ChatGPT workflow placement

Use this reference to decide where a workflow belongs in ChatGPT: Projects, GPTs, Skills, memory, tools, apps, agent mode, or runtime prompts.

Start with the workflow layer, then choose the ChatGPT surface

Use this reference to decide whether a workflow belongs in Project Instructions, project files, Project Sources, Custom GPT instructions, GPT Knowledge, Skills, Custom Instructions, Library, apps/actions, agent mode, or the runtime chat prompt.

Choose the ChatGPT experience first

Start by choosing the experience that matches the work. Then select the reasoning level, tools, persistent context, reusable workflow surface, and external integrations required by the task.

Chat

Use Chat for questions, conversation, brainstorming, quick analysis, search, file review, and tool-assisted work that remains centered in a conversation.

Work

Use Work for longer, more involved tasks that require research, connected apps or files, active steering, approval points, and finished documents, spreadsheets, presentations, reports, or Sites.

Codex

Use Codex for software-development workflows involving repositories, code changes, tests, diffs, pull-request review, terminal work, and implementation tasks.

Use the OpenAI API when the workflow must run outside ChatGPT

Use the API when the workflow must run inside a website, application, backend service, internal system, CI pipeline, or controlled production automation.

Open the API placement map →

Speed, model tier, and reasoning controls

Chat and Work expose different control models. In standard Chat, choose the balance of response speed and reasoning effort. In Work and Codex, choose a GPT-5.6 model tier and reasoning effort, with Max and Ultra available according to the product, plan, workspace configuration, platform, and rollout state.

Standard Chat: choose a speed and reasoning level

Standard Chat uses simplified controls rather than requiring the user to select every underlying model directly.

Standard Chat speed and reasoning levels
Option Use for Current model behavior
Instant Fast responses for everyday questions, short drafting, rewriting, summarization, and routine work. Powered by GPT-5.5 Instant and remains the default fast-response option.
Medium Standard reasoning for planning, structured comparison, analysis, and deliberate professional work. Powered by GPT-5.6 Sol on eligible plans.
High Extended reasoning for complex, multi-step tasks with several constraints or dependencies. Powered by GPT-5.6 Sol on eligible plans.
Extra High The highest standard-Chat reasoning effort for especially demanding tasks. Powered by GPT-5.6 Sol where available.
Pro Difficult tasks and longer-running workflows that require the highest-capability standard-Chat option. Powered by GPT-5.6 Sol Pro where available.
Automatic switching
On eligible paid plans, ChatGPT can automatically switch a request from Instant to Medium when additional reasoning is useful. Open Configure in the model picker to manage the automatic-switching control.
Managed-workspace labels
ChatGPT Business workspaces may show Pro Standard and Pro Extended rather than a single Pro label. The exact options depend on workspace access settings and rollout state.
Mobile placement
On iOS and Android, the simplified model picker can appear at the top of the conversation. The exact controls shown depend on the account plan, workspace, app version, and rollout state.

Work and Codex: choose a model tier and reasoning effort

Work and Codex expose GPT-5.6 model tiers separately from reasoning effort. These controls are not the same as the simplified Instant-to-Pro picker used in standard Chat.

GPT-5.6 model tiers and execution controls
Control Type Meaning Availability boundary
Sol Model tier The flagship GPT-5.6 tier for the most capable reasoning and professional work. Available in Work and Codex on eligible paid plans and managed workspaces.
Terra Model tier A balance of capability, speed, and cost for everyday agentic work. Available in Work and Codex according to plan and supported surface.
Luna Model tier The fastest and lowest-cost GPT-5.6 tier. Available in Work and Codex on eligible plans.
Reasoning effort Per-model control Controls how much reasoning time the selected GPT-5.6 tier receives. The available effort levels depend on the selected tier, plan, product, and rollout.
Max Reasoning effort Gives GPT-5.6 the highest available single-agent reasoning effort for demanding work. Available to users with GPT-5.6 access in Work and Codex and enabled through the relevant settings.
Ultra Multi-agent mode Goes beyond a single agent by using concurrent subagents and combining their work for complex, parallelizable tasks. Available in ChatGPT Work for Pro and Enterprise users, and in Codex for Plus and higher plans.
Max is not a model tier
Max is a reasoning-effort setting. It does not replace Sol, Terra, or Luna. Select a model tier first, then choose the available reasoning effort.
Ultra is not a normal reasoning level
Ultra is a multi-agent execution mode. It uses concurrent subagents rather than only giving one agent additional reasoning time.
Standard Chat boundary
Terra and Luna are not selectable in standard Chat conversations. They are exposed through Work, Codex, and the OpenAI API according to product and plan availability.

