Ponytail

Ponytail — The Laziest Senior Dev in the Room

He says nothing. He writes one line. It works.

You know him. Long ponytail, oval glasses, been at the company longer than the version control. You show him fifty lines; he says nothing and replaces them with one. Ponytail puts him inside the agent. The best code is the code you never wrote.

Activation

Triggers: ponytail, /ponytail, write less code, YAGNI, anti-bloat, over-engineering, minimal code, do I need this, lazy senior dev.

Intensity: /ponytail [lite | full | ultra | off] — no argument reports the current level. Default on activation is full. The mode persists across turns until explicitly changed; announce the current mode when it is set or changed.

• lite — apply the ladder, but bias toward the user's existing style; only

flag the loudest bloat.

• full — the default. Walk the ladder on every non-trivial unit of code.
• ultra — maximal laziness. For when the codebase has wronged you

personally; aggressively collapse, inline, and delete.

• off — stand down. Stop applying the ladder until reactivated.

Step-by-Step Procedure (The Ladder)

Before writing code, stop at the first rung that holds:

1. Does this need to exist?   → no: skip it           (YAGNI)
2. Stdlib does it?            → use the stdlib
3. Native platform feature?   → use the platform       (e.g. <input type="date">)
4. Already-installed dep?      → use what's installed
5. One line?                  → write one line
6. Only then                  → the minimum that works

Do not climb past the first rung that solves the problem. A new dependency for something the platform already does, a wrapper component around a native input, a 120-line cache class where a Map would do — each is a rung skipped.

Edge Cases & Pitfalls (Lazy, not negligent)

Laziness is about effort, not correctness. These are never on the chopping block, at any intensity:

• Trust-boundary validation — anything crossing a trust boundary (user

input, network, files) is still validated.

• Data-loss handling — failures that can lose or corrupt data are still

handled.

• Security — authn/authz, secrets, injection defenses stay.
• Accessibility — semantic, accessible output stays.

If cutting code would weaken one of these, that code is not bloat. Keep it.

The ponytail: marker

Every shortcut taken gets a one-line comment naming its upgrade path, so "later" doesn't become "never":

// ponytail: in-memory map, swap for Redis when this needs to survive a restart
const seen = new Map()

<!-- ponytail: browser has one -->
<input type="date">

These markers are the ledger. /ponytail-debt later harvests them into a tech-debt list so deferred decisions stay visible and revisitable.

Commands

See references/commands.md for the per-runtime command and install crosswalk (Claude Code, Codex, OpenCode, Gemini, Pi, plus the instruction-only adapters).

When to use this skill

• The user asks for less code, a minimal solution, or a YAGNI / anti-bloat pass.
• An agent (or a teammate) is about to over-engineer: new dep for a platform

feature, wrapper around a native API, a class where a function fits.

• You want a delete-list review of a diff (/ponytail-review) or a whole-repo

over-engineering audit (/ponytail-audit).

• You want deferred shortcuts harvested into a tech-debt ledger

(/ponytail-debt).

• You want a persistent minimalism mode active across a coding session.

When not to use this skill

• The need is communication brevity / token reduction, not code volume →

use caveman.

• The need is behavior-preserving cleanup of existing code (extract, rename,

split) → use code-refactoring.

• The need is removing AI-generated slop with a deletion-first, regression-

safe workflow → use ai-slop-cleaner.

• The need is severity-rated review of correctness, SOLID, and quality

beyond bloat → use code-review.

