Build Your First AI Workflow: A Step-by-Step Guide Using Our Real Stashed Example

There is a difference between using AI and having an AI workflow.

Using AI looks like this: you have a problem, you open Claude, you write something, you get a response, you copy some of it, you close the tab. Useful. Slightly random. Hard to repeat.

Having an AI workflow looks like this: every time a specific task comes up, you run a documented process - the same inputs, the same prompt structure, the same quality check - and you get a consistent output in a predictable amount of time. This is the foundation of what it means to be an AI operator - someone who runs entire business functions with AI rather than just using it occasionally.

The first is a trick. The second is a system.

Tricks save you time occasionally. Systems save you time every single time. And over a year, systems compound into a real operational advantage.

We run a full content week for Stashed in about 4 hours, using AI workflows we built and documented over about three months. If you want to understand the workflow behind Stashed as a brand-building story first, that post covers the full picture. Here is how to build your first workflow.

What Is an AI Workflow, Exactly?

An AI workflow is a documented, repeatable process that uses AI tools to produce a specific output from a specific input.

Key word: documented. If it only exists in your head, it is not a workflow. It is a habit - and habits break when you are sick, stressed, or handing work to someone else.

A workflow has:

• A clear trigger (what event starts this process)
• A defined input format (what information goes in)
• One or more AI tool interactions (what prompt, which tool)
• An output standard (what does “good” look like)
• A review step (how do you know it is ready)

Most people do the middle part - they have a prompt they like - and skip the surrounding structure. That is why their results are inconsistent. The prompt is only one component.

The 4-Step Workflow Framework

Every AI workflow we build follows this structure:

Step 1: Task - Define the output you need and the input you have. Be specific about both.

Step 2: Prompt - Write the instruction to the AI. The prompt should be templated, not written fresh each time.

Step 3: Output - Get the AI response. Do not use it raw.

Step 4: Refine - Apply your quality filter. Does this sound like us? Is the information accurate? Would we be embarrassed if this went out without changes?

That is it. Four steps. The sophistication lives inside each step, not in some complex flowchart.

Real Example: Our Stashed Instagram Caption Workflow

We post on Stashed’s Instagram account four to five times per week. That is 16-20 pieces of caption copy per month. Writing each one from scratch would take us 20-40 minutes each - 5-8 hours per month on one content task.

With our workflow, each caption takes about 8-12 minutes total. Here is the exact process.

Step 1: Input the Product Details

Before touching any AI tool, we fill in a standard input template. This template took us two weeks to refine and is now the single most important part of the workflow.

CAPTION INPUT TEMPLATE
------------------------
Product/subject: [What are we posting about? Bag, panel design, pack shot, lifestyle photo?]
Post type: [Product showcase / Story post / Community / Behind the scenes]
Key message: [One sentence. What do we want the viewer to feel or understand?]
Call to action: [What do we want them to do? Link in bio, DM, tag someone, nothing]
Tone: [Casual / Direct / Slightly funny / Emotional - pick one]
Any specific details to include: [A fact, a feature, a limited detail, anything specific]
Hashtag style: [Minimal (5-8) / Standard (15-20)]

We fill this in before opening Claude. This forces clarity. If you cannot fill in the “key message” field in one sentence, you do not know what the post is about yet. Figure that out before you ask the AI.

Step 2: Run the Caption Prompt

We have a master prompt that we paste into Claude with the filled template attached. The prompt does not change. The template input changes each time.

The prompt looks like this:

“You are writing Instagram captions for Stashed, a Made-in-India bag brand built by two 22-year-olds from Vaniyambadi, Tamil Nadu. Stashed bags have a removable Velcro front panel - you can swap designs. The brand voice is: direct, honest, slightly casual, never corporate, no buzzwords, no guru energy. The audience is Indian youth aged 18-28 who take public transport.

Using the inputs below, write 3 caption variations. Each should be different in approach (not just different words). Keep each caption under 150 words. Do not use emojis unless specifically asked. End each caption with the call to action from the inputs.

INPUTS: [paste filled template here]”

Three variations. We pick one. Sometimes we blend two. We rarely use any of them completely unchanged - but they are 80% of the way there and the remaining 20% takes two minutes.

Step 3: Review and Vibe-Check

This is the step people skip. Do not skip this step.

Before a caption goes out, we read it aloud. Actually aloud, not in our head. If we stumble on a word or it sounds like a brand deck instead of a person talking, we rewrite that line.

Our vibe-check questions:

• Does this sound like something we would actually say to a friend?
• Is there any word in here that would never appear in a text message? (Delete it.)
• Does it explain too much? (Instagram captions that explain are usually trying to justify the product, not show it. Cut.)
• Is the call to action natural or forced? (“Link in bio!” at the end of a poetic caption is jarring. Match the CTA energy to the caption energy.)

The vibe-check takes 3-4 minutes. It is the difference between content that performs and content that gets scrolled past.

Step 4: Schedule and Post

We use a scheduling tool (whatever works for your workflow - Buffer, Later, native Instagram scheduler). Approved captions go into the queue with the corresponding image.

We do this in batches. Every Monday morning, we fill out input templates for the week’s posts. We run all of them through the prompt workflow in one session. We vibe-check all of them. We schedule all of them.

