[Prepare -> Plan -> Execute -> Review]
Here's the process I’m using to create on-brand marketing assets for B2B GTM teams. It works pretty well for writing tasks, design, and some coding. I’ve included overall process and a specific example for one step in my workflow.
If you find this interesting, shoot me an email at akhil@fluxgtm.com.
Anyway, here's the overall process that I'm using:
- Prepare: give the model a heads-up of what you’re going to be working on in this session. I’ve got a detailed explanation about the project in a README.
- Plan: get into the specifics of the task at hand, building the to-do list etc. For repeated tasks I use a custom slash command (sometimes with $ARGUMENTS for variables), or just raw dog a new conversation. This is all in planning mode.
- Execute: once I’m happy with the plan, I let Claude Cook
- Review and selectively improve: this step has the biggest improvement in outputs
Tactical note: the tasks I’m working on are quite varied so accounting for every situation a single CLAUDEmd file doesn’t make sense. This README approach allows me to be more flexible.
Now here’s a specific application of this process that I’m using to create Brand Systems for clients
- Prepare
- Start a new chat, use a /new-chat slash command to prompt Claude to review the README to get up to speed with the project.
- Plan - all in plan mode
- Use a custom slash command to explain the part of the process that we’re working on i.e., /brand-system:01-start
- This explains the part of the process that we’re going to be working on with the files to expect in the next prompt
- Another custom slash command with the below inputs
- Location of the design reference images to use the brand system, which are referenced as $ARGUMENTS since the location changes depending on the client I’m working with
- A generic JSON template with the structure of the brand system
- A detailed prompt with instructions
- Since I’m in plan mode, I review Claude’s todo list to make sure it’s aligned. For the brand system, it’s usually pretty standard. Other steps in my process require more iteration.
- Execute
- Run the todo list, check twitter, grab a coffee
- I usually use Opus 4.1 for creative tasks like design and writing, especially anthing multimodel (like this example where I’m sending images)
- Review - initially in plan mode, then switch to run
- Plan mode
- Once I have the output, I have another custom slash command with a lengthy review prompt, specific to each step in the workflow. I also re-share the design reference images
- Importantly, the custom prompt focuses on just listing issues, not suggesting any fixes
- Here, I review the list of issues and choose the ones that I want to implement
- Execute mode: Implement the suggestions
- In most cases, one loop of this review/issue is enough. Some steps in my workflow are more judgement based so for those I’ll run through review/improve loop a couple more times.
- Plan mode
Questions you might have
- Why don’t you use single larger prompts instead of splitting things up?
- For my workflow, splitting tasks into these smaller steps tasks feels like it gives better outputs
- This helps me to build the relevant context and feels like it ‘primes’ the model for the primary task
- This is 100% anecdotal but it works for me
- Where do you save the custom slash commands?
- I keep the the custom commands and other things like writing guidelines, templates, etc. in the project repo so I can use github to manage versions
- Why don’t you use subagents for some of these tasks?
- For my workflow, maintaining a single context for discreet tasks works best
- Subagents spin up a fresh context each time so don’t do the trick for me. The tasks I’m working on require building context from previous steps
- How are you using custom output styles?
- I’m experimenting with these, e.g. for the copywriting that I do as part of my process
- I’ve got a business copywriting output style that helps to convert input text (e.g. call transcripts) into marketing copy for some output (e.g. case studies), but it does require me providing a guideline/template of the structure that I want to follow
I’m still building this plane while I’m flying it - would love any thoughts on this process, ways to improve, things I’ve missedm etc.