Mobile control model

Switch between Chat and Work
On mobile, use the dropdown at the top of the screen to switch between Chat and Work.
Standard Chat controls
In Chat, use the simplified speed and reasoning picker. The exact options depend on plan, workspace configuration, app version, and rollout state.
Work controls
In Work, eligible accounts can select the available GPT-5.6 model tier and effort controls. Max and Ultra appear only when the account, plan, product, and rollout provide access.
Codex boundary on mobile
Codex is not a selectable mobile mode. Supported desktop Codex tasks can be accessed from the mobile Remote tab, but this is different from selecting Work directly on mobile.

ChatGPT Work, agent mode, and Scheduled Tasks

These surfaces support different forms of extended or delegated work. Use Work to produce longer research and finished materials, agent mode to perform controlled actions, and Scheduled Tasks to run work later, repeat it, or monitor for meaningful changes.

Choose the appropriate execution surface
Surface Primary role Use it when Do not treat it as
ChatGPT Work Longer research and creation of finished materials. The task requires analysis, iteration, files, review criteria, active steering, or a completed document, spreadsheet, presentation, report, or Site. A Project, a single tool, or a synonym for agent mode.
Agent mode Delegated, multi-step online action. The task requires website navigation, forms, uploaded files, connected data, spreadsheet editing, supported commands, or another external action. A general writing mode or an unrestricted background worker.
Scheduled Tasks Future, recurring, or monitoring work. The task must run later, repeat on a schedule, or periodically check for a change and notify the user only when there is a meaningful update. Agent mode itself, Project memory, or Codex automations.

ChatGPT Work

Work is the ChatGPT experience for longer research, analysis, and deliverable creation. It can be used in a standalone conversation or inside a Project when the task needs persistent files, instructions, and related conversations.

Use Work for
Researching a topic, analyzing information, and creating finished documents, spreadsheets, presentations, reports, or Sites.
Start Work on web or mobile
Select Work, start a new conversation or open an existing Project, add the files and context required by the task, and describe the outcome or deliverable.
Start Work on desktop
Open the ChatGPT desktop app, start a standalone conversation or open a Project or local folder, then select Work from the mode switcher.
Cloud Work
Work on web and mobile runs in the cloud. Cloud Work can use the files and context supplied through supported cloud surfaces, but it cannot directly access files stored only on the user’s computer.
Desktop Work
Desktop Work can use local files and desktop apps after the user grants access. Grant access only to the files and applications required for the current task.
Work and Projects
A Project keeps related conversations, files, and instructions together. Work performs the research or deliverable task. A Project is the context container; Work is the execution experience.
Review and control
Review progress, answer questions, change direction, and approve important actions. Treat generated documents, spreadsheets, presentations, reports, and Sites as outputs that require review.
Availability boundary
Work is rolling out to eligible accounts. Availability depends on the plan, workspace configuration, supported surface, region, and rollout state.

ChatGPT agent mode

Agent mode is for complex online tasks that require ChatGPT to reason, research, and perform actions while the user retains control.

Use agent mode for
Navigating websites, working with uploaded files, accessing enabled third-party data sources, filling forms, editing spreadsheets, running supported code or commands, and completing controlled multi-step online tasks.
Start agent mode
Select agent mode from the tools menu or type /agent in the composer. Describe the required result, allowed systems, action boundaries, approval points, and stopping condition.
Guide or interrupt the task
Agent mode can pause for clarification or confirmation. The user can provide additional instructions, change direction, or interrupt the task while it is running.
Authentication
When a website requires authentication or sensitive input, use browser takeover rather than placing passwords or private authentication data in an ordinary chat message.
Approval boundary
Require confirmation before sensitive, external, high-impact, or irreversible actions. Do not use broad instructions such as “handle everything.”
App boundary
Enable only the apps required for the current task. Agent access remains subject to the user’s source-system permissions and the workspace controls applied to the enabled apps.
Security boundary
Agent safeguards reduce but do not eliminate prompt-injection, privacy, or unintended-action risk. Stop the run if a website, message, or instruction appears suspicious.
Scheduling relationship
A Scheduled Task can invoke agent work where supported, but the schedule is created and managed as a Scheduled Task. Agent mode and Scheduled Tasks remain separate capabilities.

