An AI workflow efficiency list compiled with Astra
The author lists practices such as matching models and reasoning depth to each task, planting deliberate bugs to test whether tests catch them, and writing project memory into files, each with a copyable prompt.
If you are just getting started with Ai, or would like to enhance your workflow, allow me to offer some efficiencies I have uncovered. You can tell your AI to consider implementing any or all of these into your workflow. I'd love to hear your feedback! Each method is followed by a prompt you can feed your machine (or copy and paste it all). The following was written, mostly, by ChatGPT: Astra/High ‐---- 1. Match AI models/depths to specific jobs: planning, building, and checking. Use the right brain & brain depth (high, low, etc.) for the job - conserve compute and save time. Reserve expensive reasoning for difficult problems. A separate reviewer checks finished work against the original requirements and actual test results. AI prompt: Inspect the project instructions. Add a small task queue with a suggested model and reasoning effort per task. Separate building from independent review. Require test evidence before marking work complete. Reuse existing tools. Tests required to catch deliberate mistakes 2. Passing tests can hide broken software. Introduce a deliberate mistake in a disposable project copy and check whether the tests notice. A test capable of catching a known bug provides stronger evidence than a green checkmark alone. AI prompt: Run the existing tests. In a disposable copy, introduce one realistic bug. Confirm a relevant test fails for the intended reason. Improve weak tests using written requirements. Keep deliberate bugs out of the main project. Record the results. 3.Project memory surviving a fresh chat Save project memory in files: completed work, decisions, test results, unfinished tasks, and the next action. A fresh AI chat can continue from a short handoff instead of rebuilding context from a long conversation. AI prompt: Create or update HANDOFF.md with the goal, current project version, completed work, evidence, remaining problems, and next task. Link detailed records instead of copying long logs. Add an instruction to read the handoff before continuing. 4. Publishing with automatic safeguards Publishing becomes safer when a small program checks the finished files, catches suspicious deletions, detects competing changes, and verifies the uploaded result. A saved receipt connects every release to the exact work and test evidence. AI prompt: Inspect the publishing process. Add a dry-run check listing exact files, flagging sensitive files and large deletions, and stopping on competing changes. Verify published files and save a receipt. Test in a disposable repository. Follow existing approval rules. 5. Shared tools proven in real projects Build a shared tool after a repeated need appears in real projects. Test the shared part in both projects, keep project-specific rules separate, and record the tested version. Useful reuse reduces work without forcing simultaneous upgrades. AI prompt: Find one repeated task across two real projects. Check existing tools first. Propose the smallest shared component and compare upkeep costs. Prove compatibility in both projects, record the tested version, and document independent upgrade and removal steps.