openclaw-knowledge-coach（OpenClaw 知识 coach）

OpenClaw Knowledge Coach

Create a local knowledge workflow in OpenClaw where importing knowledge also produces practice material for retention. OpenPraxis is on PyPI: use pip install openpraxis to get the praxis CLI.

CLI First

Use OpenPraxis CLI as the default execution path.

Install from PyPI (recommended):

pip install openpraxis
praxis --help

Or install from source for development:

git clone https://github.com/Sibo-Zhao/OpenPraxis.git
cd OpenPraxis
pip install -e ".[dev]"
praxis --help

Configure provider/model/API key before ingestion/practice:

praxis llm setup
praxis llm show

Use environment variables when needed (higher priority than config file):

export OPENAI_API_KEY="your_key_here"
# or ARK_API_KEY / MOONSHOT_API_KEY / DEEPSEEK_API_KEY based on provider

Core Workflow

1. Confirm scope and source

• Confirm knowledge domains, source folders, and accepted file types.
• Confirm whether to preserve existing metadata (tags, dates, project names).

1. Define import contract

• Normalize each source into a record with doc_id, title, source_path, tags, created_at, and content.
• Split long content into chunks with stable IDs such as doc_id#chunk-001.

1. Import into OpenClaw

• Ingest normalized records into the local OpenClaw knowledge base.
• Keep a deterministic mapping between source file and imported IDs for later updates.

1. Generate exercises at import time

• For each chunk, create at least one retrieval exercise.
• Prefer three exercise types:

- free-recall: ask the user to explain from memory. - qa: ask direct question-answer pairs. - application: ask scenario-based transfer questions.

• Save answer keys and concise grading rubrics.

1. Build review queue

• Group exercises by topic and difficulty.
• Schedule spaced review windows (for example: day 1, day 3, day 7, day 14).

1. Validate quality

• Reject exercises that can be answered without the imported knowledge.
• Reject ambiguous or duplicate questions.
• Ensure every exercise points back to doc_id and chunk_id.

CLI Command Playbook

Run this sequence when the user asks to import local knowledge and create practice:

1. Add a local file

praxis add "/absolute/path/to/note.md" --type report

1. List recent inputs and capture target input_id

praxis list --limit 20

1. Force-generate a new practice scene for an existing input

praxis practice <input_id>

1. Submit answer by file (preferred for deterministic runs)

praxis answer <scene_id> --file "/absolute/path/to/answer.md"

1. Inspect pipeline results and insight cards

praxis show <input_id>
praxis insight <input_id>

1. Export insights to Markdown/JSON

praxis export --format md --output "/absolute/path/to/insights.md"
praxis export --format json --output "/absolute/path/to/insights.json"

Agent Execution Rules

• Prefer praxis add for import and initial exercise generation.
• Parse IDs from CLI output, then chain praxis practice and praxis answer.
• Use praxis answer --file instead of interactive stdin in automation flows.
• If duplicate content is skipped, rerun with praxis add ... --force when user wants reprocessing.
• Use one-shot runtime model override only when requested:

praxis --provider openai --model gpt-4.1-mini add "/absolute/path/to/note.md"

• For image notes, pass image file path directly to praxis add; OCR extraction is built in.
• Always finish with praxis show plus praxis insight or praxis export so user gets concrete output artifacts.

Output Contract

When executing tasks with this skill, always provide these outputs:

• Import summary: files processed, chunks created, failures.
• Exercise summary: counts by type/topic/difficulty.
• Review plan: next due batches and estimated workload.
• Traceability map: source -> doc_id -> chunk_id -> exercise_id.

Exercise Format

Use this compact JSON-like structure per exercise:

{
  "exercise_id": "ex-...",
  "doc_id": "...",
  "chunk_id": "...",
  "type": "free-recall | qa | application",
  "question": "...",
  "answer_key": "...",
  "rubric": ["point 1", "point 2"],
  "difficulty": "easy | medium | hard",
  "next_review": "YYYY-MM-DD"
}

For more generation patterns, read references/exercise-patterns.md.