Scheduled Tasks

Scheduled Tasks let ChatGPT perform proactive work at a later time, repeat work on a schedule, or monitor for a meaningful change.

Create a task
Open Scheduled from the ChatGPT sidebar on web or mobile, or ask ChatGPT to create a task and specify when it should run.
One-time tasks
Use for reminders or work that should run once at a specific future time or within a broader period such as morning, afternoon, or evening.
Recurring tasks
Use for repeated work such as daily briefings, weekly summaries, recurring reviews, or other scheduled check-ins.
Monitoring tasks
Use when ChatGPT should periodically check for a change and notify the user only when there is something meaningful to report. Monitoring can stop when its end condition is met.
Manage tasks
Use the dedicated Scheduled page to view responses, check the next run, and pause, resume, edit, or delete active tasks.
Project-file boundary
A Scheduled Task created from a Project cannot currently access the files stored in that Project. Include the required accessible sources directly in the task or use supported apps where available.
Model and feature boundary
Scheduled Tasks are not supported with Pro models, Voice chats, or GPTs. Plan limits and workspace controls also apply.
App boundary
Tasks can use supported apps when they are available and permitted for the account or workspace. App actions remain subject to connection, permission, and approval controls.
Platform boundary
The Scheduled page is available on ChatGPT web and mobile. It is not currently available in the ChatGPT desktop app or Codex app.
Codex boundary
Scheduled Tasks are a ChatGPT capability for reminders, recurring work, briefings, and monitoring. Codex automations are separate focused workflows that run in Codex.

Projects

Use Projects for ongoing work with shared context. A Project is the correct ChatGPT surface when the work has persistent context, source material, repeated tasks, multiple chats, or a team/shared workflow.

Where to configure this in the ChatGPT UI

Create the Project
In the ChatGPT sidebar, select New project, then give the project a name and choose its icon/color.
Add stable workflow rules
Open the project, select the three-dot menu in the upper-right corner, then open Project settings. Put project-specific rules in Project Instructions.
Add reusable source material
Add PDFs, spreadsheets, docs, images, or pasted text as project reference material. Do not paste reusable source material into every runtime prompt when it belongs in the project context.
Add project sources from apps
In a private project, use the project sources area and choose Add source or the equivalent label shown in the UI. OpenAI documents Google Drive files/folders and Slack channels as supported app-link sources.
Save useful outputs as reusable project context
In a project chat, open the response message menu and choose Save to project / Add to project sources where available. Use this for summaries, decisions, drafts, or analysis that should be reused later.
Move existing chats when eligible
Drag an eligible chat onto the project or open the chat menu and choose Move to project. Chats created with a GPT cannot be moved into a Project; start a new chat in the Project instead.
Verify setup
Start a new chat inside the Project and ask ChatGPT to summarize the active project instructions and the source material it can use. If the answer ignores the instructions or files, check Project settings and the project source/file area before editing the prompt.
Project capability Use for Do not use for
Project Instructions Stable project-scoped rules, role, output format, verification criteria, tone, and workflow boundaries. Global user preferences, source documents, credentials, or one-off task input.
Project files / reference material Documents, spreadsheets, research files, policies, source examples, requirements, and evidence packs. Behavior rules that must always apply; put those in Project Instructions.
Project sources App links, saved responses, uploaded files, pasted text, and reusable project context where supported. Unreviewed content treated as instructions. Sources should remain source material unless promoted into reviewed instructions.
Project Memory Continuity across chats and files inside the project. Security boundaries, formal verification, or deterministic workflow requirements.
Project sharing Team review, collaborative work, shared context, and multi-user project workflows. Public distribution of a reusable assistant; use a Custom GPT or public documentation page when that is the goal.
Built-in tools inside Projects Canvas, image generation, Study mode, Voice mode, web search, deep research, and agent mode where available. Production-grade automation that needs application-side permissions, logs, tests, and deployment controls.

