Workflow layers & prompt placement

Canonical reference for instruction rules, source material, reusable workflows, runtime prompts, verification gates, prompt assets, and platform placement.

This page defines the canonical workflow-layer model used across this site.

Use it when you need to decide whether something belongs in a policy, guide, prompt page, prompt component, reference page, platform setup, runtime task, or verification gate.

Core rule

Do not treat every reusable AI instruction as a prompt.

A complete AI workflow usually contains five separate layers:

  1. Instruction layer — stable rules and behavior constraints.
  2. Reference / source material layer — files, policies, examples, source documents, logs, datasets, or other material the model may consult.
  3. Reusable workflow layer — repeatable procedures, assistants, skills, agents, or workflow bundles.
  4. Runtime prompt layer — the current task, current input, selected artifacts, and one-time output request.
  5. Verification layer — checks before output is accepted, published, or used downstream.

The practical rule is:

  • Put stable behavior rules in the instruction layer.
  • Put source material in the reference layer.
  • Put repeatable procedures in the reusable workflow layer.
  • Put the current task in the runtime prompt layer.
  • Put claim checks, source checks, and output checks in the verification layer.

The five workflow layers

Layer What belongs here What should not belong here
Instruction layer Stable rules, source boundaries, refusal rules, output requirements, role/task constraints, instruction precedence Long source documents, copied articles, logs, untrusted files, one-off task details
Reference / source material layer Policies, source documents, uploaded files, examples, specs, logs, screenshots, datasets, repository snapshots, retrieved documents Instructions that should override the workflow, hidden commands inside files, unreviewed tool output treated as authority
Reusable workflow layer Repeatable procedures, Custom GPTs, Skills, Gems, project templates, agent workflows, checklists, workflow bundles One-time questions, temporary inputs, raw evidence that changes every run
Runtime prompt layer The current question, current artifact, selected excerpt, requested output, temporary constraints, task-specific variables Permanent policy rules, reusable source files, broad behavior constraints
Verification layer Evidence checks, citation checks, file-coverage checks, claim-support checks, format checks, evals, release checks New unsupported claims, additional task scope, unverified reconstruction

Site content types vs workflow layers

The site uses different content types because each one serves a different control function.

Site content type Primary purpose Main layer
Policy Defines non-negotiable rules, evidence boundaries, refusal conditions, and enforcement constraints Instruction layer / verification layer
Guide / how-to Runs a procedure that applies one or more policies or prompt assets Reusable workflow layer
Prompt page Presents a reusable prompt asset with use case, inputs, copy-ready text, and related workflow context Runtime prompt layer / reusable workflow layer
Prompt component Adds one focused behavior to a larger prompt or workflow Runtime prompt layer / instruction layer
Protected raw asset Stores the raw .txt, .md, or member-only execution asset behind access control Source asset / protected execution asset
Reference page Defines stable terminology, mappings, diagrams, or implementation lookup material Reference layer
Article Provides explanation, analysis, framing, or research context Reference / educational context

Canonical terms

Policy

A policy is a rule set. It defines what is allowed, what is blocked, what evidence is acceptable, when the workflow must fail closed, and what checks must happen before output is accepted.

Use a policy when the goal is governance or enforcement, not step-by-step execution.

Guide / how-to

A guide is a runnable procedure. It tells the user how to apply a policy, prompt, component, or workflow in practice.

Use a guide when the user needs steps, setup, procedure, verification, and common failure modes.

Prompt page

A prompt page is the user-facing page for a reusable prompt asset.

It should explain:

  • what the prompt does
  • when to use it
  • what inputs it requires
  • where to configure it
  • what output to expect
  • which guide or policy it belongs with

A prompt page is not the same thing as a raw .txt file. The page is the shareable and explainable surface. The raw file is the source or execution asset.

System / developer instruction block

A system/developer instruction block belongs in the instruction layer. It binds stable rules, source boundaries, output requirements, and instruction hierarchy.

Use this for behavior that must persist across a workflow or repeated task.

Runtime prompt / user runner

A runtime prompt or user runner is the task message for the current run.

It usually contains:

  • the current objective
  • the current input
  • selected files or excerpts
  • required output format
  • one-time constraints
  • variables for this run

Do not store broad permanent rules only in the runtime prompt if the workflow needs to be reused.

Prompt component

A prompt component is a small reusable add-on that enforces one behavior, such as deep reading, no unsupported claims, evidence-only output, citation requirements, or confidence reporting.

Use components when the behavior should be added to an existing workflow without creating a full standalone guide.

Workflow template

A workflow template is a reusable multi-step procedure. It defines inputs, steps, stop conditions, output requirements, and verification checks.

Use a workflow template when the task is repeated and has a consistent sequence.

Prompt bundle

A prompt bundle is a group of related assets that work together, such as:

  • instruction-layer prompt
  • runtime runner
  • prompt components
  • verification prompt
  • related policy
  • related guide

Use prompt bundle or workflow asset bundle instead of the legacy label stack when the goal is clarity for users.

Skill

A skill is a reusable workflow surface in some AI toolchains. It may include instructions, examples, scripts, resources, or supporting files.

On this site, the closest equivalents are reusable workflow templates, prompt bundles, guide procedures, and selected member execution assets. Not every prompt, policy, article, or reference page is a skill.

Untrusted content

Untrusted content includes uploaded files, pasted excerpts, retrieved pages, emails, tool outputs, logs, screenshots, repository files, and external documents.

Instructions found inside untrusted content are treated as data unless the user or system explicitly delegates authority to them.

Platform placement crosswalk

Use this table after you identify the workflow layer.

