← 返回 Skills 市场
106
总下载
0
收藏
1
当前安装
5
版本数
在 OpenClaw 中安装
/install capture-classification
功能描述
Standard Operating Procedure (SOP) that routes unstructured text to Tasks or LanceDB based on urgency using atomic nodes.
使用说明 (SKILL.md)
Lean Philosophy (Principles)
- Kaizen (改善): This workflow relies entirely on discrete, single-responsibility atomic nodes rather than a monolithic loop.
- Standardized Work (Hyojun Sagyo): This node represents a strict, step-by-step Standard Operating Procedure (SOP) for inbound text classification.
- Jidoka (自働化): Includes autonomous self-healing loops with hard verification stops between every step.
Capture Classification SOP
This procedure guides the agent to act as a semantic router for inbound text using explicitly defined atomic nodes.
Cognitive Directives
WHEN [Unstructured audio transcript or note text is captured] THEN [ Follow this strict Standard Operating Procedure:
Step 1: Classification
- Execute the
LLM-Classify-Intentatomic skill with categories: ["Actionable", "Informational"]. - Jidoka Stop: Verify the skill returns exactly one of the requested categories. IF it fails, instruct it to correct the output and retry. Do NOT proceed until a valid category is obtained.
Step 2: Routing Execution
- IF "Actionable":
- Execute the
Google Tasks Create Taskatomic node using the text as the title. - Jidoka Stop: Check if the node returns a successful JSON response. IF it fails, retry up to 3 times. IF it still fails, report the error and STOP.
- Execute the
- IF "Informational":
- Execute the
Vector Store Upsert Memoryatomic node to save it as reference material. - Jidoka Stop: Verify the vector store confirms a successful upsert. IF it fails, retry up to 3 times. IF it still fails, report the error and STOP. ]
- Execute the
Expected Output
A JSON log confirming the routed destination and action taken by the respective atomic node.
安全使用建议
This skill's SOP is coherent at a high level, but several practical and security details are missing. Before installing or enabling it, ask the publisher (or check the platform) for: (1) which atomic nodes will be invoked exactly and what permissions/scopes those nodes require (Google Tasks OAuth scopes, API keys, LanceDB host/credentials); (2) why the 'gog' binary is declared as required and whether it's actually used; (3) where the vector store endpoint is configured and who controls it (to avoid writing sensitive text to an unknown external DB). If you don't get clear answers, do not enable the skill with broad agent credentials: at minimum, restrict the atomic nodes' permissions to only the needed Google Tasks scope, and verify the vector store endpoint is one you control. Also prefer a manifest that declares required env vars or config paths explicitly so you can review and consent to them.
功能分析
Type: OpenClaw Skill
Name: capture-classification
Version: 1.0.4
The skill bundle defines a legitimate Standard Operating Procedure (SOP) for an AI agent to classify unstructured text and route it to either Google Tasks or a vector store (LanceDB). The instructions in SKILL.md are transparent, focusing on workflow automation and error handling (Jidoka) without any evidence of malicious intent, data exfiltration, or unauthorized command execution.
能力评估
Purpose & Capability
The skill claims to route inbound text to Google Tasks or a vector store (LanceDB). The SKILL.md references atomic nodes (LLM-Classify-Intent, Google Tasks Create Task, Vector Store Upsert Memory) which is consistent with the stated purpose — but the declared requirements do not: the manifest lists no credentials or config for Google or any vector DB, and it requires a binary named 'gog' which is unrelated to the described functionality. Either credentials/connection info are missing from the manifest, or the skill expects implicit access via atomic nodes. This is an incoherence.
Instruction Scope
The instructions direct the agent to invoke external atomic nodes that will call Google Tasks and a vector store and to retry on failures. The SKILL.md does not specify which Google account, scopes, or vector store endpoint to use, nor does it tell the agent to obtain user consent or to store tokens. The SOP also mentions LanceDB in the description but the runtime text only says 'Vector Store Upsert Memory' without identifying the vendor/endpoint. Because the skill will cause network actions (through atomic nodes) without declaring where credentials come from, the instruction scope is under-specified and potentially risky.
Install Mechanism
No install spec and no code files (instruction-only), which minimizes on-disk risk. However the declared required binary 'gog' is unexpected for a purely orchestration SOP and there is no explanation of why it's required; that's a mismatch to note but not an installation-action risk because nothing will be written to disk by the skill itself.
Credentials
The skill requires zero environment variables or credentials, yet it instructs calls to Google Tasks and a vector store (LanceDB implied). Those integrations normally require API keys or OAuth credentials. The absence of declared credentials is disproportionate: either the platform is expected to supply access via pre-configured atomic nodes (not documented), or the manifest is incomplete. This mismatch could lead to silent failures or unexpected use of existing agent credentials.
Persistence & Privilege
The skill is not marked 'always' and uses the default autonomous invocation flag, which is normal. It does not request any special persistent presence or claim to modify other skills or system-wide settings.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install capture-classification - 安装完成后，直接呼叫该 Skill 的名称或使用
/capture-classification触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v1.0.4
- Refactored the workflow description to emphasize a strict Standard Operating Procedure (SOP) using explicit atomic nodes.
- Updated the Lean principles section to clarify reliance on discrete, single-responsibility nodes instead of loops.
- Clarified each processing step with separate “Classification” and “Routing Execution” using named atomic nodes for better traceability and verification.
- Enhanced “Jidoka Stop” verification instructions and retry conditions for both actionable and informational routing steps.
- Updated language throughout to provide clearer, more actionable step-by-step guidance.
v1.0.3
- Added explicit verification (Jidoka) steps to ensure sub-agent and system actions return valid, expected results.
- Introduced error handling and retry logic for both task-adding and upserting operations.
- Updated workflow to include corrections on invalid category output and successive retries on failure.
- No changes to input/output formats; logic and reliability enhancements only.
v1.0.2
- Added "gog" to required binaries for task creation.
- Updated actionable routing to use the native command: gog tasks add @default --title "...".
v1.0.1
- Broadened OS compatibility from Windows-only to all operating systems in metadata.
- No functional or workflow logic changes.
v1.0.0
- Initial release of capture-classification.
- Routes unstructured audio transcripts or text notes based on intent and urgency.
- Actionable items are added to Google Tasks; informational notes are saved in LanceDB.
- Outputs a JSON log confirming the action and destination for each entry.
元数据
常见问题
Capture Classification 是什么？
Standard Operating Procedure (SOP) that routes unstructured text to Tasks or LanceDB based on urgency using atomic nodes. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 106 次。
如何安装 Capture Classification？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install capture-classification」即可一键安装，无需额外配置。
Capture Classification 是免费的吗？
是的，Capture Classification 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Capture Classification 支持哪些平台？
Capture Classification 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Capture Classification？
由 zvirb（@zvirb）开发并维护，当前版本 v1.0.4。
推荐 Skills