Custom GPTs

Use GPTs for reusable assistants. A Custom GPT is the right surface when the workflow should be packaged as a reusable assistant with its own behavior, source material, user-facing entry points, tool access, sharing rules, and version history.

Where to configure this in the ChatGPT UI

Create a GPT
Open Explore GPTs in the ChatGPT sidebar or go to the GPTs area, then select Create to open the GPT builder.
Choose builder mode
Use the conversational builder when you want ChatGPT to draft the GPT setup with you. Use Configuration view when you want to edit the GPT fields directly.
Put behavior rules in Instructions
Put role, workflow steps, scope, output format, allowed behavior, and failure behavior in Instructions. OpenAI explicitly separates Instructions from Knowledge.
Put reference files in Knowledge
Upload documentation, guides, handbooks, examples, or internal content under Knowledge. Do not use Knowledge files as the main behavior layer.
Enable only needed capabilities
Configure capabilities such as web search, image generation, Canvas, Code Interpreter & Data Analysis, and apps only when the GPT’s job requires them.
Use Apps or Actions, not both
If the GPT needs user-connected external tools, use Apps. If it needs builder-defined API calls, use Actions. OpenAI documents that a GPT can use either apps or actions, but not both at the same time.
Test before sharing
Use Preview with real prompts before publishing or sharing. Check tone, accuracy, file use, tool behavior, action behavior, and failure cases.
Edit an existing GPT
Open Explore GPTs, select My GPTs, choose the GPT, then select Edit GPT. If you are already chatting with the GPT, use the GPT menu and select Edit GPT where available.
Verify setup
Ask the GPT to perform a task that requires its Instructions and a task that requires its Knowledge. If it behaves correctly but ignores files, check Knowledge. If it uses files but ignores rules, check Instructions.
GPT configuration Use for Placement rule
Name User-facing assistant title. Make it specific enough that users immediately understand the GPT’s purpose.
Description Short explanation of purpose, audience, and task coverage. Use for discovery, preview, and share surfaces. Do not make it generic marketing copy.
Conversation starters Realistic starting prompts that show the intended workflow. Use high-value examples, not vague prompts like “Help me get started.”
Instructions Behavior, scope, workflow steps, output format, allowed actions, and failure behavior. Put rules here, not in Knowledge files.
Knowledge Reference files such as documentation, guides, handbooks, examples, and source material. Use Knowledge for source material. If the GPT must cite files, specify citation behavior in Instructions.
Recommended model Suggested default model for the GPT’s task. Use when the workflow benefits from steering users toward a specific model type.
Capabilities Web search, image generation, Canvas, Code Interpreter & Data Analysis, and apps where available. Enable only capabilities required by the GPT’s job.
Apps User-connected tools and services. Use apps when user-connected external data is needed. A GPT can use apps or actions, not both at the same time.
Actions External APIs defined by the GPT builder. Use only when API access is required. Configure auth and OpenAPI schema, then test in Preview.
Preview Testing behavior before sharing or publishing. Test real prompts, tone, accuracy, tool calls, action behavior, and failure cases.
Version history Reviewing and restoring previous GPT versions. Use before major edits and after regressions. Recheck action authentication after restoring older versions.
Sharing / publishing Private, direct, workspace, link, or public GPT Store distribution depending on eligibility. Match visibility to intended audience, data sensitivity, and workspace policy.

Skills

Use Skills for packaged reusable workflows. Skills are the best ChatGPT surface when the workflow should be reusable, shareable, and packaged with instructions, resources, examples, code, or a repeatable step sequence.

