Codex 101 for Construction PMs and VET Leaders: Working Inside Your Project Folder
Published 26 September 2026
Codex is not just a chat box. It works inside a project folder on your computer, where the real work already lives: RFIs, variations, site photos, modules, mapping tables, learner evidence, templates and review notes. This article explains the idea, a five-level ladder for building up to it safely, and what has changed in OpenAI’s tools since our guide was written.
You do not need to code to use it. You do need to decide what it may read, what it may change, and who reviews the result.
The key idea: the folder is the workspace
In a browser chat you copy pieces of your work into a window. With Codex you bring the AI to the work folder instead. It can read the files there, compare them, create new ones and follow standing instructions across sessions, using the permissions you give it.
That is the difference from browser-based tools such as ChatGPT or Gemini in chat mode, which mainly talk with you. It matters whether you manage variation registers, SWMS documents and apprentice logbooks, or a corpus of PDFs and interview transcripts. The underlying capability is the same: read what is in the folder, work across several files, and produce something you can review.
Two familiar scenarios
Load RFIs, variations, site instructions, subcontractor quotes, daywork sheets, photos, claims, email threads and the contract schedule. Ask Codex to trace what changed, what it affected, what evidence supports it, and where the cost impact may sit.
What it is doing: building an evidence trail of event, source documents, dates, cost category, missing records, assumptions and items needing QS review.
The human still owns: entitlement, valuation, the contract position, and anything submitted commercially.
Load units of competency, assessment tools, mapping documents, training and assessment strategy (TAS) files, learner samples, validation records, trainer credentials and existing modules. Ask Codex to map coverage, identify gaps, flag duplicate coverage and prepare a review pack.
What it is doing: building a coverage map of requirement, module location, assessment task, evidence source, weak or duplicate coverage, and review questions.
The human still owns: compliance judgement, assessment decisions, validation, and RTO accountability.
The professional boundary: Codex can prepare the evidence trail. It does not own entitlement, valuation, safety sign-off, assessment judgement, compliance approval or trainer competency decisions.
The five levels
The Prompt User
What it is: you open Codex and ask isolated questions. It suits quick drafting, summarising and plain-English explanation, such as a clearer email about a delay or a simpler learner instruction.
Why people stall: every prompt starts from zero. Codex has no access to the folder, project history, template or evidence.
Move up: open the actual project folder and give Codex a real task with context.
The Folder Worker
What it is: Codex can now see the job pack, not just your question. Give it the relevant folder and a clear definition of done: what to read, what to ignore, what output you want, and what must be flagged for review. A QS might ask for a draft change-event register with document references and missing evidence. A VET leader might ask for a draft coverage map and compliance questions.
Why people stall: they dump files into a folder and hope. Codex still needs instructions, boundaries and output rules.
Move up: ask Codex to trace every claim back to source documents. Evidence mapping is where the value starts to build.
The Evidence Mapper
What it is: Codex compares, traces and groups, and shows its working. It links claims to source documents, shows gaps, and separates evidence from assumptions. Good output includes source references, confidence levels, missing evidence, contradictions, assumptions and items that need a human decision. The quality test is simple: a table is only useful if it leads back to the evidence.
Why people stall: they accept the table without checking the trace.
Move up: write standing instructions, and turn the mapping task into a repeatable workflow.
The Workflow Builder
What it is: you stop explaining the same review method every time. A standing project brief (the AGENTS.md file) holds your rules, and reusable Skills hold common jobs: for a QS, an RFI-to-variation impact review or a progress claim support checker; for a VET leader, a unit-to-module mapper or a validation pack reviewer. A good Skill defines the input folder, evidence rules, output table, review flags, file naming and the human sign-off point.
Why people stall: they get one good result but never save the process, so the next review starts from memory again.
Move up: choose one low-risk recurring review and make it routine, with a clear human approval point.
The Governed Operator
What it is: Codex becomes part of a controlled review system. The question is no longer “Can Codex draft this?” but what it is allowed to read, what it can draft, what it must never change, who reviews the output, and what audit trail you keep. Approvals, sandboxing, controlled connectors, Skills, automations, role-based access and review logs keep the workflow reviewable.
Why people stall: trust. The tool is not the hard part. The hard part is deciding where human accountability sits.
Best first routine: start with one valuable but low-risk review. Keep the first version boring, reviewable and human-approved.
What has changed since the guide
Our guide was last reviewed in May 2026, and OpenAI’s tools have moved on since. Checked against OpenAI’s documentation in September 2026:
- New home. OpenAI describes the standalone Codex app as merging into a new ChatGPT desktop app. Codex remains a separate view there, alongside Chat and a general-purpose agent called ChatGPT Work. Screens and menus differ from older screenshots.
- Not only for developers. OpenAI says non-developers are a growing share of Codex users, and it has added role-specific plugins for knowledge work.
- Permissions come first. The current documentation covers permission profiles, sandboxing, agent approvals and an auto-review feature. Understand these before pointing Codex at live project files.
- Plans and limits vary. OpenAI states that Codex is included with ChatGPT plans, with usage limits that differ by plan. Check the current pricing page.
- Versions and models change often, sometimes weekly. We have left version numbers and model names out on purpose.
Names, plans and features change quickly, so treat this section as a snapshot and confirm details on OpenAI’s own pages before you rely on them:
Before you start: guardrails
Codex acts on files, so the habits matter more than they do in a chat window:
- Start with a practice folder. Use a harmless folder of sample documents before you touch live job or compliance files.
- Check the folder first. Look at what is inside before asking Codex to read everything.
- Read first, write later. Begin with read-only summaries, and allow file creation only when the proposed output makes sense.
- Protect your data. Keep confidential, commercially sensitive and personal information out unless your policy and account settings allow it, and know where it goes.
- Keep a human in charge. Review every output, and keep accountability for commercial and compliance decisions with a person.
Read the full guide
The interactive guide covers the key idea, a first session, the five-level ladder and a printable reference summary, with examples for quantity surveyors, project managers and VET leaders. It takes about 10 to 15 minutes.
Open the guide: Codex 101 for Construction PMs and VET LeadersThe guide ties its Codex claims to public OpenAI documentation. The product names and screens in it reflect its May 2026 review.
Skills, Not Prompts: build a procedure once, run it in any AI
Codex reads the same SKILL.md folders as Claude, Cursor and Copilot. Learn what a skill is, where to find real ones, and which jobs to build first for construction and VET work.
Read the skills guideKeep reading
More on AI, construction, education and project management.
Use AI as a guide, not the final word
AI is powerful, and tools like these help keep answers on track, but they can still be wrong or out of date. Always review the outputs yourself. Make sure any codes, standards or regulations they quote are current and match your local specifications before you rely on them.