Course reference · Review quarterly

The supervised AI work playbook.

A compact reference for turning a professional task into a bounded, verifiable workflow in Codex or Claude Code.

Goal

Name the result, audience, and intended use.

Context

Choose current, relevant, minimal evidence.

Boundaries

Protect facts, privacy, scope, and external action.

Done when

Define the check and ask the agent to show evidence.

Choose the surface

Use focused chat when…	Use an agentic workspace when…
The answer is contained in the prompt.	The result depends on several project files.
You need explanation, ideas, or a draft.	The agent must use tools or connected systems.
You can review the result in one pass.	The work has multiple steps and needs iteration.
No persistent project context is required.	A test, calculation, comparison, or other check can close the loop.

Copyable brief

GOAL
Create [artifact] for [audience and use]. Lead with [priority].

CONTEXT
Use [current sources]. Take [specific evidence] from each.

BOUNDARIES
Keep [approved facts] unchanged. Flag [conflicts or gaps]. Do not [external or irreversible action].

DONE WHEN
Check [required fields, calculations, sources, tests, or visual result]. Show the evidence and list anything unresolved.

WORKING AGREEMENT
Inspect the available context first. State material assumptions. Ask before any external, irreversible, or consequential action. Produce a reviewable draft, run the available checks, and report the evidence used.

Permission ladder

Level	Suitable for	Human gate
Read and summarize	Approved internal sources; no state change.	Review source access and the final summary.
Draft for review	Files, proposed messages, plans, analysis.	Approve before sending, publishing, or committing.
Act inside a boundary	Reversible, well-tested, repetitive operations.	Approve scope, limits, recovery, and verification first.

Verification checklist

Every material number traces to a named source.
Conflicting sources and version differences are visible.
Required fields, owners, dates, and sections are present.
Unverified claims are qualified or removed.
The agent ran the available check and reported the result.
A human reviewed consequential recommendations and actions.

Glossary

Agent: An AI system that can work through steps and use tools inside defined boundaries.
Workspace: The selected folder, project, files, tools, and instructions available for a task.
Context: The information that can materially change the result.
Boundary: A limit on files, tools, scope, data, or external action.
Artifact: A reviewable output such as a document, spreadsheet, presentation, website, or code change.
Verification: A check that produces evidence about whether the requested result is correct and complete.
Durable guidance: Project instructions reused across tasks, such as AGENTS.md or CLAUDE.md.
Skill: A packaged, repeatable workflow with instructions and supporting resources.

Current primary references

OpenAI Codex best practices · OpenAI prompting · Claude Code best practices · Claude Code security

Product interfaces and feature names change. Review this playbook and its source links quarterly.