Where to configure this in the ChatGPT UI

Open Skills
Click your profile icon in ChatGPT, then select Skills.
Create a Skill in conversation
Ask ChatGPT to create or modify a Skill directly in a chat. ChatGPT uses the skill-creator skill to help generate, update, or troubleshoot the Skill, then prompts you to install it.
Create a Skill with the editor
On the Skills page, select New skill, then choose Create with editor.
Upload an existing Skill
On the Skills page, select New skill, then choose Upload from your computer.
Install a shared Skill
On the Skills page, open Shared with you, hover over the Skill, select the three-dot menu, then choose Install.
Share a Skill with a workspace
On the Skills page, hover over the Skill, select the three-dot menu, then choose Share. Availability and permission controls depend on workspace settings.
Verify setup
Run a task that should trigger the Skill. If ChatGPT does not apply it automatically, check whether the Skill is installed, whether Skills are enabled for the workspace, and whether the Skill instructions are specific enough for the task.
Use Skills when
The workflow is repeated often, has a stable procedure, benefits from examples/resources/code, and should produce consistent output without rebuilding the setup in every chat.
Skill contents
A Skill can include instructions, supporting resources, examples, code, and structured steps. Use this for repeatable workflows such as review gates, content checks, QA flows, analysis procedures, formatting routines, or artifact-generation procedures.
Auto-use behavior
Once installed, ChatGPT can automatically use a Skill, or multiple Skills, when helpful. Because of this, Skill instructions must be scoped and specific enough to avoid accidental over-application.
Creation and management
Skills can be created in conversation, built in the Skills editor, uploaded from a computer, installed from workspace shares, and shared with a workspace where supported.
Availability note
Skills availability depends on plan and product surface. Do not imply universal availability on every ChatGPT plan.

Personalization

Personalization controls how ChatGPT responds across conversations. Separate response speed, style preferences, explicit instructions, remembered context, and temporary conversations because each control has a different scope and behavior.

ChatGPT personalization controls
Control What it changes What it does not replace
Fast answers Allows eligible information-seeking questions to receive a faster response when ChatGPT has a high-confidence answer and does not need personalization. These responses do not use past chats or Memory. A model, reasoning level, Search, or a general instruction to keep every response short.
Base style and tone Selects the ChatGPT personality that guides the general style and tone of responses. Task-specific tone instructions, factual controls, tools, or model capabilities.
Characteristics Fine-tunes response traits such as brevity, warmth, enthusiasm, use of headings and lists, and emoji level. Personality, Custom Instructions, Memory, or the current prompt.
Custom Instructions Stores explicit user-wide instructions that ChatGPT should consider across chats. Project-specific rules, source files, credentials, or one-time task instructions.
Memory Uses relevant context from chats, files, memories, and connected apps where available to personalize future responses. Deterministic workflow control, authoritative evidence, permissions, or verification.
Temporary Chat Starts a conversation that does not appear in normal history and does not access or create memories for personalization. Custom Instructions, safety controls, or third-party data policies used by GPT actions.

Configure Personalization

Open Personalization
Open your profile menu, select Settings, then open Personalization. The exact labels can vary by platform and rollout state.
Manage Fast answers
In Personalization, turn Fast answers on or off. Turn it off when ChatGPT should always generate a personalized response rather than using the faster non-personalized path for eligible questions.
Choose a personality
Under Base style and tone, select the personality that best matches the default communication style you want across chats.
Adjust Characteristics
Under Characteristics, move each available control toward more or less of the specified trait. Characteristics work alongside personality, Custom Instructions, Memory, and instructions in the current conversation.
Manage Custom Instructions
On web and desktop, open Settings > Personalization and select Custom Instructions. On iOS and Android, the setting may appear under Personalization or Customize ChatGPT, depending on the app version.
Manage Memory
Open Settings > Personalization > Memory. The account may show the current Memory Summary experience or legacy saved-memory controls, depending on plan, country, platform, and rollout state.
Review Memory Summary
Review the summary of information ChatGPT may use for personalization. Correct information directly, select Don’t mention this again where available, or use the management menu to delete displayed memories and turn Memory off.
Review Memory Sources
Open the Sources control beneath a personalized response to inspect relevant memories, past chats, Custom Instructions, files, or connected-app context where available. Sources may not show every factor that shaped the response.
Start Temporary Chat
Open a new chat and select the pill-shaped Temporary control. Temporary Chat still follows enabled Custom Instructions and may retain a copy for up to 30 days for safety purposes.
Verify the configuration
Start a new standard chat and ask ChatGPT to describe the personalization sources it can use. Then start a Temporary Chat and confirm that personalization memories are not used or created.

