Prompt Engineering

An LLM without a prompt is a chatbot. A prompt gives it direction, purpose, and constraints. For agents, the prompt is the operating system. It tells the agent what problem to solve, which tools to reach for, how to interpret results, and what counts as done.

A poor prompt leads to hallucinated tool calls, incorrect data summaries, and noise in the output. "Audit a business" without guidance might result in the agent calling every endpoint in the catalog, drowning the user in raw data. A well-engineered prompt says: "Audit NAP consistency across 5 directories, identify critical mismatches, prioritize fixes, and suggest a 30-day correction plan." The agent now knows which tools to call, what patterns matter, and what to emphasize in the response.

Prompt engineering is especially critical in agent workflows because the cost of a bad prompt compounds. Each incorrect tool call wastes credits. Each misdirected reasoning step wastes time. Each vague output requires a follow-up conversation. Good prompts reduce friction and improve both accuracy and speed.

Effective prompts for local SEO agents follow a consistent structure:

Role and context — Define the agent's perspective. "You are a local SEO analyst auditing a multi-location plumbing business."

Task and success criteria — Be explicit about what success looks like. "Identify the top 3 ranking barriers across all locations. Prioritize by impact on visibility. Provide concrete fixes."

Tool guidance — Hint at which tools are relevant without over-constraining. "You have access to ranking, review, citation, and keyword tools. Use them to build a complete picture."

Output format — Define the shape of the response. "Structure your findings as: Problem, Root Cause, Recommended Fix, Expected Impact. Include specific data points (e.g., rank position, review count, citation gaps)."

Constraints and style — Set boundaries. "Surface only actionable findings, not curiosities. Use specific numbers. Avoid generic advice." For agents, style matters because the output flows directly to clients or dashboards.

A local SEO agent prompt might look like: "Audit [business] in [market]. Check rankings, NAP consistency, review volume, and keyword gaps. Summarize the top 2 issues and recommend fixes. Format as a bullet list with specific data points. Omit endpoints that return no new insights."

Vagueness is the most common mistake. "Analyze the business" leaves the agent guessing about scope. Does that mean rankings? Reviews? Citations? All three? The agent might call too many tools or miss critical ones. Instead, be specific: "Analyze rankings and citation consistency."

Over-specification is the opposite error. If you script every tool call, the agent becomes a machine running your commands, not a reasoning system. It loses the ability to adapt when data is incomplete or when a different tool would be more efficient. Instead of "Call the local-pack tool, then the review-velocity tool, then the citation-audit tool," say "Use ranking and reputation tools to build a competitive analysis."

Missing success criteria leads to rambling outputs. "Tell me about the business" produces a data dump. "Identify the one ranking issue holding back visibility" produces focused findings. Agents reason better when they know what matters.

Ignoring token budget causes truncated responses. A prompt requesting "all possible insights" from a 40-tool catalog can exceed context limits. Constrain scope: "Surface the top 3 issues only."

Neglecting tool context wastes calls. If the agent doesn't understand that geogrid costs 50 credits while local-pack costs 1, it might choose the expensive option when the cheap one would suffice. Hint at relative costs and tool trade-offs.

Prompt engineering is not a one-time task. After running a prompt against real workflows, review the outputs. Did the agent choose the right tools? Did it prioritize findings correctly? Was the output consumable or bloated?

Common iterations:

Tool selection issues — If the agent called unnecessary endpoints or missed critical ones, adjust the tool guidance in the prompt. "Use ranking and review tools (not keyword tools) for this audit."

Output noise — If the summary included interesting but not actionable findings, tighten the constraints. "Only include findings that directly impact ranking or visibility."

Reasoning gaps — If the agent made logical leaps or incorrect inferences, add reasoning steps. "Explain how citation inconsistencies affect local pack ranking before recommending fixes."

Token overflow — If responses were truncated, reduce scope. "Audit the top 3 locations only, not all 15."

Speed or cost issues — If the workflow ran too many expensive tool calls, hint at prioritization. "Start with low-cost checks (local-pack, NAP audit). Only escalate to geogrid if ranking is unexpectedly low."

Good prompt engineering is collaborative between human and machine. The human defines success; the agent discovers the path. Iterate until the path is reliable, the findings are relevant, and the output is actionable.