked *

Previous Post
Headroom : compression de contexte LLM pour agents de code

Headroom LLM context compression for coding agents

Next Post
ai-memory : mémoire agents de code en Rust, MIT et supervision

ai-memory coding agent memory in Rust: MIT licence and monitoring

Related Posts