1
Archetype
2
Triggers
3
Outputs
4
Customize
5
Preview

What should this workflow do?

Choose the type of agentic workflow you want to create.

What triggers it?

Select one or more events that start this workflow.

What can it write?

Select the actions the workflow is allowed to perform.

Any extra context?

Optional β€” add details to make the workflow fit your project. Leave blank for a solid default.

Preview & Download

Choose how you want to use your workflow.

workflow.md

⚠️ Anti-Patterns & Lessons Learned

Based on analysis of 679 workflows Γ— 4,888 runs. Avoid these patterns to improve workflow reliability.

πŸ”΄ Critical 0–2% success

slash_command trigger

The slash_command trigger has a near-zero success rate across all combos (n=204). It is effectively broken.

Fix: Remove slash_command. Use issues + workflow_dispatch instead for user-initiated workflows.
πŸ”΄ Critical 13% success

workflow_run chaining

Using workflow_run as the sole trigger results in 13% success (n=76). Chain-of-workflow patterns are fragile.

Fix: Use schedule + workflow_dispatch with pre-steps for data gathering instead of workflow chaining.
🟑 High 47% vs 67%

codex model selection

The codex model has a 47% success rate vs 67% for the default model. Use default unless you specifically need code generation.

Fix: Remove model: codex. For complex synthesis, use claude-opus-4.5 instead.
🟑 High βˆ’21pp success

Missing workflow_dispatch

Workflows without workflow_dispatch have 21 percentage points lower success (n=3,476 vs n=1,412). Also blocks manual re-runs for debugging.

Fix: Always add workflow_dispatch: to your trigger config.
🟠 Medium 32% prevalence

Template cloning without adaptation

32% of all workflows are unmodified template copies. Copied workflows have lower success because permissions, context, and timeouts don't match the target repo.

Fix: After copying a template, customize triggers, timeout, safe-outputs, and all repo-specific references.
🟠 Medium Context overflow

Oversized prompt on default model

Prompts exceeding 30KB on the default model risk context limits and degraded performance. The sweet spot is 3–8KB.

Fix: Split into multiple workflows, or upgrade to claude-opus-4.5 for prompts >20KB.
πŸ“Š Data: analysis-report.json Β· πŸ“– Full guide: Best Practices Β· πŸ”¬ Research: Issue #885