Workflow-First Custom Agent Design
Build automation around the team's real workflow instead of a generic template.
- Difficulty
- Moderate
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 97%
Workflow-first agent design begins with operational reality rather than a prebuilt automation template. The builder explains the current process, recurring source material, desired outputs, review needs, and collaboration environment to a coding agent. The first implementation covers the smallest useful path and is tested with a real example. The team then customizes it for its own formats, clients, brand requirements, and handoffs. Manual actions can remain when they already happen naturally, while repetitive transformations and routing become automated. A downloadable markdown implementation may accelerate the build, but it is treated as a starting point because each team knows its workflows best. This approach lets a non-technical operator turn deep process knowledge into a tailored tool through guided conversation.
Origin
Extracted from Marketing Against The Grain, where a non-technical agency operator explains how she creates client-specific agents with Claude Code.
Core principles
- 01Begin with the current workflow and desired result.
- 02Customize the agent to the client or team using it.
- 03Use conversation to translate operational knowledge into software.
- 04Retain manual steps when automation would not add meaningful value.
- 05Treat shared templates as starting points rather than final solutions.
How to run it
- 1
Map the Existing Process
Document what currently happens before and after each event, including tools, inputs, outputs, owners, and review points.
Pro tip Describe what the team already does rather than designing an idealized process from scratch.
Watch out Automating an unclear process usually preserves its confusion.
- 2
Define the Final Result
State exactly what the agent should produce and where those outputs should go.
Pro tip Use a successful prior deliverable as a concrete reference.
Watch out Vague goals encourage generic outputs.
- 3
Explain It Conversationally
Give the workflow description to Claude Code or first draft the tool specification in Claude Chat and transfer it as markdown.
Pro tip Use ordinary operational language; technical implementation can be guided by the coding agent.
Watch out Do not assume the build will be correct in one exchange.
- 4
Build the Minimum Useful Path
Implement the recurring transformation before adding optional integrations or full automation.
Pro tip Leave naturally convenient steps, such as an existing transcript download, manual at first.
Watch out Premature automation adds failure points.
- 5
Test With Real Work
Run an actual transcript through the agent and inspect whether the package supports the team's next actions.
Pro tip Judge usefulness within the real publishing workflow, not just whether files were generated.
Watch out A technically successful run can still produce operationally useless content.
- 6
Customize Per Team or Client
Adjust formats, context, brand rules, and handoffs for the people who will use the agent.
Pro tip Use the shared implementation as a foundation rather than accepting it unchanged.
Watch out Overstandardization may erase important client differences.
In the wild
An agency maps one client's webinar process, required articles, social outputs, email sequences, and Google Docs handoff. It explains the process to Claude Code, tests the resulting app on a recent transcript, and then stores the client's rules and context alongside the agent.
→ The agency gains a tailored tool in a few hours without forcing the client into a generic automation.
Common mistakes
Adopting the Demo Unchanged
A shared implementation can accelerate setup, but it may not reflect the team's formats, tools, or review process.
Automating Every Click
A step that already occurs naturally may not deserve an integration until it becomes a measurable bottleneck.
Expecting a One-Shot Build
Conversational software development still requires testing, clarification, and iteration.
Is it for you?
Best for
It is best for domain experts who understand their process but lack conventional software-development skills.
Not ideal for
It is not ideal for poorly understood processes that vary completely on every run.
From the transcript
“But look at your own workflow. And if something needs automating, do it.”
“just describe what you need and what your final result should be, and it will tell you how to do that.”
“I really encourage you to also try to build it from scratch because you know your workflows best currently and you know what you need.”
From the episode
This Content AI Agent Runs My 0-Employee Marketing Agency