Five Levers for Optimizing a Repeated Workflow
CoreOptimize workflows for efficiency and effectiveness · Difficulty 2/5
Explanation
Beyond Fixing One Output
Troubleshooting a single weak output is reactive. Optimization is proactive: it makes a workflow that runs repeatedly faster, cheaper, and more reliable every time it runs, not just the one time something went wrong. Five concrete levers do this work:
| Lever | What it buys |
|---|---|
| Right-size the model per step | A fast model for bulk simple work, a stronger one only where reasoning is genuinely hard -- efficiency and cost together |
| Persist reusable context in a Project | Instructions and knowledge aren't rebuilt every session -- consistency and speed |
| Standardize proven prompts | Quality doesn't depend on re-inventing the prompt each time -- templates or system-level instructions carry the win forward |
| Calibrate human review | Review effort matches the stakes -- kept where it counts, streamlined where it doesn't |
| Decompose long tasks | Checkable steps catch errors early instead of letting them compound across a long task |
These Are Workflow-Level, Not Output-Level
Each lever targets how the workflow is *structured* for repeated use, distinct from diagnosing any single bad output (Task Statement 7.1) or iterating on one prompt (Task Statement 7.2). Right-sizing the model and standardizing prompts, in particular, are the same moves used to fix a single output -- but applied here as a standing policy for every future run of the workflow, not a one-time correction.
Common exam traps
- Assuming decomposition, model right-sizing, or Projects are one-off fixes rather than standing workflow design choices meant to pay off on every subsequent run.
- Treating human review as purely a speed cost to be minimized -- calibrating review to stakes, not eliminating it, is the actual lever.
Key Takeaways
- Optimization targets the whole repeated workflow, not just one weak output
- Right-sizing the model per step and persisting context in a Project both buy efficiency and consistency across runs
- Standardized prompts and decomposition into checkable steps make quality and error-catching repeatable, not one-time wins
- Human review should be calibrated to stakes -- kept where it counts, streamlined elsewhere -- not simply minimized
Glossary Terms
A persistent workspace in claude.ai that bundles standing Instructions (role, tone, format, rules) and Project Knowledge (uploaded reference documents) so every conversation inside it draws on the same configuration without re-supplying it each time. Scales beyond a single context window because large Project Knowledge is retrieved rather than loaded in full on every turn.
The practice of directing requests to different Claude model tiers based on assessed complexity and requirements. A common pattern uses a fast, cheap model (Haiku) to classify task complexity, then routes to Sonnet or Opus accordingly.
The process of breaking a complex task into smaller, independently executable subtasks that can be assigned to specialized subagents or processed sequentially. Good decomposition creates subtasks with clear boundaries, independent execution, and verifiable outputs.
Related Concepts