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.

GPT systems are very good at sounding continuous. They maintain a consistent tone, fill in plausible details, and give the impression that they are operating from a coherent shared understanding. That fluency is useful in casual conversation. In Project Forge, it is a risk. The system may appear to work smoothly while actually pulling from prior chat history, ambient session memory, or implied continuity that was never explicitly established. Session hygiene is the discipline of preventing that from happening — and catching it quickly when it does.
If you have to choose between smoothness and explicitness, choose explicitness.

​
The Core Hygiene Problem

Project Forge is built on explicit context. Every preparation run is supposed to derive its authority from the canonical files and the artifacts you explicitly provide — not from what the model happens to remember, not from what was discussed three messages ago, and not from momentum in the current conversation. When session hygiene breaks down, the system does not immediately fail in an obvious way. It continues producing output. The output looks reasonable. But the basis for that output has shifted from explicit files to implicit inference, and that shift is not visible in the response. You only discover the problem when an artifact is wrong in a way that is hard to trace, or when the target project starts and cannot reconstruct what it was supposed to start from. The hygiene rules below exist to prevent that gap from opening.

​
Best Practices

For any work that matters, start from a new, clearly bounded chat or project. Do not rely on conversational carry-over from earlier in the same session unless that carry-over has already been externalized into an explicit artifact or file.A clean start means: new chat, explicit files, no assumed continuity. If you are not sure whether the current session is clean, treat it as potentially contaminated and prefer starting fresh over repairing the doubt.
GPT environments can carry context in several places you may not be thinking about: saved memories, project-level instructions, chat history references, and default continuity within a project. When you need a clean run, be deliberate about what is active.If the environment can reference previous chats, prior sessions, or persistent memories, assume contamination is possible unless you have deliberately disabled it or isolated the session. Do not assume that because you did not consciously provide something, the model does not have access to it.
If something matters to the current preparation run, put it in a file or artifact. Do not trust:
  • Memory of an earlier answer in the same session
  • A vague sense that “the model already knows this”
  • Previous chat continuity
  • Remembered tone or framing from a prior run
Trust: canonicals, explicitly provided artifacts, validated source material. If the artifact does not exist yet, it does not exist. Do not act as if it does.
A good startup pack is:
  • AI_START.md
  • The 4 canonicals (00_SCOPE.md, 01_RULES.md, 03_ARTIFACTS.md, 02_PROTOCOL.md)
  • Only the artifacts that actually exist
  • Only the source material you intend to expose
Do not give ten files “just in case.” Volume creates noise before it creates clarity. Do not give unclassified notes, old chat excerpts, or mixed drafts that have not been evaluated against the canonicals. Every file you add to a session is a potential basis for something the AI infers — make sure every file you add is a file you actually want informing the run.

​
Drift: What It Looks Like

Drift is when the AI is no longer operating from the explicit current package. It happens gradually and is often not obvious from the AI’s output alone — the responses still sound coherent and on-topic.
Six signs of drift to watch for:
  1. The AI brings in facts or examples from previous chats you did not provide
  2. The AI acts as if an artifact exists when it has not been provided in this session
  3. The AI speaks as if state is already known even though it was never externalized
  4. The AI answers from chat momentum instead of from the canonicals and artifacts
  5. The AI resolves ambiguity without stopping when the ambiguity should have changed the move
  6. The AI speaks from prior discussion as if it were stable, validated basis
Any one of these is a signal to stop and verify what the AI is actually operating from.

​
The Drift Reset Procedure

When drift appears, do not try to argue with it inside the drifting session. Do not add more explanation or context on top of the contaminated state. Use the smallest reset that restores explicit context.
1

Stop the Current Move

Do not continue producing output from the current state. Stop before the drift compounds. An artifact built on contaminated context is not a useful artifact.
2

Restate the Current Object

Say explicitly what the preparation run is for — what target project is being prepared, what opening decision is being made. Anchor the session back to a named, stated object.
3

Restate the Active Authoritative Files

Name the files that are authoritative in this session: the canonicals that were provided, the artifacts that actually exist. Make the basis explicit rather than assumed.
4

Restate What Does Not Exist

Be explicit about what has not been provided and what has not been created yet. “We have not built the SSOT yet. We have not provided a handoff. These do not exist in this session.” This prevents the model from filling gaps with invented state.
5

Reopen in a Clean Session if Trust Is Gone

If the session has clearly become contaminated and the above steps do not restore confidence in the current context, a clean restart is often cheaper than continuing to patch trust into a compromised session. Open a new chat, provide the explicit startup pack, and begin from verified ground.

​
The Reset Strategy

The rule is: use the smallest reset that restores explicitness. Not the most thorough possible intervention — the most targeted one that brings the session back to explicit, verifiable ground. If restatement is enough, restate. If the session needs a new context window, open one. If the prior artifacts are still valid, use them. If the prior artifacts are now suspect, treat them as unverified and re-validate before using them as basis. The cost of a clean restart is always lower than the cost of an artifact built on contaminated context.

​
AI_START.md Startup Hygiene Rules

AI_START.md encodes the hygiene contract for every Project Forge session. Its rules are worth knowing because they are the baseline the AI is supposed to hold — and when drift appears, it often means one or more of these rules was not enforced:
  • Do not assume — do not treat past chat memory as stable input, do not treat chat history as authority, do not assume continuity from previous sessions, do not assume an artifact exists unless explicitly provided
  • Session hygiene — prefer explicit files over implicit memory, prefer provided artifacts over remembered context, treat chat as temporary unless externalized, keep verified, inferred, and missing states distinct
  • First move rule — before doing anything else, close the real object, close the target, determine required outputs, determine whether project-opening artifacts already exist, determine current state
  • First question rule — ask a question only if the answer changes target, authority, readiness, or required artifacts; if no such question is needed, proceed without asking
When a session drifts past any of these rules, the drift reset procedure above brings it back.

Build docs developers (and LLMs) love

Get started for freeTalk to us