context-engineering（情境工程）

Context Engineering Skill

Modern LLMs bill you for every token you send. Large, unstructured prompts lead to higher cost and worse reasoning (lost-in-the-middle). This skill defines practical Context Engineering techniques and shows how to apply them in your current setup, regardless of which editor or AI product a user has.

0. Non‑Negotiable Pre‑Actions

Before any user applies these techniques, always do the following:

1. Define the objective in one sentence. Example: “Add input validation to app.py and factor shared logic into utils/transform.py.”
2. Limit the scope. Only include the files that matter for the current change, not the whole repo.
3. Wrap context with hard boundaries. Use the XML tags in this skill (<global_rules>, <file>, <task>) so any LLM knows what is rules, what is code, and what is the task.
4. State negative constraints. Always end prompts with instructions like “No preamble. No explanations. Output code or SEARCH/REPLACE blocks only.”
5. Test on a small example first. Try the pattern on one small function or file before applying it to a large refactor. If the model output is noisy, tighten constraints and retry.

These pre‑actions are model‑ and tool‑agnostic and should be followed by every user to make the skill reliable.

1. Skeleton-of-Thought (SoT) Prompting

Goal: Give the model the *shape* of your code instead of full implementations.

How it looks (blueprint instead of full file):

def process_data(input_str: str) -> dict:
    """Parses raw string into a structured dict."""
    ...

class ReportBuilder:
    """Builds JSON reports from normalized data."""

    def build(self, data: dict) -> str:
        """Return JSON summary."""
        ...

Prompt pattern:

I am working on this project. Here is the skeleton:
<file name="pipeline.py">
def process_data(input_str: str) -> dict:
    """Parses raw string into a structured dict."""
    ...
</file>

Please implement process_data to parse JSON, handle errors, and return a dict.

Execution steps:

• Strip function bodies (keep signatures, docstrings, and class layout).
• Send only the skeleton when you want new logic or refactors.

2. Token-Efficient Diff Updates

Goal: Stop paying for full-file rewrites; request focused changes.

Search/replace format:

Do not rewrite the whole file.
Respond only with SEARCH/REPLACE blocks in this format:

SEARCH:
old_code_block
REPLACE:
new_code_block

Example usage:

Here is the current code in handler.py:
<file name="handler.py">
def handle(event):
    user = get_user(event["id"])
    # TODO: add logging and error handling
    return user
</file>

Update it to add logging and catch KeyError. Use SEARCH/REPLACE blocks only.

LLMs then return minimal patches instead of re-typing the file.

3. XML Tagging for Modular Context

Goal: Create hard boundaries between rules, files, and tasks so the model does not mix contexts.

Pattern:

<global_rules>
  - Follow PEP8.
  - Use snake_case for functions.
</global_rules>

<file name="utils.py">
def slugify(name: str) -> str:
    ...
</file>

<task>
Add a logging decorator to functions in utils.py and apply it to slugify.
</task>

Why this works: LLMs are trained to respect XML-like tags as structure. Wrapping each file or rule block in tags keeps the model from blending unrelated parts.

4. Symbolic Variables (Reusable Rules)

Goal: Define long rule sets once, then reference them by handle.

Header section:

[RULE_STRICT_TYPES]:
  - Always use Python type hints.
  - Use Pydantic models for external I/O.

[RULE_STYLE]:
  - Use snake_case for functions and variables.
  - Add a one-line docstring to public functions.

Later in the prompt:

Implement a user login function. Apply [RULE_STRICT_TYPES] and [RULE_STYLE].

This keeps the task-specific part short and readable while still enforcing strong constraints.

5. Negative Constraints (Token Filter)

Goal: Remove unneeded prose; pay only for code or the minimal output you want.

Common constraints:

Constraints:
- No preamble.
- Do not explain the code.
- Output code only.

Prompt example:

Refactor the following function to use early returns and clearer naming.
Constraints:
- No preamble.
- No explanation.
- Output code only.

6. LLMBundle and Context Packages

LLMBundle (e.g. Context-Engineering/llmbundle) acts as a context manager for multi-file projects.

Desired behavior:

• Wrap each file in <file name="...">...</file> tags.
• Include a header with <global_rules> and symbolic rule blocks.
• Omit noise files like node_modules/, build artifacts, and lockfiles.
• Optionally add a table of contents at the top:

<bundle_index>
  - app.py
  - utils/transform.py
  - tests/test_transform.py
</bundle_index>

With LLMBundle, your prompt to an AI agent becomes:

<global_rules>
  - Follow PEP8.
  - Prefer pure functions when possible.
</global_rules>

<bundle_index>
  - app.py
  - utils/transform.py
</bundle_index>

<file name="app.py">
  ...code...
</file>

<file name="utils/transform.py">
  ...code...
</file>

<task>
Add input validation to app.py and factor shared logic into utils/transform.py.
</task>

Using llmbundle (optional helper)

If you have the external tool llmbundle installed:

• Run llmbundle on the files or directories you want to share (see its README for exact commands and security guidance).
• Open the generated script (for example project.sh) and paste its contents into your LLM conversation.
• Treat the outer bash structure as a container and only modify file contents between the heredoc EOF markers, as described in the llmbundle header comments.

This skill does not require llmbundle and never executes it. It only describes how to use the tool if the user has chosen to install and verify it independently.

---

## 7. How To Apply This In Your Setup

1. **For new features:** Send a SoT skeleton of relevant files instead of full bodies.\n2. **For edits:** Ask for SEARCH/REPLACE style diffs, not entire files.\n3. **For multi-file work:** Wrap rules, files, and tasks in XML tags.\n4. **For repeated rules:** Define symbolic variables like [RULE_*] once at the top and reference them later.\n5. **For cost control:** Always end prompts with negative constraints (no preamble, no explanations) when you only need code or diffs.\n6. **With LLMBundle:** Use it to automatically package your repo into XML-tagged, token-efficient context before sending it to an AI agent.