Hi all,
I am building a complex workflow-based agent and have run into what looks like a hard limit of around 8,000 characters per text field in the workflow builder, specifically on the instruction_template for a step (a Respond step).
My use case:
- One step is responsible for rendering a fairly rich, deterministic document for the user.
- The instructions for that step currently need to encode:
- Detailed Markdown section and table structure.
- Strict formatting rules for specific sections (e.g., how to format IDs and acceptance criteria).
- A standard support footer that should always appear at the end.
- Optional export format rules (e.g., how to produce a matching Confluence/storage‑format version of the same content).
Because of this, the instruction_template for that single step ends up over 8,000 characters (around 9K), and the workflow cannot be saved or published.
I have a few questions:
- Is there a documented / fixed character limit per
instruction_template (or per text box) in the workflow builder today? - Is that limit coming from the workflow configuration / UI, or from underlying LLM constraints?
- What patterns are recommended for cases where a single step genuinely needs detailed, deterministic formatting rules? For example:
- Moving long, stable conventions into a separate document and referring to it briefly from the step.
- Splitting behaviour across multiple steps while still keeping the dependency tree valid.
- Using any existing or planned mechanism for shared “prompt snippets” or reusable instruction templates across steps.
The goal is to keep the agent’s behaviour tightly specified to minimise hallucinations, but at the moment the per‑field size limit is the blocker rather than the model’s context window.
Any guidance or best practices on how to manage this within the current product limits would be much appreciated.
A simulated screenshot for your reference:
Thanks!
