Project file with a ChatGPT plugin: method and limits

Install the ChatGPT plugin: Project file →

Mrs Durand (a fictional client) emails the practice on a Tuesday evening: she wants to extend her house in Toulouse by about 40 m², she states a budget of €120,000 excluding VAT, she "will send the surveyor's drawings later". Three weeks later a shower room is added to the brief. In the meantime the practice has started a planning note and a draft permit notice. When the file is picked back up, nobody can say which version of the brief fed which analysis, or whether the survey was received, read, or only promised.

This article introduces the ChatGPT plugin Project file (technical id archi-projet, version 0.1.1), an Educasium resource for Codex: what it produces, the context of the mission, the step-by-step method, what sets it apart from a plain prompt, how to install it and its limits. It is built for the French construction framework, so French terms are kept with a short gloss.

Contents

1. What the Project file plugin actually does
2. The context: a client's request is not yet a project
3. Using the plugin step by step
4. What changes compared with a ChatGPT prompt or skill
5. Installing the plugin in Codex
6. When this plugin is not enough
7. Training in AI-assisted project management
8. Frequently asked questions

What the Project file plugin actually does

The plugin is a coordinator. It does not analyse a PLU (local planning scheme) or write a CCTP (technical specification) itself: it opens the file from the client's request, tracks documents and versions, and organises the hand-overs between the specialised missions of the Educasium suite.

Three components in the package

• A local client form (formulaire-client.html) that opens in a web browser. The client (the maître d'ouvrage in French practice) describes the operation, the municipality, the brief, the intended works, how far along they are and, optionally, an indicative budget excluding VAT, a target date and the documents they can supply. The operation reference is provided by the practice. The form exports a JSON file; nothing is sent to a server.
• Method instructions (the projet-operation skill) that guide the assistant to tell apart documents that are announced, received and actually read, and prepare the relevant missions.
• An optional local register in Python (3.10 or later, standard library). It creates the operation folder, copies documents into it, computes their fingerprints and saves successive JSON states without deleting earlier ones.

What you get as output

In French, as Markdown and JSON: an operation sheet, a mission map with statuses, the sources, pending human decisions, the changes and the next useful action.

Eight coordinated missions

The plugin does not launch the other plugins automatically.

The context: a client's request is not yet a project

The maître d'ouvrage (the client or project owner) defines needs and the brief; the maître d'œuvre (the architect or design team) designs and oversees the project within the mission they are given. That mission runs through phases (sketch design, preliminary design, permit, tender, construction, handover), each relying on documents produced or received earlier.

The first risk is documentary. A client request is a statement: it proves neither an existing floor area, nor a planning rule, nor that the person signing the email may make each decision. The form is recorded as an unverified source; announced documents stay "declared" until a file is actually received; unknown information stays empty rather than guessed; an indicative budget never becomes a contract amount. Personal data entered is kept to what the mission requires.

Using the plugin step by step

Step 1: Collect the request. The client fills in the local form, or the practice gathers the same information in conversation.

Step 2: Initialise the file. The assistant creates a new, empty operation folder. An existing file is never overwritten, nor linked just because the name looks similar: it is resumed by checking its status.

Step 3: Inventory the documents. Each document is announced, received but unread, or reviewed; gaps become clarification requests.

Step 4: Prepare the relevant missions. The assistant suggests the missions that matter, without forcing all eight or an artificial order. Incomplete work moves forward with its limits written down, without stepping over a decision not yet made.

Step 5: Record each hand-over. Each mission result becomes a hand-over: sources read and unread, upstream missions used, status (proposal, incomplete or bounded comparison), open points and established results.

Step 6: Resume after a change. Every time the file is resumed, the assistant checks its identity, the latest state and the fingerprints. If a source changes, the missions depending on it are flagged for re-examination.

A fictional example to test the method: "Create the project file for the fictional Maison Durand operation: a 40 m² extension to a house in Toulouse, stated budget €120,000 excl. VAT, no survey received. List the relevant missions and the decisions the client still has to make." Then check that the existing floor area, the survey and the deadlines stay flagged as missing and that nothing is presented as approved by the client.

