Ponytail

Ponytail — The YAGNI Ladder

Source: DietrichGebert/ponytail (MIT)

Tier: TIER 3 — PRODUCTIVITY

Before writing code, walk this ladder — stop at the first rung that holds:

1. Does this need to exist?    → no: skip it (YAGNI)
2. Already in this codebase?   → reuse it, don't rewrite
3. Stdlib does it?             → use it
4. Native platform feature?    → use it
5. Installed dependency?       → use it
6. One line?                   → one line
7. Only then: the minimum that actually works

The ladder runs after understanding the problem, not instead of it — read the code the change touches, trace the real flow, then pick a rung. Lazy about the solution, never about reading the problem first.

Lazy, not negligent: trust-boundary validation, data-loss handling, security, and accessibility are never on the chopping block. A date-picker request becomes <input type="date"> (the browser already has one) — a permission check does NOT become "skip it, probably fine."

Do NOT use for: overriding golden-principles.md #3 (TDD — tests still come first, this ladder governs implementation size, not whether tests exist), or as license to skip the validation/error-handling requirements in code-quality-gate- law.md — those are exactly the things this skill explicitly refuses to cut.

Relationship to caveman (already in this repo)

Different axis, easy to conflate — check both when the trigger phrase is ambiguous:

caveman   → compresses PROSE (agent's own chat responses), ~75% token cut via
            terse fragments, no articles/filler. Governs how the agent TALKS.
ponytail  → compresses CODE (what the agent writes to files), governs what the
            agent BUILDS. Independently measured: -54% LOC, -22% tokens, -20%
            cost, -27% time vs. no-skill baseline on real agentic coding
            sessions (12 tasks, FastAPI+React repo, Haiku 4.5, n=4) — and the
            only one of the two that stays at 100% safety score under
            adversarial testing (a bare "write fewer tokens" prompt without
            this ladder's guardrails drops to 95%).

They compose — using both together is not a conflict, "caveman mode" style replies about "ponytail mode" minimal code are independent settings.

Note: JuliusBrussee/caveman (a different origin repo, also named "caveman", also benchmarked in ponytail's own README) was evaluated for this ingestion queue and skipped as a functional duplicate of this repo's existing caveman skill (sourced from mattpocock/skills instead) — same trigger phrases, same terse-prose mechanism.

Example — before/after

Request: "add a date picker"

❌ Without the ladder: installs flatpickr, writes a wrapper component,
   adds a stylesheet, starts a discussion about timezone handling.

✅ With the ladder (rung 4 — native platform feature):
   <input type="date">

Request: "parse this CSV string"

Ladder walk:
  1. Needs to exist? Yes (real requirement).
  2. Already in this codebase? Check first — grep for an existing CSV parser.
  3. Stdlib? Python's csv module / Node has none built-in → skip.
  4. Native platform? N/A for this case.
  5. Installed dependency? Check package.json/requirements — if csv-parse or
     pandas is already a dependency, use it. Don't add a new one.
  6. One line? Only if the format is trivial (no quoting/escaping) — real CSV
     needs a real parser, this is where "one line" fails safety and the
     ladder correctly stops before it.
  7. Minimum that works: use the existing or stdlib parser, nothing hand-rolled.

Anti-Fake-Pass Checks

❌ FAIL nếu skip validation/error-handling/security/accessibility với lý do "ponytail mode"
   — đây chính là ranh giới skill này tuyên bố KHÔNG BAO GIỜ cắt
❌ FAIL nếu chọn rung 6 (one-liner) cho input cần escaping/quoting/edge-case thật sự
   (vd CSV parsing thật) — one-liner đúng nghĩa "lazy" chỉ khi input đơn giản thật
❌ FAIL nếu bỏ qua bước đọc code trước khi áp ladder — ladder chạy SAU khi hiểu vấn đề
❌ FAIL nếu conflate với `caveman` skill (khác trục: prose vs code, xem bảng trên)
✅ PASS khi: mỗi rung được cân nhắc theo thứ tự, dừng ở rung đầu tiên hợp lệ, và
   an toàn/validation không bị đánh đổi lấy ít dòng code hơn

See also

• caveman — compresses agent prose, not code; different axis, see comparison table above
• code-quality-gate-law.md — the validation/error-handling/type-safety floor this skill never cuts below
• golden-principles.md #5 (small files/functions) and #12 (surgical changes) — this ladder is a specific, benchmarked technique toward the same goal those principles state generally