You want an agent to prepare a weekly report. You explain the format, remind it not to send anything without review, then discover it cannot read the data. A long prompt may solve this once, but it mixes three needs that are easier to manage separately.
Our practical approach is to keep stable context where the agent regularly reads it, turn a recurring procedure into a skill, and connect a source through a tool when access is needed. Every task does not need all three.
AGENTS.md: the rules of the project
In , AGENTS.md supplies project instructions. It is a suitable place for facts such as where content lives, which command validates it and which actions require approval.
An editorial project might specify that articles have Spanish and English versions and need individual review before publication. These rules affect many assignments. The file does not need the complete writing manual for every article type.
Make instructions observable. Asking for good writing leaves considerable room for interpretation. Asking the agent to preserve numbers in translation and keep sources beside factual claims gives a reviewer something to check. When a rule depends on a tool, explain how to confirm that tool is available.
Instruction filenames and discovery vary between products. A file that reads will not necessarily apply automatically in another application.
A skill: a procedure worth repeating
The Agent Skills specification defines a directory containing SKILL.md, with optional supporting resources. This suits a reusable method such as reviewing an article, preparing a proposal or analyzing an export.
The practical difference from a general rule is the sequence. An editorial review can require reading the source, finding factual claims, checking links and returning edits with unresolved questions. The agent needs that procedure for content reviews, not every conversation.
Try the process manually before creating a skill. Notice which instructions you repeat. Formalizing a method too early may preserve its mistakes. Keep examples that let you evaluate whether the procedure helps.
MCP: access to tools and data
MCP is a protocol for connecting AI applications to external systems. It can expose a tool that reads data. A connection does not replace access permissions or instructions for interpreting the result.
For the weekly report, the connector retrieves results, the skill explains how to compare them, and project instructions retain the approval requirement. If an authorized CSV already contains what you need, a connector may be unnecessary for that task.
More style instructions will not repair a failed data request. Check the connected account, allowed operation and actual response. Keep credentials out of instruction files.
Divide one report sensibly
Consider a shop reviewing orders every Monday. This is a design example, not a client case. Stable context defines currency, timezone and separate treatment of refunds. The procedure compares complete weeks, identifies changes and asks questions where data is missing. The tool retrieves orders and refunds with read-only access.
The output could contain one paragraph on the main change, a small table and missing-data questions. If access fails, the report should say so. It should not quietly replace real figures with estimates.
Test a normal period, a period with no orders and one containing a refund. Check that the explanation matches the data. You are evaluating the process’s result, not whether the agent names all three technologies.
Choose without adding unnecessary complexity
Start with project instructions when a stable rule is missing. Write a skill when you repeat a procedure. Investigate a connection when the agent lacks a source or action. Add each component to solve an identifiable problem.
Before designing the procedure, review what agent skills are.
To apply these distinctions to a development environment, see how to evaluate Google Antigravity.
Our Marketing Skills article shows a specific reusable method. To separate rules, procedure and access in your own business process, tell us how it works today.