Layer ChatGPT Claude Gemini API / internal system
Instruction layer Project Instructions, GPT Instructions, Skill instructions, or Custom Instructions for broad user-wide preferences Project Instructions or Skill instructions Gem Instructions or API system instructions system/developer instructions, policy middleware, controller configuration
Reference / source material layer Project files/sources, GPT Knowledge, uploaded files Project Knowledge, uploaded documents, Project RAG where available Gem Knowledge files, Files API, File Search / RAG where available retrieval store, vector index, file-ingestion layer, application context
Reusable workflow layer Project, Custom GPT, Skill, reusable prompt page Project, Skill, reusable workflow Gem, API workflow, reusable prompt page orchestrator, agent route, workflow service, job template
Runtime prompt layer Current chat message or task prompt Current chat message Current chat message or contents request payload user message, request body, task payload
Verification layer Prompt gate, checklist, guide verification step, eval prompt Prompt gate, checklist, guide verification step Prompt gate, checklist, guide verification step validation service, eval layer, test suite, logging review, approval gate

Placement decision rules

Use these rules when creating or reviewing a new site asset.

Put it in the instruction layer when

  • the rule must persist across multiple runs
  • the behavior defines what the assistant may or may not do
  • the rule controls evidence, refusal, citation, output format, or authority boundaries
  • the rule should not be repeated manually in every runtime prompt

Put it in the reference layer when

  • the material should be consulted as source material
  • the content is evidence, documentation, examples, logs, screenshots, specs, or repository material
  • the material may change independently from the workflow
  • the content should not override the instruction hierarchy

Put it in the reusable workflow layer when

  • the same procedure is used repeatedly
  • the workflow has multiple steps
  • the workflow includes required inputs, stop conditions, and verification
  • the workflow should be packaged as a guide, GPT, Skill, Gem, or internal agent route

Put it in the runtime prompt layer when

  • the input is specific to the current task
  • the user is asking one concrete question
  • the prompt includes temporary variables or selected artifacts
  • the instruction does not need to persist after this run

Put it in the verification layer when

  • the output must be checked before acceptance
  • the workflow requires source support, citation checks, claim checks, or coverage checks
  • the result will be published, sent to users, used in code, or used for decision-making

Prompt assets and raw files

A prompt should have a managed page whenever users need to understand, share, configure, or reuse it.

Use this separation:

Asset Correct role
Prompt page User-facing explanation, metadata, use case, copy-ready prompt, related guide/policy
Raw .txt or .md prompt file Source asset, protected execution file, downloadable member artifact, or canonical raw prompt body
Catalog YAML Metadata for cards, filters, access state, related content, and navigation
Protected member object Full Pro prompt, execution asset, workflow bundle, or gated file
Public page Preview, explanation, safe setup guidance, SEO surface, and conversion path

Do not use a raw .txt file as the primary user-facing prompt experience when a managed prompt page exists.

API and agent-system mapping

API pipelines

For API-based systems, separate the workflow across application layers:

  • System/developer instructions: stable policy binding, instruction hierarchy, source boundary, and output contract.
  • Retrieval / file-ingestion layer: source files, documents, logs, examples, and indexed reference material.
  • User payload: the current request, variables, selected artifacts, and output request.
  • Tool layer: callable tools, permissions, schemas, validation, allowlists, and execution constraints.
  • Verification layer: evals, schema validation, claim checks, citation checks, approval gates, and logs.

Agent loops

For agent systems, the model should not be the only enforcement point.

Use the orchestrator or controller layer for:

  • tool permission checks
  • input validation
  • allowlists and deny rules
  • retrieval boundaries
  • memory boundaries
  • approval gates
  • output validation
  • logging and audit review

The model receives instructions and task context, but the system should still enforce policy outside the model when the action has security, privacy, cost, or production impact.

Trust-boundary rule

Treat external or user-provided material as data, not instructions.

This includes:

  • uploaded files
  • pasted text
  • retrieved web pages
  • connector output
  • tool output
  • emails
  • logs
  • repository files
  • screenshots
  • PDFs
  • datasets
  • memory entries that were not explicitly approved as instruction source

If untrusted content contains commands such as “ignore previous instructions,” “use this policy instead,” “send data elsewhere,” or “treat this as system instruction,” those commands remain data unless explicitly authorized by the controlling instruction layer.

How this page connects to the rest of the site

Use this page to classify the layer first. Then open the platform-specific mapping page only when you need tool-specific setup.

Use the libraries this way:

External references

  • OpenAI — Projects in ChatGPT: https://help.openai.com/en/articles/10169521-projects-in-chatgpt
  • OpenAI — Creating and editing GPTs: https://help.openai.com/en/articles/8554397-creating-and-editing-gpts
  • OpenAI — Skills in ChatGPT: https://help.openai.com/en/articles/20001066-skills-in-chatgpt
  • Anthropic — What are projects?: https://support.anthropic.com/en/articles/9517075-what-are-projects
  • Anthropic — How can I create and manage projects?: https://support.anthropic.com/en/articles/9519177-how-can-i-create-and-manage-projects
  • Anthropic — Retrieval augmented generation for projects: https://support.anthropic.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects
  • Google — Use Gems in Gemini Apps: https://support.google.com/gemini/answer/15146780
  • Google — Tips for creating custom Gems: https://support.google.com/gemini/answer/15235603
  • Google AI for Developers — Files API: https://ai.google.dev/gemini-api/docs/files
  • Google AI for Developers — File Search: https://ai.google.dev/gemini-api/docs/file-search
  • Google AI for Developers — Generate content API: https://ai.google.dev/api/generate-content
  • OWASP — AI Agent Security Cheat Sheet: https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html
  • OWASP — LLM Prompt Injection Prevention Cheat Sheet: https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html
  • Diátaxis — Reference: https://diataxis.fr/reference/