ChatGPT · Where each workflow layer lives
This Step 2 mapping tells you where to put each workflow layer inside ChatGPT: user-wide instructions, project-level instructions and sources, reusable GPTs and Skills, run-time prompts, and verification gates. Projects are the main workspace surface for long-running workflows (instructions + files + chats + memory).
1) ChatGPT surfaces you can configure
- Projects (workspace)
- Projects group chats, upload reference files, and add custom instructions for the project so ChatGPT “stays on-topic”. Projects support memory, including project-only memory. In some workspaces, Projects use a tabbed layout with Chats and Sources; project files may appear under Sources.
- Project Instructions (project-scoped rules)
- A project-level instruction surface: use it for workflow rules that should apply across all chats in that project.
- Project Sources / Project files (project reference materials)
- A project-level source surface: upload files as reference material. In some workspaces, “Sources” is a dedicated tab. In some workspaces you can also add sources from apps, conversations, or quick text inputs.
- Project memory / project-only memory
- Project memory keeps context in the project. With project-only memory, the project does not reference saved memories from outside the project and cannot reference conversations outside the project.
- Custom GPTs (reusable assistants)
- Custom GPTs are configured for a specific purpose using Instructions (behavior) and Knowledge (reference files), plus selected capabilities. Instructions apply to every conversation with that GPT.
- Skills (reusable workflows)
- Skills are reusable, shareable workflows that can bundle instructions, examples, and even code. Once installed, ChatGPT can automatically use a skill when it’s helpful (availability depends on plan/workspace).
- Custom Instructions (user-wide defaults)
- Custom Instructions apply immediately to all chats. Use only for rules that should be user-wide (not project-specific).
- Memory (account-level)
- Memory has two core controls: saved memories and reference chat history. Turning off “Reference saved memories” also turns off “Reference chat history”. To chat with a blank slate (no memories/history), use Temporary Chat; it still follows Custom Instructions if enabled.
- Library (file reuse)
- Uploaded and created files can be saved to Library so you can reuse them later. To add a saved file to a chat: open the attachment menu and choose “Add from library”.
- Apps (connectors)
- Apps bring external tools and data into ChatGPT. You can invoke an app in chat via @mention, or click “+” then “More”.
- Deep research (documented research)
- Deep research produces a structured report with citations/source links and lets you choose which sources it can use (websites, uploaded files, connected apps). Use it when the workflow requires evidence-backed research outputs.
2) Layer → where to put it in ChatGPT
Use this mapping after Step 1 (layer classification). The goal is not “more features”, but correct placement: rules go into instruction surfaces, reference material goes into sources/knowledge/file surfaces, repeatable procedures become reusable GPTs/Skills/Projects, current requests stay in the current chat, and verification stays as a gate/checklist.
- Instruction layer
-
Put stable workflow rules here (rules that must persist across messages).
- Project Instructions: project-scoped rules for workflows you run inside a Project.
- GPT Instructions: reusable assistant rules when the workflow is packaged as a Custom GPT.
- Skill instructions: packaged workflow rules when implemented as a Skill.
- Custom Instructions: only for user-wide defaults that should apply across all chats.
Memory is not an instruction surface. If the workflow must not be shaped by account memories/history: prefer project-only memory in Projects (where available) or run it in Temporary Chat.
- Reference / source material layer
-
Put grounding material here (policies, docs, examples, evidence the model should consult).
- Project Sources / Project files: upload reference files to the Project; in some workspaces this is the “Sources” tab.
- Add sources from apps/conversations/text: where available, build a “living” project knowledge base by adding sources from supported connectors, conversations, or quick text inputs.
- GPT Knowledge: attach reusable reference files directly to a Custom GPT.
- Library: reuse saved files across chats via “Add from library”.
- Apps: reference external systems when the workflow requires data that lives outside ChatGPT.
Retention: files uploaded as GPT Knowledge are retained until the GPT is deleted; files uploaded in chats are retained per chat retention.
- Reusable workflow layer
-
Put the repeatable procedure here (so you don’t rewrite it every time).
- Project: best when the workflow depends on shared project context (instructions + sources/files + project memory).
- Custom GPT: best when you want a reusable “assistant/tool” with its own Instructions + Knowledge.
- Skill: best when you want a packaged workflow; skills can bundle instructions/examples/code and can be auto-used when helpful.
- Runtime prompt layer
-
Put the one-off task input for the current run here.
- Current chat message inside the relevant Project / GPT / Skill run (the actual task request).
- Per-run attachments only if they are specific to this run (don’t promote to Knowledge unless reusable).
- Verification layer
-
After you get an answer, run a verification pass before you use it.
Either (a) ask ChatGPT to verify the output against your checklist/sources in a second message, or (b) enforce a required verification section via Project/GPT/Skill instructions.
- Guide checklist / review gate: a fixed final pass before publish/send/use.
- Skill verification step: embed verification as an explicit workflow step when packaged as a Skill.
- Deep research citations: for research workflows, use the report’s citations/source links as verification substrate.
UI note: Project “Sources” may appear as a dedicated tab in some workspaces; treat it as the official surface for project reference material. User-wide Custom Instructions and account Memory are separate surfaces; use them only when the rule is meant to apply beyond a single project/workflow.