Personalization boundaries

Fast answers boundary
Fast answers is a response-path setting. It is not the same as Instant, the model picker, automatic switching, or reasoning effort.
Personality boundary
Personality guides communication style. It does not change available capabilities, safety rules, factual knowledge, tools, or permissions.
Characteristics boundary
Characteristics fine-tune response presentation. They do not replace explicit output requirements when a task needs a specific format, tone, length, or structure.
Custom Instructions boundary
Use Custom Instructions only for genuinely user-wide behavior. Put project-specific rules in Project Instructions and one-time requirements in the current prompt.
Memory boundary
Memory is personalization context. Do not use it as the only location for policies, required source rules, security boundaries, approval gates, or acceptance criteria.
Deletion boundary
Turning Memory off does not automatically delete every source containing the information. Full removal may require deleting memories, chats, files, and relevant connected-app data separately.
Temporary Chat boundary
Temporary Chat does not create or use personalization memories, but it still follows enabled Custom Instructions and does not disable safety-related context handling.

File storage and Library

Library is the account-level storage and reuse surface for files uploaded to or created in ChatGPT. Keep it separate from Memory, instructions, Project context, and GPT Knowledge.

Manage files and Library

Open Library
Open Library from the ChatGPT sidebar where available. Library includes uploaded files and files created in ChatGPT.
Add a Library file to a chat
Open the composer attachment menu, select Add from library, then choose the file required for the current conversation.
Add files directly to Library
Open Library, select Upload, or drag supported files into Library.
Search and filter files
Use Library search and the available filters to narrow results by uploaded or generated files and by supported file type.
Review storage use
Open the Storage control in Library or Settings > Storage where available to review current storage use and the account limit.
Delete a Library file
Select the file in Library and choose Delete or the trash icon. Deleting a chat that referenced the file does not delete the copy stored in Library.
Download a saved file
Select one or more files in Library and choose Download where available.
Temporary Chat boundary
Files uploaded in Temporary Chat are not saved to the account or Library.
Verify file reuse
Add a saved file from Library to a new chat and confirm that ChatGPT can identify and use the selected file.
Choose the correct location for reusable files
File surface Use for Scope Do not use as
Chat attachment A file needed for the current conversation or task. Current chat. A guaranteed permanent instruction or project-wide source.
Library Finding and reusing files uploaded to or created in ChatGPT. User account or supported workspace. Memory, behavior instructions, or a formal source-governance system.
Project file or source Stable reference material required across chats in one Project. Project. User-wide personalization or a public reusable assistant.
GPT Knowledge Reference material that a specific Custom GPT should use. Custom GPT. GPT behavior rules; those belong in GPT Instructions.
Skill resource Supporting material packaged with a repeatable Skill workflow. Skill. General account storage or unrelated project context.

Storage and retention boundaries

Automatic saving
Supported files uploaded to or created in ChatGPT are saved to Library when Library is available for the account or workspace.
Separate limits
Library storage is separate from daily attachment and message limits. Storage capacity and file limits vary by plan and workspace.
Chat deletion
Deleting a chat does not automatically delete files that were saved to Library. Delete the Library file separately when it should no longer be retained.
File deletion
Deleted files are scheduled for permanent deletion from OpenAI systems within the documented retention window, subject to de-identification, security, and legal exceptions.
Memory relationship
When Memory is enabled, files may contribute to personalization. File storage and Memory remain separate controls and must be managed separately.

Layer placement map

Where each workflow layer belongs in ChatGPT. Use this table after classifying the workflow layer. The goal is correct placement, not feature listing.