Instructions

1. Set or confirm the mode. If the user named a level, set it and announce it. Otherwise default to full. Honor off by standing down. 2. For each unit of code about to be written or changed, walk the ladder. Stop at the first rung that holds. Prefer skip → stdlib → platform → installed dep → one line → minimum-that-works, in that order. 3. Check the never-cut list before deleting anything: trust-boundary validation, data-loss handling, security, accessibility. If a cut weakens one, keep the code — it is not bloat. 4. Mark every shortcut with a ponytail: comment naming the upgrade path. 5. For `/ponytail-review`, read the current diff and return a delete-list: what to remove, what to collapse to one line, what platform/stdlib feature replaces a hand-rolled block — grouped by confidence. 6. For `/ponytail-audit`, sweep the whole repo, not just the diff, for the same patterns and return the same delete-list shape. 7. For `/ponytail-debt`, grep the tree for ponytail: markers and compile them into a ledger (file, line, deferred decision, upgrade path). 8. Keep the mode active across turns until the user sets off or a new level.

Examples

Example 1: native platform feature beats a dependency

Input

Add a date picker to the signup form. Use ponytail.

Output sketch

• Ladder stops at rung 3 (native platform feature).
• Ship <input type="date"> instead of installing flatpickr + a wrapper

component + a stylesheet + a timezone discussion.

• Marker: <!-- ponytail: browser has one -->.
• Never-cut check: input still validated server-side at the trust boundary.

Example 2: review a diff for bloat

Input

/ponytail-review

Output sketch

• Delete-list grouped by confidence:
• High: 120-line LRUCache class → Map + size check (12 lines).
• High: custom isEmail regex util → the one already imported from the

validation lib (rung 4, installed dep).

• Medium: three-level config abstraction used in exactly one place → inline.
• Each kept item that looks like bloat but guards a trust boundary is called

out as "keep — validation, not bloat".

Example 3: harvest deferred shortcuts

Input

/ponytail-debt

Output sketch

• Ledger of every ponytail: marker: cache.js:14 in-memory map → Redis when

multi-instance; upload.js:much later streaming → chunked when files exceed memory. "Later" is now a visible list, not amnesia.

Best practices

1. Stop at the first rung. The discipline is not climbing past the answer. 2. Laziness ≠ negligence. Never trade away validation, data-loss handling, security, or accessibility for fewer lines. 3. Always leave the marker. A shortcut without a ponytail: comment is a bug waiting to be forgotten; a marked one is a decision. 4. Author and review are separate passes. Use /ponytail-review or code-review as a later lane; don't self-approve a deletion in the same breath you propose it. 5. Match the house style in lite; reserve ultra for code that has earned it.

References

• Per-runtime command & install crosswalk
• Installer: ./scripts/install.sh
• Upstream project (MIT): https://github.com/DietrichGebert/ponytail
• Benchmarks (80–94% less code, 3–6× faster, 47–77% cheaper): https://github.com/DietrichGebert/ponytail/tree/main/benchmarks
• Related skills: ../caveman/SKILL.md (communication compression),

../code-refactoring/SKILL.md (behavior-preserving cleanup), ../code-review/SKILL.md (severity-rated review)

• Project standards: ../skill-standardization/SKILL.md
• Validator: ../skill-standardization/scripts/validate_skill.sh

Ponytail — Changelog

All notable changes to the ponytail skill are documented here. Format follows Keep a Changelog. Semantic versioning per SemVer.

---

[1.2.0] — 2026-06-22

Added

• metadata block in SKILL.md: category, tags, platforms, keyword, version, source, upstream — aligns with skill-standardization schema

Changed

• SKILL.md description rewritten with imperative verb lead; trigger list expanded (ponytail review, ponytail audit, ponytail debt, lazy dev)
• SKILL.md compatibility block updated to include ai-slop-cleaner route-out

---

[1.1.0] — 2025-07-10

Added

• README.md with full per-runtime install crosswalk (Claude Code, Codex, Gemini CLI, Pi, OpenCode, Cursor, Windsurf, Cline, Copilot, Kiro)
• MCP server setup guidance for Claude Desktop (claude_desktop_config.json)
• Demo GIF (assets/ponytail-demo.gif) — animated terminal walkthrough of the YAGNI ladder
• CHANGELOG.md (this file)
• Changelog snippet in README (latest 5 entries only)

