Setting up and configuring the agent

Today's topics

• What would you like the agent to know about your work? What would you like it to do consistently?
• Today, write your own instructions and try them on a task you choose.
• Then learn how memory stores past decisions and how Skills describe reusable procedures.

Task and setup

Choose a task you want to work on: writing, data analysis, organizing files, or something else. You can also use last week's life expectancy example. If you still need to finish setup or run the analysis, use the Week 1 exercise and the setup guide.

Bardach: The eightfold path

• Bardach and Patashnik, A Practical Guide for Policy Analysis, the overview of the eightfold path. Use it as a reference when discussing your project today.

What instructions files do

• What instructions would you like to write for your agent? Think about your preferred language, writing style, tools, file organization, or how you want it to approach a task.
• Instructions files carry your preferences and requirements into future sessions.
• If you do not state a choice, the agent uses its default. That default may differ from the choice you would have made.
• There are three places to write things down:
  • the instructions file, for your explicit preferences and recurring requirements,
  • memory, for your preferences and past decisions,
  • a Skill, for a procedure you call by name.
• The following sections cover memory and Skills separately.

Global vs. project instructions

• The scope depends on what you mean. “Use R” belongs in global instructions if you want R across projects. It belongs in project instructions if this particular project requires R.
• A project can have requirements that differ from your usual preferences. Write those exceptions in its project instructions.

Global and project file locations

• ~ means your home folder. These are the default global locations.
• On macOS and Linux, folders starting with a dot, such as .codex, are normally hidden. In Finder, Command–Shift–period shows hidden files.
• Read your existing instructions. Which match how you want to work? Which would you add, change, or remove? Decide whether each requirement belongs in global or project instructions.
• You can open and edit the files yourself, or ask the agent to show their contents and make the changes you specify.
• Codex documentation on AGENTS.md →
• Claude Code documentation on CLAUDE.md →

Checking whether the agent follows your instructions

• Run a task that gives the agent a chance to use the instructions you wrote.
• Check the result against your requirements. Which did it follow? Which did it miss or interpret differently?
• Revise an instruction if it does not express what you mean, then try again in a new session.
• Outputs can vary across runs. Check whether your requirements are met; a difference alone does not show that an instruction is unclear.

What memory is for

• Memory helps the agent continue earlier work without making you explain the background again.
• It can retain decisions and their reasons, problems you discovered, approaches you already tried, and work left to do.
• For example, suppose you discover that the 2023 data are incomplete. You decide to end the analysis in 2022 and check the missing records later. A useful memory records that decision, its reason, and the unfinished check.
• In a later session, that information helps the agent continue the analysis without repeating the same investigation.

Instructions vs. Memory: When to use each

• A past decision can become a project requirement. If all future analyses must use the agreed period, add “Use data through 2022” to the project instructions.
• Review memories when you resume work. A previous decision may need updating when the data or task changes.

Memory vs. conversation history

• Conversation history is the record of your chat: the questions, answers, and discussion about the incomplete data.
• Memory can keep the conclusion: why you chose 2022 and what still needs checking.
• Use history to revisit the original exchange. Use memory to carry useful information into later work.

Reviewing and managing memory

• What gets saved and reused depends on the tool and your memory settings. Check what was actually saved.
• Local memory files and controls depend on the tool:
  • Codex: ~/.codex/memories/. Use /memories to control whether a chat can use memories or contribute to future memories.
  • Claude Code: memory files under ~/.claude/projects/. Use /memory to review and manage memory.
• Local memory is separate from the shared project folder. Put decisions your collaborators need in shared project documentation.
• Codex documentation on memories →
• Claude Code documentation on memory →

What a Skill is

• A Skill is a written procedure the agent follows when you ask for it by name.
• Writing a Skill is the same act as writing down your rules. The instructions file holds the rules that always apply; a Skill holds a procedure you call when you want it.
• A Skill lives in a folder with a SKILL.md file that states what it does and the steps to follow.
• Choose a procedure you expect to repeat. Checking a saved analysis is one example.
• The SKILL.md format is an open standard, so the same folder works in both tools. Each tool reads its own location:
  • Codex reads ~/.agents/skills/, ~/.codex/skills/, and .agents/skills/ inside the project,
  • Claude Code reads ~/.claude/skills/ and .claude/skills/ inside the project.
• To list the Skills the agent can see:
  • Codex in the terminal or in the VS Code extension: /skills.
  • Codex in the ChatGPT desktop app: ask for the list in plain language.
  • Claude Code in the terminal: /plugin.
  • Claude Code in the desktop app: type /, or click + and choose Slash commands.
• Codex documentation on Skills →
• Claude Code documentation on Skills →

Skill sources and safety

• A Skill written by a stranger runs on your machine.
• Read its SKILL.md line by line before you use it, and know which files and commands it can touch.
• Check the source and installed version of a Skill before updating it.
• Each tool ships an official catalog:
  • Codex: the openai/skills catalog, installed by typing $skill-installer and a skill name inside the session,
  • Claude Code: the Anthropic catalog, installed with /plugin install in the terminal, or in the desktop app by clicking + and choosing Plugins.

1. Your initial analysis

• Before using AI, write a basic problem definition for your own project: what is happening, to whom, where, and who could act on it?
• Make a first pass through the eight steps. What evidence might you need, what policy alternatives could be considered, and which questions are still open?
• Save this version so you can compare it with your revision.

2. Share with the class

• Present your problem definition and initial analysis to the class.
• Ask for feedback on the problem, possible alternatives, and questions you have not yet resolved. Record the suggestions.

3. Revise with AI and class feedback

• Give the agent your initial analysis and the feedback you received. Ask it to suggest revisions and explain its reasons.
• Decide which suggestions to use and revise your analysis.

4. Compare the two versions

• What changed in your problem definition or your application of the eight steps? What stayed the same?
• Which changes came from class feedback, AI suggestions, or your own reconsideration? Did they improve the analysis?
• Record 1–2 observations for your first memo, including a suggestion you accepted or rejected and why.

Weekly memo · due Sept 24 before class. Write 1–2 specific observations about using AI to define your project’s problem. Did it help, partly help, or fail to help? Use examples from your initial analysis and revision: what did AI suggest, and why did you accept or reject it? If class feedback shaped your revision, explain its contribution too. Aim for 150–200 words total. Write in your own words. Do not use AI to write the memo. Begin in class and submit to Canvas Discussions. See assignments and grading.