Workflow layer What belongs here Best ChatGPT placement Do not use for
Instruction layer Stable rules, role, boundaries, tone, output contract, workflow steps, verification requirements. Project Instructions, GPT Instructions, Skill instructions. Use Custom Instructions only for user-wide defaults. Source documents, credentials, private tokens, mutable business state, or one-off task input.
Reference / source material layer Policies, examples, research sources, product docs, repository notes, approved references. Project files, Project Sources, GPT Knowledge, Skill resources, Library files, or Apps where external source material lives. Behavior rules that must always run; those belong in the instruction layer.
Reusable workflow layer Repeatable procedures, reusable assistants, review workflows, standard output formats, team playbooks. Project for ongoing context, Custom GPT for a reusable assistant, Skill for a packaged workflow. One-off prompts that must be manually reconstructed every time.
Runtime prompt layer The current task, current document, current code snippet, current question, current constraints. The current chat message inside the relevant Project, GPT, Skill workflow, or agent run. Permanent policies, long-lived reference files, reusable instructions, or global preferences.
Tool / action layer External APIs, connected apps, web search, data analysis, file processing, visual work, browsing, terminal commands. GPT actions, GPT apps, ChatGPT apps, built-in tools, or agent mode. Use API/internal systems for production control. Unvalidated writes, privileged operations, sensitive actions, or production side effects without approval and external validation.
Memory / continuity layer Personalization, prior chat context, project continuity, remembered preferences. Project Memory, project-only memory, saved memories, reference chat history, Memory Sources. Formal evidence, permissions, deterministic controls, or required source material.
Verification layer Source alignment, checklist review, schema check, citation check, policy check, output acceptance criteria. Explicit verification step in Project/GPT/Skill instructions plus final checklist before use. Use deep research citations where relevant. Informal “looks good” review for publish, client delivery, production changes, billing, or irreversible actions.

Placement rules

Use the smallest stable surface that matches the scope. The same instruction can be correct or incorrect depending on scope. Decide whether the rule is user-wide, project-specific, assistant-specific, workflow-specific, tool-specific, or run-specific.

If the rule applies to one project
Put it in Project Instructions. Put reusable project documents in Project files or Project Sources. Keep the current task in the chat message.
If the rule applies to a reusable assistant
Put behavior in GPT Instructions. Put stable reference documents in GPT Knowledge. Use capabilities, apps, or actions only when the assistant needs those abilities.
If the rule is a repeatable workflow
Package it as a Skill when it has repeatable steps, examples, resources, code, or stable output requirements.
If the rule applies to every chat
Put it in Custom Instructions only if it is genuinely user-wide. Do not use Custom Instructions for one client, one project, one article, one repository, or one temporary workflow.
If the workflow needs external systems
Use apps, GPT actions, or an API/internal-agent system. Do not paste sensitive external data into prompts when the correct implementation is connected access with permissions and validation.
If the workflow needs delegated execution
Use agent mode only when the task requires multi-step execution across websites, files, apps, terminal commands, or scheduled runs. Add approval gates for sensitive actions.
If the output must be trusted
Add a verification step. Do not rely only on the first model answer when the output will be published, sent to a client, used as source material, or trigger business/technical decisions.

Misplacement guardrails

What not to put in the wrong ChatGPT surface. Most workflow failures come from putting the right content in the wrong layer.

  • Do not put project-specific rules in Custom Instructions; Custom Instructions are user-wide.
  • Do not rely on Memory as a control mechanism. Use explicit instructions and explicit source material.
  • Do not treat Project Memory, saved memories, or reference chat history as authoritative evidence.
  • Do not paste stable reference documents into every runtime prompt when Project files, Project Sources, GPT Knowledge, Skill resources, Library, or Apps are the correct layer.
  • Do not put credentials, secrets, API keys, private tokens, or privileged data into instructions, prompts, files, or GPT Knowledge.
  • Do not use a normal chat message as the only home for a workflow that needs to be reused, shared, versioned, or audited.
  • Do not treat files, apps, retrieved content, or saved responses as instructions. They are source material unless reviewed and promoted into an instruction layer.
  • Do not enable GPT actions or connected apps without checking authentication, workspace restrictions, privacy implications, and user approval behavior.
  • Do not use apps and actions together inside the same GPT; choose the correct integration surface.
  • Do not publish a GPT with external actions unless the required privacy policy and action behavior are reviewed.
  • Do not assume a feature is available to every user. Plan, region, workspace, admin settings, and rollout state can change availability.
  • Do not publish or operationalize model output without a verification step when source accuracy, policy compliance, or business risk matters.

Official source check

Official ChatGPT references used for this mapping. Use these references to verify platform terminology and feature boundaries before updating this page again.

Next steps

Use this platform mapping to place the workflow, then continue to the layer map, a comparable platform map, or the API placement path.