01 / less
Less code to maintain
Skip speculative features, wrappers, and abstractions that don’t earn their keep.
01 / an idea worth stealing
The best code is the code you never wrote.
Ponytail puts a lazy senior developer inside your AI coding agent: understand the problem, then stop at the smallest solution that actually works.
read firstbuild lessship correct
02 / tl;dr for humans
A small set of rules that makes AI coding agents less wasteful.
It is not a library that writes code. It is a behaviour layer that changes what the agent reaches for first.
02 / native◌
More platform leverage
Use the browser, standard library, installed dependencies, or existing project patterns before adding anything new.
03 / safe✓
Small without being careless
Validation, security, error handling, accessibility, and data-loss protection stay in scope.
03 / the failure mode
AI agents are very good at building the thing you didn’t need.
The point isn’t to chase the lowest line count. The point is to avoid paying for code that adds no value, while keeping the code that protects users.
04 / the instinct
Ask for a date picker. Watch two instincts compete.
agent-without-ponytail
// install a package import DatePicker from "date-picker-lib"; // wrapper, styles, state, options… <DatePicker minDate={new Date()} locale="en-AU" timezone="Australia/Sydney" />
More code. More surface area. More things to own.
→
agent-with-ponytail
<!-- ponytail: browser has one --> <input type="date">
Start native. Add complexity only when the task proves it’s needed.
05 / how it thinks
Stop at the first rung that holds.
06 / the guardrail
Lazy about the solution. Never lazy about reading.
Ponytail explicitly keeps the work that protects people, systems, and data.
What stays
✓Trust-boundary validation
Untrusted input still gets checked.
Untrusted input still gets checked.
✓Security
Path traversal, injection, forged tokens: not optional.
Path traversal, injection, forged tokens: not optional.
✓Accessibility + error handling
Small code is not an excuse to make a broken interface.
Small code is not an excuse to make a broken interface.
The benchmark’s useful distinction
A bare “YAGNI + one-liners” prompt produced the fewest lines on one safety task, but dropped a path-traversal guard once in four runs. Ponytail kept the guard in all four.
Minimalism is a judgement system, not a command to delete lines.
07 / measured, with caveats
The honest benchmark result.
Select a metric. Bars are Ponytail and control arms relative to the no-skill baseline.
On average across 12 feature tasks, Ponytail left 46% of baseline LOC: a 54% reduction. This is an aggregate, not a promise for every task.
08 / where it earns its place
It shines where there is an over-build trap.
Strong fit
- Greenfield feature work where agents may reach for a package first.
- Projects with a healthy standard library or capable platform primitives.
- Teams that want a shared review language for “why is this here?”
- Multi-agent workflows where behaviour should travel with the project.
Not a magic wand
- Irreducible domain logic still needs its real code.
- A command palette does not become native just because you wish it would.
- Smaller code is not automatically faster on every reasoning model.
- A benchmark is evidence, not a security certification.
09 / not locked to one agent
One behaviour layer. Many hosts.
Ponytail keeps the core skills portable, then uses thin adapters or project instruction files for each host.
Claude CodeCodexGitHub CopilotGemini CLIHermes AgentOpenCodePiCursorWindsurfClineQoderCodeWhaleZedKiro…and more
The portable fallback is simple: put the repo’s AGENTS.md rules in the project so any compatible agent can read them.
10 / install in minutes
Pick your agent. Copy the smallest install.
CODEX / PLUGIN
codex plugin marketplace add DietrichGebert/ponytail codex plugin add ponytail@ponytail
Then open /hooks, review and trust the two lifecycle hooks, and start a new thread.
11 / step-by-step implementation guide
Bring the behaviour into your own project.
Use a native adapter, or copy AGENTS.md into the project root.
Start with full. Use lite for a gentler adoption.
Describe the outcome. Let the agent read the touched code and trace the flow first.
Ask @ponytail-review or /ponytail-review what can be cut.
Run the audit and track deliberate simplifications with their upgrade trigger.
test the smallest working pathkeep safety guardsmeasure before claiming savingsadd complexity only when evidence demands it
12 / your own operating rule
Turn the ladder into a project instruction.
If you don’t need the plugin, this is the portable kernel. Save it as AGENTS.md, then let your agent apply it on every task.
AGENTS.md / minimal kernel
Before writing code: 1. Does this need to exist? 2. Can we reuse what’s already here? 3. Does the standard library or platform do it? 4. Is the dependency already installed? 5. What is the minimum correct implementation? Read first. Keep validation, security, error handling, accessibility, and data-loss protection.
13 / make it fit your workflow
Four levels. A few useful commands.
lite
Build what’s asked
Name the lazier alternative, but don’t push the full ladder into every decision.
full / default
The full ladder
YAGNI → reuse → stdlib → native → installed dependency → one line → minimum.
ultra
Deletion before addition
Challenge speculative requirements and ship the smallest thing that survives scrutiny.
@ponytail-reviewReview the current diff for over-engineering.
@ponytail-auditAudit the whole repo, ranked by biggest cut.
@ponytail-debtHarvest deliberate shortcuts into a debt ledger.
14 / the handoff
Use AI to write less. Think more clearly.
Ponytail’s real contribution is a repeatable judgement loop: understand the problem, use what already exists, and stop before the code becomes somebody else’s maintenance burden.
Lazy means efficient, not careless.
MIT licensed · open source · agent-portable