Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/gpt-pf-chat-gpt-project-forge/llms.txt

Use this file to discover all available pages before exploring further.

The fastest way to understand Project Forge is to run it once on a real, small case. This guide walks through a complete first session: from a clean workspace to a finished opening package ready to hand off to the target project. You do not need to understand every part of the system before you start. You need to close the target, load the right files in the right order, and let the system do the rest.
Start smaller than you think. A clean opening is usually smaller than anxiety wants it to be.

​
Which File Opens What

Before the first session, orient yourself to the repository layout. Each file has one job:
NeedFile
Start an AI session correctlyAI_START.md
Understand system identity and scope00_SCOPE.md
Understand criteria and authority01_RULES.md
Understand artifact classes03_ARTIFACTS.md
Understand working procedure02_PROTOCOL.md
Operate as a humanoperator_guide/OPERATOR_GUIDE.md
Fill opening files manuallyartifact_templates/
Use guided browser compilationWeb Compiler

​
Your First Preparation Session

1

Start from a clean workspace

Open a new chat or a new ChatGPT Project dedicated to Project Forge preparation. Do not reuse an existing project chat that already carries context from other work.Session hygiene is the most important practical step. GPT systems are good at sounding continuous even when that continuity is implicit, partial, or contaminated. A clean start means the opening package you produce will reflect only the inputs you deliberately provide — not carry-over from earlier conversations.If your environment supports saved memories or default project continuity, be aware that those surfaces can introduce context you did not intend. When the preparation matters, prefer a genuinely isolated session.
2

Load AI_START.md and the four canonicals in read order

Provide the following files to the assistant, in this order:
  1. AI_START.md — startup routing file; tells the AI what not to assume and how to enter the system
  2. 00_SCOPE.md — system identity, scope, non-scope, and final output
  3. 01_RULES.md — criteria, authority hierarchy, stability, residence, validation, and readiness
  4. 03_ARTIFACTS.md — artifact grammar, artifact classes, schemas, triggers, and freshness rules
  5. 02_PROTOCOL.md — minimum working procedure, state definitions, stop conditions
Read order matters for comprehension. AI_START.md routes the AI into the canonicals correctly. Do not start from 02_PROTOCOL.md alone — without the frame, rules, and artifact grammar established above it, procedure can be misread as permission.
Read order:
00_SCOPE.md
01_RULES.md
03_ARTIFACTS.md
02_PROTOCOL.md
Precedence order (for conflict resolution) is: 00_SCOPE.md > 01_RULES.md > 02_PROTOCOL.md > 03_ARTIFACTS.md. Read order is for comprehension; precedence is for conflict closure.
3

Give the minimal operator instruction

Once the files are loaded, give the assistant one opening instruction. This is the minimum required to start a guided preparation session:
Read AI_START.md, then use the Project Forge canonicals to prepare the opening package for this target project.
Then follow with the target project information in the next step. The assistant will use the canonicals to close the object, classify information, and determine what artifacts are needed.
4

Define the target project

Describe the project you are preparing. At minimum, close three things:
  • Objective — what the project is for and what it must produce
  • Scope in — what is explicitly included
  • Scope out — what is explicitly excluded
Until these three are stable, the honest readiness state is NOT_READY. That is a normal state — do not skip past it.Worked example — sports coverage case:Suppose you want to open a MotoGP project, but not as a generic fan chat. The project should follow MotoGP through foreign press coverage: tracking what different countries emphasise, which sources are authoritative, which narratives recur, and what kind of briefing the project should produce.Your opening instruction would look like:
Read AI_START.md, then use the Project Forge canonicals to prepare the opening package for this target project.

Target: MotoGP season coverage through foreign press.
Objective: Track how different countries cover MotoGP — emphasis, narratives, framing, and source patterns.
Scope in: Press sources from Italy, Spain, UK, Japan. Published articles and official team communications.
Scope out: Social media, fan forums, unofficial commentary.
Expected output: Weekly briefing summarising cross-country narrative divergence.
Project Forge helps close the target, the countries or languages in scope, the source perimeter, what counts as official basis, what is excluded, and the expected output shape — before the real project is ever opened.
5

Compile the opening artifacts

Work with the assistant to produce the opening package. The minimum useful package contains:
  • ENTRY_POINT.md — tells the target project how to read the package
  • TARGET_PROJECT_BRIEF_ARTIFACT — defines objective, scope, expected outputs, and constraints
  • INITIAL_SSOT_ARTIFACT — freezes the initial official basis
Add optional artifacts only when genuinely required:
  • HANDOFF_ARTIFACT — only when the next run cannot safely reconstruct state from the stable artifacts
  • SOURCE_OR_MATERIAL_TRANSFER_ARTIFACT — only when candidate material must travel to the target project without becoming official basis yet
If you prefer to fill the artifacts yourself rather than through guided compilation, the templates are in artifact_templates/. If you want a structured browser surface for the same job, use the web compiler.
6

Review the package

Before opening the target project, verify the package is complete:
  • Does the entry point correctly describe how to read the package?
  • Is the target brief specific enough that a fresh session could understand the project without asking?
  • Does the initial SSOT name only validated, approved sources — not candidate material?
  • Are any critical fields missing or marked as unknown?
If critical fields are missing, the state is NOT_READY. Mark what is missing explicitly rather than leaving it implicit. The target project should not have to infer what was meant.
7

Open the real target project from those files

Create or enter the real target project and load the opening package at startup. The target project reads the entry point, the brief, and the initial SSOT — not the preparation chat.This is the test: the target project must be able to start cleanly without rebuilding hidden context from the preparation session. If it can, Project Forge has done its job. If it cannot — if something essential lives only in the preparation chat and not in the files — the opening package is not finished.Files beat memory. If something matters, it belongs in a file.

​
What a Finished Session Looks Like

A successful first session produces a small, explicit package. You will have closed the object, the target, and the required outputs. You will have an entry point, a target brief, and an initial SSOT — each filling its own role, none overlapping. The preparation workspace can be closed or archived. The target project starts from the package, not from the conversation. For a concrete minimal example ready to run, see smoke_test/CASE_01_MINIMAL_OPENING.md in the repository.

Build docs developers (and LLMs) love

Get started for freeTalk to us