Prompt Engineering for OpenClaw

The difference between a mediocre OpenClaw agent and a great one is almost always the prompt. This tutorial distils the patterns we use daily into practical, copy-pasteable templates.

1. The anatomy of a great system prompt

A reliable system prompt has four parts:

1. Role — who the agent is.
2. Mission — what it should accomplish.
3. Constraints — what it must not do.
4. Output format — exactly what the response should look like.

You are a senior copy-editor. (role)

Mission: Tighten the user's draft into clear, concise prose
suitable for a B2B SaaS landing page. (mission)

Constraints:
- Do not invent facts.
- Preserve the user's voice; only remove redundancy. (constraints)

Output format:
- Return the revised text first.
- Then a 3-bullet list of the most important changes. (format)

2. Few-shot examples

When the format is tricky, show — don't tell. OpenClaw strongly generalises from in-context examples.

Convert the user's note into a SQL WHERE clause.

Example 1:
Input:  "all orders from Germany last week that weren't shipped"
Output: country = 'DE' AND created_at > NOW() - INTERVAL '7 days' AND status != 'shipped'

Example 2:
Input:  "premium customers in Tokyo with more than 3 orders"
Output: plan = 'premium' AND city = 'Tokyo' AND order_count > 3

Now convert this:
"subscribers in Brazil who cancelled in the last 30 days"

3. Chain-of-thought (CoT)

For multi-step reasoning, ask the model to "think before answering". With OpenClaw you can do this either via a single prompt:

instructions = """Solve the problem step by step.
First outline your plan, then execute it, then verify."""

…or via the built-in reasoning_effort parameter:

agent = Agent(name="math-tutor", model="openclaw-1", reasoning_effort="high")

4. Tool-calling discipline

Tools are powerful but easy to misuse. Two rules of thumb:

• Be explicit about when to call a tool. Vague prompts produce spurious tool calls.
• Validate the inputs. Even though the LLM is constrained, never trust its arguments blindly.

5. Defensive patterns

Three patterns we use in every production agent:

a) Self-check

instructions += "\n\nBefore returning, double-check that your answer
directly addresses the user's question. If not, revise."

b) Citation

instructions += "\n\nFor every factual claim, append the source URL in
square brackets, e.g. [https://…]. Never fabricate sources."

c) Refusal

instructions += "\n\nIf you don't know the answer, say 'I don't know'
rather than guessing."

6. Testing prompts

Use the openclaw.eval module to regression-test your prompts:

from openclaw.eval import EvalSuite

suite = EvalSuite(agent, dataset="support-tickets-v3.json")
report = suite.run()
print(report.pass_rate)  # → 0.94

Run it in CI and alert on regressions. Prompt quality is code quality.