What changes compared with a ChatGPT prompt or skill

A well-written prompt produces a good answer inside one conversation. The plugin leaves a trail on your disk that outlives the conversation.

Cumulative, versioned files

Each state of the file is a new document pointing to its parent. A modified file creates a new version and the old one is kept; replaying the same operation creates no duplicate. In the Durand scenario, the added shower room shows that the planning note and the draft notice rest on the previous brief.

Four human reviews per hand-over

Each hand-over opens four reviews: H1 for scope and source corpus, H2 for facts and cross-checks, H3 for the choices specific to the mission, H4 for the concrete deliverable. All of them start as pending. The script approves no decision and never imports the sign-off of another mission: a client can approve their brief without approving a hazard analysis. There is no global sign-off for the suite.

Installing the plugin in Codex

The download is reserved for participants in Educasium training courses. From the plugin page, download the ZIP archive and unzip it into a stable folder on your computer that you will not move afterwards. Open a terminal and register that folder as a catalogue with codex plugin marketplace add "path-to-folder", replacing path-to-folder with the real location. Then install the plugin with codex plugin add archi-projet@educasium-projet. Finally, open a new Codex session so the plugin is picked up.

You install the plugin once, then create one operation folder per project. Without Python, the instructions remain usable, with their limits stated.

When this plugin is not enough

The plugin prepares and keeps records; it does not decide. Everything it produces is a draft for the responsible architect to review.

• Decisions stay human. It confirms no mandate or sign-off, signs, submits or sends nothing, and changes no model.
• No certification. Not legal advice, not a technical inspection, not legally binding evidence.
• Not a collaboration tool. No multi-user access, no authentication, no hosted portal: it is a local folder.
• No BIM required, but exports depend on your tools. You do not need Revit; no Word or Excel export is guaranteed, so check the files actually created.
• Analyses depend on the specialist plugins. The coordinator replaces neither a planning note nor a CCTP.

Training in AI-assisted project management

A register is only useful if the practice knows which document is authoritative, who decides what, and when to reopen an analysis. Our Architectural Management with AI training builds these habits on concrete practice cases.

Frequently asked questions

Does the client need to install anything to fill in the form?

No. The form is a plain HTML file that the client opens in their browser. They enter the information, then download a JSON file and send it to the practice through the agreed channel. Nothing is sent to a server and no account is needed. If the client prefers, the practice can also gather the same information during a conversation and enter it themselves. Either way, the request remains an unverified statement: it opens the file, but it proves neither the floor areas, nor the applicable rules, nor the signatory's authority.

What happens when the brief changes during design?

The new version of the brief is recorded as a new version of the source, and the old one is kept. The register then flags every mission that depended on that source for re-examination, followed by the missions that depended on those. It deliberately invalidates on the cautious side: it does not decide on its own that a change is minor or purely cosmetic. It is up to the practice to assess the real scope of the change and redo the affected analyses, with a sourced review if it wants to narrow that scope.

Do I need all eight plugins to use this one?

No. The Project file plugin works on its own to open the file, inventory documents and prepare decisions. It suggests the missions relevant to the project without forcing all eight. When a mission needs a specialist plugin that is not installed, it clearly reports the missing capability and continues with whatever document framing is possible, without inventing results. Specialist plugins are installed once, when you need them, and then serve all your operation files.

Do the register's fingerprints count as evidence?

They help detect an error or a modified file; they do not prove anything legally. The register has no electronic signature, no qualified timestamp and no tamper-proof log: someone able to edit every file in the folder could recompute a history. The fingerprints are therefore an internal rigour tool, useful for knowing which version fed which analysis. For legally binding evidence, rely on the procedures and tools designed for that purpose.

Once the file is open, two downstream missions build directly on its sources and versions: our ChatGPT method for the CCTP by trade for drafting specifications, and our ChatGPT method for site meeting reports for cumulative visit follow-up.

Funding may be available depending on your eligibility (OPCO, FIF PL). Qualiopi-certified training provider. To structure how you track your operations with AI as part of our Architectural Management with AI training, contact Educasium and tell us your status (employee, self-employed, business owner) and your goal. You will find the download and the installation tutorial on the dedicated ChatGPT plugin page.