Workflow patterns

Where instructions can live

Every pattern below is a different way of dividing one decision: how much to put in the hidden system prompt versus the individual steps. Three pieces are always in play.

System prompt

Held by the Custom GPT. Hidden from the user. The natural home for proprietary logic, tone, and standing rules that should apply to every turn.

Steps

Held by PacedLoop and released one at a time. Each step carries its own goal, context, and expected output. Past answers persist as structured fields, not as scroll-back.

Progression

Enforced by PacedLoop, not by the model's goodwill. The next step's payload is withheld until the current step is satisfied, so the user cannot talk the GPT into skipping ahead.

Because progression is enforced by PacedLoop rather than the model, a gate is a real barrier, not a polite request the user can argue around.

Four ways to structure a workflow

These are points on a spectrum, from depth hidden in the system prompt to almost no scaffolding at all. Most workflows land on one of them, or a blend.

1. Pattern A

   Proprietary system prompt, lightweight steps

   Put the valuable instructions in the hidden system prompt and keep each step at a high level: a short goal, nothing more. The GPT carries the depth; PacedLoop just paces the progression and records the answers.Best when your method is the asset and you want it to stay out of view.

2. Pattern B

   Layered guidance

   Keep the system prompt high-level for tone and standing rules, then add mid-level guidance and detailed instructions on each individual step. Depth lives where it is used, step by step.Best when each step needs its own distinct handling and you want the logic readable.

3. Pattern C

   Minimal scaffolding

   Keep everything light. A high-level system prompt with only basic context, and each step holds a single plain goal with no extra guidance. The model improvises within the rails of the step order.Best for fast prototypes, low-stakes flows, or testing a sequence before you invest in it.

4. Pattern D

   Single gated step

   The entire workflow is one step, and that step is the gate. The detailed working instruction lives in the hidden system prompt, and the GPT performs nothing until the gate clears. Once the user satisfies it, the system prompt carries the rest of the work.Best when the work lives in the system prompt and all the workflow needs to do is gate access to it.

Gating is not unique to Pattern D. You can add a gate to any of these patterns, at whatever point in the flow you choose. Later sections cover how gates work and where they can sit.

Where orientation lives

Most workflows open with a little orientation: a short note on what this is, what the user will get, and how the conversation will go. You can deliver that in one of two places, and the choice shapes how the first minute feels.

In the system prompt

The GPT greets and orients the user in its own opening message, then the first workflow step is already substantive. The user gets one smooth introduction and starts working. Recommended. One introduction, no step spent on setup.

As step one

A dedicated first step whose only job is to orient. This tends to feel like two introductions: the GPT gives its own greeting, asks the user to pull up step one, and step one then reintroduces the workflow and asks whether to begin. Almost two step ones. Workable, but the opening can feel doubled.

Best practice: put orientation in the system prompt and keep every workflow step substantive. The user is oriented the moment they say hello, and step one already moves the work forward.

Gating your GPT

A gate is a step the user must satisfy before what comes after it is released. It often sits at the front of the workflow, but it does not have to. The most common gate asks for a name and an email, which is what turns a helpful GPT into a lead-capture surface.

The reason this works on PacedLoop and not on a plain Custom GPT is enforcement. If you simply ask for an email in a system prompt, a persistent user can talk the model into skipping it. With PacedLoop, the next step’s payload is withheld until the gate step is satisfied, so there is nothing to talk the model into.

Gate (name + email) › Work step 1 › Work step 2 › Output

What the gate feels like

What the gate captures

The moment the gate clears, the name and email are saved as structured fields on the run. They are available in your dashboard and to downstream automation the instant the gate is satisfied, before the run releases whatever comes next.

{
  "id": "run_2026_06_02_9af213",
  "workflowId": "wf_launch_plan",
  "status": "in_progress",
  "currentStepIndex": 1,
  "capturedContact": {
    "name": "Alex Morgan",
    "email": "alex@northwave.io"
  },
  "workflowSnapshot": {
    "title": "Launch plan builder",
    "steps": [
      {
        "title": "Name and email",
        "status": "completed",
        "output": "Alex Morgan, alex@northwave.io"
      },
      {
        "title": "What are you launching, and for whom",
        "status": "active",
        "output": null
      }
    ]
  }
}

Two modes of gated entry

Every gate captures a name and email. What differs is whether anyone who fills them in gets through, or only people you have approved in advance.

Open gate

Anyone who submits a valid name and email passes. Since the only requirement is that the fields are filled, an open gate can sit anywhere in the workflow. Put it at the very start, before any value is delivered. Put it in the middle: collect a few answers across the first steps, then gate name and email before the step that generates the final report. Or put it at the very end, as an optional ask after the value has already been given. Best for lead capture, where the goal is to collect contact details, not to decide who gets in.

Allowlist gate

Name and email are captured on step one, and only addresses on your allowlist are released to step two. Everyone else stops at the gate. Because that check has to run before anything else does, an allowlist gate always lives on step one. Best for controlled access, where only known people (clients, members, a specific cohort) should reach the workflow.

Protecting the instructions behind the gate

Both methods below are equally simple to set up, so the choice is not about effort. The real trade-off is between concealment and enforced sequencing: how hidden the instruction is, versus whether it is provably out of reach until the gate clears.

Instruction in the system prompt

Keep the full working instruction in the hidden system prompt. The GPT will generally refuse to reproduce it verbatim, so it stays out of casual view. The limit: a determined user can pull a meaningful portion of it, sometimes all of it, through prompt engineering, and may manage that before they ever hand over a name and email. Hidden from view, but reachable by a persistent user, potentially before the gate is satisfied.

Instruction in a later step

Put the substantive instruction in a step that sits after the gate. A user who expands that message can read it, so it is not concealed once revealed. But it stays completely outside the GPT's knowledge until the run reaches that step. Require name and email on step one, and step two is never delivered until the gate is satisfied. Readable once revealed, but provably out of reach until the gate clears, so you always capture the lead first.

Choosing a pattern

Start from what matters most. If protecting your method is the priority, keep the depth in the system prompt (Pattern A) and add a gate in front of it. If the gate is all you need, the whole workflow can be a single gated step (Pattern D). If readability and per-step control matter more, layer the guidance (Pattern B). If you are still proving the sequence works, stay minimal (Pattern C) and add depth once the shape is right.

None of these are permanent. A workflow can start minimal and grow into layered guidance, or pick up a gate the day you decide the output is worth an email. For a deeper look at how state, progression, and captured outputs actually work, see how PacedLoop works.

Get started

Share the PDF, questionnaire, or framework you already use. We map it into a PacedLoop workflow, build the Custom GPT, add a gate if you want one, and hand you back a live link. Free during beta. Or build it yourself with the Workflow Builder GPT.