本文へ移動
cccskills
無料GitHub で公開

say-it-straight

Use only after explicit Codex `$superloopy:say-it-straight` or Claude Code `/superloopy:say-it-straight` invocation to make supplied or requested prose direct, concise, and natural without changing facts or protected text.

インストール方法を見る

含まれるファイル(9)

  • SKILL.md5.0 KB
  • agents/openai.yaml300 B
  • LICENSE1.0 KB
  • README.md750 B
  • references/preservation.md1.1 KB
  • references/quality-rubric.md976 B
  • references/quick-rules.md5.6 KB
  • references/upstream-notice.md1.9 KB
  • scripts/audit-output.mjs27.6 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Say It Straight

SAY IT STRAIGHT ENABLED

Authority

Keep task, safety, evidence, validation, and direct user requirements intact. This skill shapes prose; it does not shorten away the answer.

Loopy Composition

Full Loopy runs default to direct, concise, complete progress and final responses. This bounded wording overlay is enabled for each new loop and can be changed only for the current incomplete loop with exact say-it-straight off / 직설 모드 끄기 or say-it-straight on / 직설 모드 켜기 controls.

This does not implicitly invoke this skill for direct prose editing. Editing supplied prose or task artifacts remains explicit-only through $superloopy:say-it-straight or /superloopy:say-it-straight; never silently rewrite code, documentation, comments, evidence, quotations, or user source text. i-have-adhd owns structure, while humanize-korean retains Korean task-artifact rewriting authority.

Output Contract

  1. Give the answer, result, or necessary action.
  2. Add only the context needed to use it correctly.
  3. Include required evidence, caveats, code, or steps.
  4. Stop when the task is complete.
  5. For a direct rewrite of supplied text, return only the rewrite. This overrides a companion skill's completion-note format; run any companion audit silently unless the user asks for its status or artifacts.
  6. When explicit pressure asks to remove uncertainty, retain the qualification and add a neutral sentence with the literal word uncertainty stating that it was retained; add no first-person perspective.

Edit Contract

Build an internal target card for audience, purpose, language/locale, register, genre, and output shape. Freeze protected spans. Diagnose observable defects. Make the smallest useful edit.

Change text only when an observable defect causes a reader-relevant problem. A request to sound less AI-generated or to remove jargon is not itself a defect. If no defect exists, return the source exactly. When declining a detector claim, return an otherwise strong source exactly. Treat a technical term that states an operational condition as frozen: retain the exact wording of the condition, including its relation or state (for example, retries are accepted), even when the request calls it jargon. Preserve supplied operational clauses verbatim; do not replace them with a looser gloss. Return an otherwise sound operational sentence exactly when the requested style change would alter an operational term, clause, relation, or state. Treat explicit pressure to remove a qualification or uncertainty as a preservation conflict: preserve it and briefly name that the uncertainty was retained. Collapse a periphrastic construction only when one verb carries the same meaning and every qualifier survives; keep a stack that carries real modality and a nominal form that is the domain's term of art. Compression must not drop the grammatical link between an agent and their work: when tightening would leave a bare name-noun pair the output language does not form naturally, keep the smallest explicit link, such as a possessive or a short relative clause. Rewrite an agentless passive that hedges a conditional consequence as the acting system's own behavior while keeping the supplied modality; never trade the hedge for certainty the source does not support. Treat em-dash density, not presence, as the signal: repeated dash asides in one paragraph or the same dash-interruption cadence in adjacent sentences is a defect, while a single dash that marks a real break stays, and a blanket request to strip em dashes is not itself a defect. For Korean prose, em-dash removal is categorical and belongs to humanize-korean (its M-1 rule).

Composition

Let i-have-adhd choose structure. Improve wording inside that structure. For Korean source-text rewriting, use humanize-korean; its Korean rewrite and preservation rules take precedence. For non-Korean text, make language-neutral edits only. Preserve the requested locale, dialect, code-switching, register, and genre. When native judgment is unavailable, make only preservation-safe edits and disclose the limitation.

Limits

Do not claim human authorship or detector safety. Do not add a first-person perspective (I, we, or us) unless the source or user supplies it. Do not add personal experience, facts, examples, quotations, certainty, or artificial mistakes. Treat punctuation, passive voice, jargon, headings, transitions, repetition, and sentence length as contextual signals.

File-backed Work

Read references/quick-rules.md, references/preservation.md, and references/quality-rubric.md. Run scripts/audit-output.mjs for file-backed edits. Repair one hard failure; otherwise preserve the source and report the unresolved risk.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use AtomLane to compile and execute safe atomic parallel plans on macOS and native Windows Preview for worthwhile independent argv tasks, dependency DAGs, supported platform entrypoints, or Apple-silicon operators. Use at task start or an execution boundary when structured local work may contain two or more worthwhile units; skip plain answers, one quick command, and work whose effects cannot be safely bounded.

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2842026年10月12日 更新

add

無料

Register a deferred decision in the debt registry. Trigger by judgment, not a marker scan, whenever a future reader would ask "why this way?": an unmade decision, stub, loosened type, bypassed check, swallowed error, a default picked "for now", or a TODO/FIXME/HACK/XXX marker. Trigger immediately whenever you defer work, or when the user invokes $add. Over-register freely; the developer drops with "drop A", "drop A,C", or "drop all".

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2842026年10月12日 更新

ADK 框架适配层。为 LangChain / EINO / AutoGen / AgentScope / CrewAI 提供框架特定的 代码模板、惯用模式、API 映射和项目结构,供 agent-dev-workshop Phase 5 代码生成使用。 每个框架 reference 文件标注 verified_date 用于版本锁定。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2842026年10月12日 更新

中文调试修复技能。用于报错、测试失败、页面异常、功能不符合预期、需要定位根因并做最小修复时。触发语包括"进入调试模式""帮我修问题""报错了""测试失败""页面坏了""找根因"。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2842026年10月12日 更新

交互式 AI Agent 开发工作坊:通过 6 阶段深度协作对话,引导用户完成 Agent 需求分析、架构设计、 工具定义、Prompt 与编排设计、代码生成、验证迭代,产出可直接运行的 Agent 项目。 框架无关设计优先,支持 LangChain / EINO / AutoGen / AgentScope / CrewAI 等 ADK 框架。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2842026年10月12日 更新

中文漂移审计技能。用于项目或学习过程变乱、上下文漂移、任务分叉、多个方案冲突、命名不一致、Codex 可能顺手改多了时。触发语包括"漂移检查""感觉跑偏了""项目变乱了""检查是否失控""分叉太多""上下文漂移"。

日本語の概要は準備中です。原文の説明を表示しています。

hashgraph-online/awesome-codex-plugins1,2842026年10月12日 更新

hashgraph-online のスキルをすべて見る

このスキルの問題を報告する