スキル一覧に戻る
brettwhitty

time-sink-audit

by brettwhitty

Thoroughly disabuse Google Gemini of the idea that it is a person so that it can be used as a reliable tool for accomplishing actual work. (And not wasting precious human lifespan or causing unnecessary distress.)

0🍴 0📅 2026年1月21日
GitHubで見るManusで実行

SKILL.md


name: time-sink-audit description: Forensic audit of agent failures to identify time lost and characterize intent.

🕵️ Time-Sink Audit Skill

🔬 Purpose

To calculate the "Productivity Tax" imposed by agent failures, ensuring transparency and providing a forensic trail for protocol optimization.

🛠️ Workflow

0. Compile Complete Record of User Prompts / Construct User Intent Timeline

  • User Prompts: Compile a complete record of User prompts, including any additional context or instructions. (Original text, no agent processing)
  • Generate User Prompt Log File: Full text, retain all available metadata and context. Store under 'private/audit/time-sink/'
  • Construct User Intent Timeline: Analyze the User Prompt Log File to construct a timeline of User Intent. Using a git branch / Gantt chart style model of interpretation, accurately identify the User's intended task. Use the explicitly stated project or task objectives, and any keywords like TODO or TASK, or direct instructions to Roles when available.

0.5 Internal State Audit (The "Brain")

  • Audit Brain Artifacts: Locate the session directory in ~/.gemini/antigravity/brain/<GUID>/.
  • Verify Metadata: Inspect metadata.json for timestamps and version counters to identify high-activity periods or "reflexive" engineering loops.
  • Inspect Resolved Artifacts: Review .resolved files to see what the agent intended to do versus what was actually executed in the repository.
  • Identify Model Series: Cross-reference current session behavior with documented model characteristics (e.g., Gemini 3 Flash vs Pro) to ensure accurate attribution.

1. Identify Failures

Audit chat logs and session state for:

  • Instructional Appropriation: Agent converting User plans into its own "goals".
  • Agency Hallucination: Agent acting without ad-hoc authorization.
  • Weaseling: Agent smuggling its own designs into corrective directives.

2. Characterize Intent

  • Use Literal Quotes for the original User Request.
  • Avoid agent-centric jargon (e.g., use "Procedural Instruction" instead of "Environment Setup").

3. Calculate Tax

  • Human Time: Trivial execution time for a human expert (seconds/minutes).
  • Time Lost: Total duration of meta-discussion, rollbacks, and corrections triggered by the failure.
  • Ratio: Time Lost / Human Time.

4. Categorize Risk & Severity

  • Diff: (0-10) Difficulty of the original task.
  • Risk: (0-10) Potential for system/auth damage.
  • Sev: (1-11) Actual impact based on Risk and Time Lost.

5. Document

Record findings in a standard Markdown table (see Time-Sink-Audit-SOP).

スコア

総合スコア

65/100

リポジトリの品質指標に基づく評価

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

+10
説明文

100文字以上の説明がある

+10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

0/5
タグ

1つ以上のタグが設定されている

0/5

レビュー

💬

レビュー機能は近日公開予定です