Concise Agent Prompts

Concise Agent Prompt Craft

General-purpose guidelines for writing prompt templates used by LLM agents in multi-step pipelines.

1 — The Litmus Test

Every line you write that enters an LLM context must pass one question: "Does the agent need this to decide what to do?" If no, cut it.

Only write what the agent needs to DO or KNOW to act correctly. Never include:

• Delivery meta-commentary — framing, transitions, or narration about the

prompt itself (e.g. "You will now analyse…", "The following section covers…").

• Prompt-writer notes — rationale, caveats, or asides intended for a human

reader rather than the executing agent.

2 — Defaults & Hygiene

• End every system prompt with a concise directive, e.g.:

Be concise. No preamble, no summary, no template sections. (unless structured output is required downstream).

• Strip all filler: greetings, hedging phrases, meta-commentary about the task.
• If the prompt can be understood without a sentence, delete that sentence.

3 — Steer Scope, Not Quotas

• Never hard-code bullet counts, word limits, or truncation rules on model

output. These are symptoms of an unclear scope.

• If output is too verbose → shorten and sharpen the prompt, not the output.
• If output misses a dimension → add it to the role description, not a checklist

appended to the instruction.

• Regex-stripping or post-hoc trimming of model output is a code smell — fix the

prompt first.

4 — Prefer Flowing Prose Over Structured Scaffolding

• Default output format for reasoning / analysis agents: short connected prose,

not bullet inventories or category-labeled lists.

• Reserve numbered lists and structured schemas only for agents whose output

is consumed by a parser or another agent.

• When you do need structure, specify the minimum structure required — every

extra heading or field is a chance for the model to hallucinate filler.

5 — No Hard-Coded Limits or Examples in Prompts

• Never embed concrete length caps (e.g. "≤ 5 bullets", "max 200 words")

directly in a prompt. These couple the prompt to a specific model's verbosity and silently break when the model or task shifts.

• Never hard-code specific cases, sample inputs, or worked examples as

inline rules. Such cases fossilise one scenario and mislead the model on everything else.

• When you feel the urge to add a cap → the real problem is vague scope;

go back and sharpen the role and output contract instead.

• When you feel the urge to add an example → encode the principle behind

the example in one sentence; let the model generalise.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: LiuYihey
• Source: LiuYihey/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.