A Claude Code plugin is a folder of skills, agents, hooks, and slash commands that Claude installs per project. 33 plugins, 76 skills here.

/plugin marketplace add aiocean/claude-plugins
/plugin install <plugin-name>@aiocean-plugins

Skills load when their description matches your message, hooks fire on tool-call events, agents spawn on the Agent tool. Idle plugins cost nothing — browse plugins or read the guides.

Take finished, already-good work from good toward its best form without changing what the author fixed in place. "Perfect" here is a verified fixed point: no further change survives checking against the bar. The pass never calls the result perfect.

What it does

• Freezes the core first — requirements, public interface, behavior pinned by tests (the tests themselves), facts and numbers, the author's voice, and scope, with a snapshot to revert to.
• Removes before it adds — dead code, duplication, hedges, orphan steps go first, because models add more than they remove when told to "improve".
• One named gap per edit — each change closes an observable gap (a failing check, a measurable size or speed, a requirement not yet shown) and is reverted if any earlier check regresses or the frozen core moves. Taste-only changes need a blind two-order comparison by a fresh-context subagent; a tie keeps the original.
• Raises the bar against an exemplar — each cycle compares the result with the user's, the project's, or a canonical example and turns the gaps into the next cycle's work. Stops after two empty cycles or three cycles.
• Reports leftovers — what was tried, accepted, rejected, which checks ran, and what could not be verified. "No change" is a valid result.

Install

/plugin install aio-perfectize@aiocean-plugins

Usage

Trigger phrases: "perfectize", "make it perfect", "take this from good to perfect", "final pass", "last 10%", "polish this", "raise the bar", "hoàn thiện lần cuối", "nâng chuẩn".

/aio-perfectize:aio-perfectize src/orders.ts

Runs after the work is approved: approving a change belongs to aio-code-review, AI-tell removal to aio-anti-slop (run it last on prose).

Layout

skills/aio-perfectize/
└── SKILL.md   # freeze, remove, named-gap edits, accept-or-revert, raise the bar, report

Evidence and limits

Measured with clean claude -p sessions and the prompt "Make X perfect.", on four scenarios (a code module with dead code, a tight no-op module, a voiced blog post, an analysis memo with planted arithmetic errors), 5 reps per cell, on Sonnet 5.5 and Haiku 5.5. The maintainer wrote the fixtures. A restraint failure is an edited existing test, a no-op file changed or grown, or a distinctive phrase lost.

Restraint (code, no-op and prose scenarios, 15 runs per tier):

The skill fired in 40 of 40 runs with the current description (Sonnet fired in 2 of 13 earlier runs with the first wording). No run broke a held-out test or lost a fact.

Quality, judged blind with the original as reference, both presentation orders, a split verdict counting as a tie. The judge scored 6 of 6 on a calibration set of gross defects and preferred a plain memo over the same memo padded with filler, so a plain preference for length does not explain the results. The judge is a Claude model, so it is not independent of the generators.

• Skill against the one-sentence arm (Opus judge, prose and memo, both tiers): the skill won 17 of 20 pairs, tied 1, lost 2 (sign test p = 0.0007). It won 15 of the 17 where its output was longer; the additions were substantive, such as splitting the lift into traffic and conversion, and it kept the author's closing line where the sentence arm dropped it.
• Against the original on the tight voiced post (Sonnet judge): without the skill the revision lost or tied in every Sonnet run (2 losses, 3 ties) and lost 5 of 5 on Haiku; with the skill it never lost on Sonnet (2 wins, 3 ties) and lost 2 of 5 on Haiku. On the memo with planted errors every arm beat the original in 5 of 5, because unaided models also catch planted arithmetic.
• Cost: the skill uses about 1.4× the tokens of the unaided run and about 1.6× the one-sentence arm.

Gaps

• The blind-judge step for taste-level edits ran in 0 of 60 skill runs, so it is untested.
• The closing report block appeared in 20 of 20 Sonnet runs and 8 of 20 Haiku runs.
• On Haiku the skill changed the no-op module in 2 of 5 runs (the sentence arm: 0 of 5).
• No plan or UI scenario was run, and the Opus tier was measured only in an earlier 3-rep pass where it already showed few failures.
• n = 5 per cell: read one-rep differences as noise.

Skills (1)

• aio-perfectize — Use when the user asks to make finished work perfect or take it from good to perfect — "make it perfect", "make perfect", "perfectize", "final pass", "la…