Codex Project Operations Template: Requests, AGENTS.md, Config, Skills, and Verification
A practical operating template that combines Codex task requests, AGENTS.md, config, skills, subagents, and verification.
AI Engineering Hub: Coding Agent Operations and Verification
A hub for making AI coding agents more repeatable through harness design, context control, verification, permissions, and security boundaries.
This page treats AI coding agents as operational systems, not just prompt targets. It organizes existing posts by the problem a reader is trying to solve.
Problems Covered Here
- Codex or Claude Code gives different results for similar requests.
AGENTS.mdandCLAUDE.mdkeep growing without a clear boundary.- Long logs, plans, and memory consume the context budget.
- Build and test pass, but agent work still needs a verification loop.
- MCP, hooks, settings, permissions, approvals, and guardrails need security boundaries.
Short Answer
Reliable agent work comes from separating responsibilities: task request, instruction file, config, tool permissions, trace, and validation loop. Keep always-on documents short, move repeated procedures into templates or skills, and use settings, permissions, hooks, and CI for enforceable boundaries.
Core Concepts
- Harness Engineering: start with What Is Harness Engineering?.
- Context/Token Management: start with Why Token Management Matters in Harness Engineering.
- Agent Workflow: pair How to Write Your First Codex Task Request with Operating Codex Plan-First.
- Verification Loop: use Build and Test Are Not Enough to Validate an Agent.
- Guardrail/Approval: use Approval Boundaries and Guardrails.
Reading Order
- Why Do AI Coding Tools Produce Different Results?
- What Is Harness Engineering?
- Project Instruction Files Should Not Be Control Planes
- How to Write AGENTS.md for Codex
- Why Token Management Matters in Harness Engineering
- Build and Test Are Not Enough to Validate an Agent
Problem-Based Paths
When Codex results keep drifting
- Treat Codex as a Work Execution Agent
- Why Codex Needs a Harness
- How to Write Your First Codex Task Request
- Operating Codex Plan-First
When AGENTS.md is unclear
- How to Write AGENTS.md for Codex
- Why a Good AGENTS.md Should Be Short
- Project Instruction Files Should Not Be Control Planes
- Shared Problems in AGENTS.md, CLAUDE.md, and System Prompts That Burn Tokens
When context and token pressure grow
- Why Token Management Matters in Harness Engineering
- Long Logs, Long Plans, Long Memory
- How to Design State Summaries That Save Tokens
- How Token Management Strategies Differ Between Codex and Claude Code
- Token Management Series
When agent output needs verification
- Build and Test Are Not Enough to Validate an Agent
- From Prose to Schema: Turning Handoff Into a Contract
- Why Trace Matters More Than Results
- From Documents to an Observable Harness
- How to Read a Jenkinsfile: agent, stages, steps, and post
- PR/MR Collaboration Flow and Review Criteria
When MCP, hooks, and permissions need boundaries
- Approval Boundaries and Guardrails
- From Principles to Enforcement
- Future candidates: Claude Code hooks, MCP data exposure checks, task-level tool permissions, and prompt injection as a harness problem
Practical Templates
- AI agent operations templates
- Includes a minimal
AGENTS.md, minimalCLAUDE.md, Codex task request prompt, agent work review checklist, and Claude Code permissions/settings checklist.
Related Posts
Codex Subagent Usage Criteria: When Parallel Agents Are Worth It
Explains when Codex subagents are worth using and when a single-agent flow is safer.
Codex Skill Usage Criteria: When to Standardize Repeated Workflows
Explains when to move repeated writing, review, and release procedures into Codex skills.
Codex Config Operations: Fix Model, Permissions, Sandbox, and MCP Defaults
Explains why model, permission, sandbox, and MCP defaults should be fixed in Codex config for repeatable work.
Codex Plan-First Operations: Fix Scope and Verification Before Editing
Explains how to run complex Codex tasks plan-first by clarifying scope, risk, and verification before edits.
AGENTS.md Length Criteria: Why Shorter Project Instructions Work Better
Explains why a long AGENTS.md increases token cost, duplicate instructions, and mixed responsibilities.
AGENTS.md Writing Guide: Short Operating Rules for Codex
Explains how to write a Codex AGENTS.md with repository purpose, inspection paths, working rules, and verification criteria.
Codex Task Request Template: Goal, Scope, Constraints, and Verification
Shows how to write a first Codex task request with goal, scope, constraints, completion criteria, and verification.
Why Everything Claude Code Is Trending: The Repository Turning Claude Code into an Operating System
A quick analysis of the 168K-star Everything Claude Code repository, based on its code structure, Claude Code’s official extension model, external articles, ...
Codex Harness Design: Why Better Prompts Are Not Enough
Explains why reliable Codex work needs project rules, permissions, and verification, not only better prompts.
Codex Operating Model: Treat It as a Work Execution Agent, Not a Code Generator
Explains why Codex should be treated as a repository work agent, not just a code generator.
Observable Harness Migration: Making Document-Centered Agent Ops Verifiable
A practical roadmap for moving from document-centered operations to an observable harness.
Agent Approval and Guardrail Boundaries: Why Empty Boundaries Break Systems
Explains why agent systems become unstable when approval boundaries and guardrails are missing.
Agent Trace Design: Why Execution Records Matter More Than Results
Explains why results alone are not enough and why trace matters for operations, debugging, and evaluation.
Harness Enforcement Design: Turning Principles into System Rules
Explains how to turn natural-language principles into concrete system rules and enforcement.
Multi-Agent Usage Criteria: Treat It as an Option, Not the Default
Explains why multi-agent setups are not the default answer and when a single agent is the better baseline.
AI Agent Verification Loop: Why Build and Test Are Not Enough
Explains why build and test alone are not enough to validate an agent and what extra checks are needed.
Handoff Schema Design: Turning Free-Form Notes into a Verifiable Contract
Explains why handoffs should move from free-form prose to schema-based contracts and how to design them.
Project Instruction File Scope: Why AGENTS.md and CLAUDE.md Should Not Be Control Planes
Explores what a project instruction file should own and where it should stop inside a larger harness.
Harness Engineering Concept: Designing the Agent Execution Environment
Explains harness engineering as the design of the tools, permissions, and verification around AI.
AI Coding Tool Result Drift: Why the Harness Changes the Output
Explains why AI coding tools produce different results by looking beyond prompts to the execution environment.