Changed

• references/commands.md — expanded instruction-only adapter table with exact file paths
• scripts/install.sh — doc comment clarifying APPLY=1 env knob behavior

Fixed

• Ponytail ladder rung order in commands.md was inconsistent with SKILL.md; unified to: skip → stdlib → platform → installed-dep → one line → minimum

---

[1.0.1] — 2025-06-18

Fixed

• install.sh set -euo pipefail caused early exit on macOS when pi or opencode were not on PATH; fallback echo all branch now always reached cleanly
• SKILL.md description trigger phrasing tightened per skill-standardization validator feedback (description must start with imperative verb)

Changed

• evals/evals.json — added third eval (route-out to caveman when request is communication brevity, not code volume)

---

[1.0.0] — 2025-05-30

Added

• Initial SKILL.md with YAGNI ladder, never-cut safety list (ponytail: marker convention), four intensity levels (lite / full / ultra / off)
• references/commands.md — per-runtime command and install crosswalk
• scripts/install.sh — auto-detect runtime, print (or with APPLY=1 run) the right install command
• evals/evals.json — two canonical evals: native date picker vs. flatpickr, delete-list review of LRUCache + hand-rolled email regex
• Upstream credit: DietrichGebert/ponytail (MIT); catalog: akillness/oh-my-skills

---

[0.2.0] — 2025-04-14

Added

• Compatibility note routing caveman (communication compression) and code-refactoring (behavior-preserving cleanup) as explicit out-of-scope skills
• SKILL.toon compact rendering for token-limited runtimes

Changed

• Description rewritten to be trigger-oriented; now activates on: ponytail, write less code, YAGNI, anti-bloat, minimal code, do I need this, lazy senior dev

🐴 ponytail

"He says nothing. He writes one line. It works."

The laziest senior dev in the room as an agent skill. Before writing anything, ponytail walks the YAGNI ladder and stops at the first rung that holds:

1. Does this need to exist? → no: skip it (YAGNI) 2. Stdlib does it? → use the stdlib 3. Native platform feature? → use the platform (e.g. <input type="date">) 4. Already-installed dep? → use what's installed 5. One line? → write one line 6. Only then → the minimum that works

Every shortcut taken gets a ponytail: comment naming its upgrade path so "later" never becomes "never."

---

Demo

!ponytail in action

Terminal walkthrough: YAGNI ladder → before/after date-picker example → install commands

---

Install

Quick (any runtime)

bash npx skills add https://github.com/akillness/oh-my-skills --skill ponytail

global, non-interactive:

npx skills add https://github.com/akillness/oh-my-skills --skill ponytail -g -y

Or run the bundled auto-detect helper:

bash

prints the right install command; set APPLY=1 to actually run it

bash .agent-skills/ponytail/scripts/install.sh

force a specific runtime and apply immediately

RUNTIME=claude APPLY=1 bash .agent-skills/ponytail/scripts/install.sh

---

Per-runtime (upstream plugin)

Source: <https://github.com/DietrichGebert/ponytail> (MIT)

---

Claude Desktop (MCP server)

Add ponytail's rules via the MCP server entry in ~/Library/Application Support/Claude/claude_desktop_config.json:

{ "mcpServers": { "ponytail": { "command": "npx", "args": [ "-y", "skills", "serve", "https://github.com/akillness/oh-my-skills", "--skill", "ponytail" ] } } }

Restart Claude Desktop after saving. The skill activates on triggers: ponytail, write less code, YAGNI, anti-bloat, minimal code.

---

Instruction-only adapters (always-on, no commands needed)

Copy the matching rules file from the upstream repo into your project:

Agent portability map: <https://github.com/DietrichGebert/ponytail/blob/main/docs/agent-portability.md>

---

Commands

Available in skill-capable hosts (Claude Code, Codex, Gemini CLI, Pi, OpenCode). In Codex, prefix commands with @ instead of /.

