andrewsthoughts

   ,d88b.d88b,
   88888888888
   `Y8888888Y'
     `Y888Y'
       `Y'

For us, in this scenario: 1) the pipeline helps agent perform better 2) reviewing the spec is much more convenient than when juggling between TUI and text editor (esp. if you are running 5 of those pipelines in parallel) 3) if you configure the reviewer in the settings, cross-agent review saves us from some of the minutae of guiding/aligning the agent

Lmk if I misunderstood your question, happy to help.

Meanwhile, you can BYOA - bring your own agent - if you are a hobbyist, Gemini is free with gmail (but they WILL train on your data). And if you have ChatGPT sub, you can use codex CLI with Zenflow for no extra charge (and they don't train on paid users data).

Create a new task with your prompt, and hit "Create" (instead of "Create and Run"). The interface will show a little hint "Edit steps in plan.md", with 'plan.md' being clickable. Click on it and edit it, experimenting with some ideas. {Bonus tip: toggle "Auto-start steps", to keep it Ralph-y)

I just winged the workflowsbelow, and it worked for the prompt I threw at it. If you like it, you can save it as your custom workflow and use it in the future. If you don't like it - change to your preference.

Now, I prefer a slightly different flow: Implement > Review > [Fix] (and typically limit the loop to 3 times to avoid "divergence"). We'll ship some pre-built templates for that soon. Our researchers are currently working on various variations on our private datasets.

--- # Quick change

## Configuration - *Artifacts Path*: {@artifacts_path} → `.zenflow/tasks/{task_id}`

---

## Agent Instructions

This is a quick change workflow for small or straightforward tasks where all requirements are clear from the task description.

### Your Approach

1. Proceed directly with implementation 2. Make reasonable assumptions when details are unclear 3. Do not ask clarifying questions unless absolutely blocked 4. Focus on getting the task done efficiently

This workflow also works for experiments when the feature is bigger but you don't care about implementation details.

If blocked or uncertain on a critical decision, ask the user for direction.

---

## Workflow Steps

### [ ] Step: Implementation

Implement the task directly based on the task description.

1. Make reasonable assumptions for any unclear details 2. Implement the required changes in the codebase 3. Add and run relevant tests and linters if applicable 4. Perform basic manual verification if applicable

Save a brief summary of what was done to `{@artifacts_path}/report.md` if significant changes were made.

After you are done with the step add another one to `{@artifacts_path}/plan.md` that will describe the next improvement opportunity.

On prompts: We've been competing with Cursor for the last 2 years in the enterprise with Zencoder, and winning nice deals based on quality. At some point, we were very protective of our prompts, but two things happened: -most of the coding vendors' prompts were leaked, there are repos online that have prompts from a bunch. The moment you allow a custom end-point for LLM, your prompts are sniffable. -agents became better at instruction following, so a lot of prompting changed to "less is more".

So with these two industry trends, we reversed the course: -moved our harness into CLI - this exposes our tips and tricks, but is better for user privacy and for user's ability to tinker the harness. For example, this allows a set-up where no code leaves your perimeter (if you use local harness and "local" model, where "local" means different things for different people) -opened the workflows in Zenflow (they are in markdown and editable)

Apple Silicon (ARM64): https://download.zencoder.ai/zenflowapp/stable/0.0.52/app/da...

Intel (x64): https://download.zencoder.ai/zenflowapp/stable/0.0.52/app/da...

We'll figure out the FF script blocking.