Ponytail makes a coding agent check seven things before it writes any new code

Ponytail is a rule set that changes what a coding agent does before it starts writing. Instead of going straight to new code, the agent has to work down a ladder of cheaper answers first. The project is MIT licensed, was created on 12 June 2026, and had 127,785 stars and 6,833 forks when the GitHub API was checked on 5 September. It appeared on Product Hunt on 4 September, which is what put it back in front of people.

The ladder

The rules give the agent seven questions to answer in order. Does this need to exist at all? Is it already somewhere in this codebase? Can the standard library do it? Is there a native platform feature for it? Is it in a dependency that is already installed? Can it be one line? Only after all seven does the agent build something, and then only the smallest version that works.

The project is explicit that the laziness applies to the solution and never to the reading. The agent is expected to trace the real code path before it decides which rung it is on, so "already in this codebase" is a claim it has checked rather than guessed. Four things are marked as never cuttable for the sake of brevity: trust boundary validation, data loss handling, security, and accessibility.

What the project measured, by its own account

The repository publishes a benchmark, and it is the project's own, so read it as the author's claim rather than an independent result. The author says they ran twelve feature tickets through a headless Claude Code session on Haiku 4.5, with and without the rules, on tiangolo's full-stack-fastapi-template, at n=4, and scored the git diff each run left behind.

On those runs the author reports 54 percent fewer lines of code, 22 percent fewer tokens, 20 percent lower cost, and 27 percent less time, with safety held at 100 percent. The caveat is stated in the same place: the saving is largest where there was a real over-build trap and close to zero on code that was already minimal. The author also notes that an earlier single-shot benchmark overstated the gap, and calls that a conversational-baseline artifact they corrected by moving to the agentic test.

The rules ship for more than twenty hosts, including Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, OpenCode, Cursor, Windsurf and Cline, injected through lifecycle hooks where the host has them and through a persistent rule file where it does not.

A coding agent is measured on whether it produced working code, and nothing in that loop pushes back on how much code it produced. The result lands on whoever reviews it: a date picker written from scratch passes the same tests as three lines calling a library, and it is the reviewer who then owns it for years. What Ponytail does is put the argument for reuse in front of the model at generation time, where it is cheap, rather than at review time, where it means someone reading a diff and asking for it to be smaller. Whether the specific percentages hold on a different codebase is untested, but the direction is one a team can check for itself on a single pull request.

Source

DietrichGebert/ponytail, GitHub, MIT licensed, created 12 June 2026. Launched on Product Hunt on 4 September 2026.