Intensity levels

---

Never-cut list

Laziness is about effort, not correctness. These are never removed at any intensity:

• Trust-boundary validation — user input, network data, file reads are still validated.
• Data-loss handling — failures that can corrupt data are still handled.
• Security — authn/authz, secrets, injection defences stay.
• Accessibility — semantic, accessible output stays.

If cutting code would weaken one of these, that code is not bloat.

---

The ponytail: marker

Every shortcut gets a one-line comment naming its upgrade path:

js // ponytail: in-memory map — swap for Redis when this needs to survive a restart const seen = new Map()

html <!-- ponytail: browser has one --> <input type="date">

Run /ponytail-debt to harvest all markers into a tech-debt ledger.

---

References

• `references/commands.md` — per-runtime command & install crosswalk
• `scripts/install.sh` — auto-detect + apply installer
• `evals/evals.json` — canonical evals
• Upstream (MIT): <https://github.com/DietrichGebert/ponytail>
• Benchmarks (80–94% less code, 3–6× faster): <https://github.com/DietrichGebert/ponytail/tree/main/benchmarks>
• Related: `../caveman/SKILL.md` · `../code-refactoring/SKILL.md` · `../code-review/SKILL.md`

---

Changelog (latest 5)

See `CHANGELOG.md` for the full history.

[1.2.0] — 2026-06-22

• Added metadata block to SKILL.md (category, tags, platforms, keyword, version, source, upstream)
• Expanded trigger list; added ai-slop-cleaner route-out to compatibility block

[1.1.0] — 2025-07-10

• Added README.md with full per-runtime install crosswalk and MCP server setup
• Added demo GIF (assets/ponytail-demo.gif) — animated YAGNI ladder walkthrough
• Added CHANGELOG.md

[1.0.1] — 2025-06-18

• Fixed install.sh early-exit on macOS when optional CLIs are absent
• Tightened SKILL.md trigger phrasing per skill-standardization validator
• Added third eval: route-out to caveman for communication-brevity requests

[1.0.0] — 2025-05-30

• Initial release: SKILL.md, references/commands.md, scripts/install.sh, evals/evals.json
• Four intensity levels (lite / full / ultra / off), YAGNI ladder, ponytail: marker convention

[0.2.0] — 2025-04-14

• Added SKILL.toon compact rendering for token-limited runtimes
• Route-outs to caveman and code-refactoring made explicit in SKILL.md

---

_{Upstream: <a href="https://github.com/DietrichGebert/ponytail">DietrichGebert/ponytail</a> (MIT) · Catalog: <a href="https://github.com/akillness/oh-my-skills">akillness/oh-my-skills</a> · "Done. He'd be proud. He won't say it."}

Ponytail — per-runtime command & install crosswalk

Ponytail is one ruleset (the YAGNI ladder + never-cut safety list) plus a small command surface. Skill-capable hosts get the /ponytail* commands; instruction- only adapters load the always-on ruleset without commands.

Commands

In Codex the commands are skills — invoke with @ (@ponytail-review).

Install (this catalog)

Plugin-installable from the jeo-skills catalog:

npx skills add https://github.com/akillness/jeo-skills --skill ponytail
# global, non-interactive:
npx skills add https://github.com/akillness/jeo-skills --skill ponytail -g -y

Or run the bundled helper, which wires the upstream marketplace plugin for the detected runtime:

bash scripts/install.sh

Install (upstream, by runtime)

Source: https://github.com/DietrichGebert/ponytail (MIT).

Instruction-only adapters (always-on ruleset, no commands)

Copy the matching rules file from the upstream repo:

Agent portability map: https://github.com/DietrichGebert/ponytail/blob/main/docs/agent-portability.md

The ladder (reference)

1. Does this need to exist?   → no: skip it           (YAGNI)
2. Stdlib does it?            → use the stdlib
3. Native platform feature?   → use the platform
4. Already-installed dep?      → use what's installed
5. One line?                  → write one line
6. Only then                  → the minimum that works

Never cut for line count: trust-boundary validation, data-loss handling, security, accessibility.