Total time for a week of Stashed Instagram captions: 45-60 minutes.

We used to spend that much time on a single caption, second-guessing ourselves.

How to Document a Workflow So Anyone Can Reproduce It

The test of a workflow is: could someone who does not know your business run it and produce an output that meets your standard?

If the answer is no, it is not a workflow yet. It is notes.

Good workflow documentation has:

1. The trigger condition. “Run this workflow every time we need to create Instagram captions for a new post.”

2. The input template. Exactly as shown above - a form with specific fields, not a paragraph description.

3. The prompt, complete and unabridged. Paste the full prompt. Do not summarise it. If the prompt changes, update the document.

4. The output standard. “A caption under 150 words, in the Stashed voice, with the specified CTA, that passes the read-aloud test.”

5. The review checklist. The 4 vibe-check questions. Specific. Not “make sure it sounds good.”

6. The destination. Where does the output go? Scheduling tool, Google Doc, direct post?

We keep our workflow documentation in a shared Notion. Each workflow is one page. It takes about 30 minutes to write up a workflow properly after you have run it 5-6 times and know the kinks.

When to Use Claude vs ChatGPT vs Gemini

We get this question constantly. Here is our honest take.

Claude is what we use for most things. Brand voice work, longer-form copy, nuanced instructions, anything where we need the AI to actually understand context and constraints rather than pattern-match. Claude follows complex prompts more faithfully than the alternatives in our experience.

ChatGPT is fast and good for structured output tasks - tables, lists, extraction, simple rewrites. When we need a quick first draft of something with no voice requirements, ChatGPT is fine.

Gemini is useful when we are working inside Google Workspace (especially Sheets and Docs) because of the native integration. Also better than Claude for some image-to-text tasks in our experience.

The practical rule: use Claude for anything that requires brand voice, judgment, or complex instructions. Use the others for mechanical tasks where consistency is more important than nuance.

Our Full Content Week Workflow Stack

Here is the complete picture. This is how we run a full week of Stashed content in about 4 hours:

Total: roughly 4 hours. For five posts per week, one blog post, and a weekly review.

Before AI tools, this was a full-time job for one person. Now it is half a day.

Signs Your Workflow Is Working

You will know the workflow is working when:

Time saved is consistent. Not “sometimes fast, sometimes slow” - consistently fast. If some runs take twice as long as others, there is variability in your input template or prompt that needs fixing.

Quality is consistent. This is actually harder to achieve than time savings. Consistent quality means your output standard is well-defined and your review step is actually filtering.

You do not think about it anymore. The workflow runs on autopilot. You are not making creative decisions at the execution stage - you made them when you designed the workflow. Execution should be mechanical.

Your stress goes down. Content week used to give us mild anxiety every Monday. Now it is an admin task. That is the real sign a workflow is working.

Common Workflow Mistakes

Mistake 1: No output standard. “Good captions” is not an output standard. “Captions under 150 words, in the Stashed voice, that pass the read-aloud test” is an output standard. Vague standards produce vague results.

Mistake 2: Skipping the vibe-check. We have seen what happens when you publish AI-generated content without a proper review. It sounds like AI-generated content. Your audience can tell. The vibe-check is not optional.

Mistake 3: Over-automating. Not everything should be a workflow. One-off tasks, highly contextual decisions, anything that requires deep judgment - these should stay manual. Workflows are for repeatable, standardised tasks. Applying workflow thinking to everything is how you produce a lot of mediocre output very efficiently.

Mistake 4: Never updating the prompt. Your first prompt will not be your best prompt. Build in a weekly 15-minute review where you look at what the workflow produced and ask: what would make this 10% better? Update the prompt. Document the change and why you made it.

Mistake 5: Building the workflow before you understand the task. This is the most common mistake. People who barely understand what makes a good Instagram caption try to build a caption workflow. The AI makes up for some of the knowledge gap, but not all of it. Understand the task first. Build the workflow second.

How to Turn a One-Time Prompt Into a Permanent Skill

If you have run the same prompt more than three times, it should be documented.

Here is the three-step process:

1. Extract the template. Take the prompt you have been using and identify the parts that change each time. Replace them with placeholder fields (like our input template).
2. Write the output standard. Describe what a good output looks like in concrete terms. If you would have to think about whether something “counts” as good, your standard is not specific enough.
3. Add the review checklist. 3-5 yes/no questions that an output must pass before it is used.

That is it. That is a skill. You just built something you can hand to your future self or a teammate.

Your First Workflow Challenge

Here is a concrete starting point.

Pick one task you do repeatedly. Something you do at least once a week. It could be:

• Replying to customer enquiries
• Writing product descriptions
• Summarising supplier emails
• Creating social media content
• Drafting follow-up messages

Now build a 3-step workflow for it:

1. Input template (what information does the prompt need?)
2. Prompt (templated instruction to Claude or your tool of choice)
3. Output standard (what does “done and good” look like?)

Run it five times. Refine it once after run 3 and once after run 5. On run 6, you should have something that reliably works.

This is how the old 2BFT Academy skill library was built. We ran a task, refined the process, documented it, and packaged it. Every skill was a workflow template. Every workflow template came from work we were already doing.

See the learning side of 2BFT