Building Your Personal AI Workflow
Devops
Your workflow: incident triage → log analysis → runbook draft → postmortem outline. Chain the right tools.
Data Eng
Pipeline design → SQL generation → docs → data quality checks. One workflow, multiple AI touchpoints.
Building Your Personal AI Workflow
TL;DR
- A workflow isn't "use ChatGPT sometimes." It's a repeatable sequence: trigger → tool → output → next step.
- As of 2026, ~85% of developers use AI tools. EY (2025): 96% of orgs report productivity gains; 57% say significant. The gap is having structured workflows, not access.
- Start with one high-frequency task. Automate it end-to-end. Add more from there.
An AI workflow is a habit, not a feature. It's "when X happens, I do Y with AI, then Z." Once you have a few of these, you stop thinking about "should I use AI?" — it's just how you work. Reality check: developers often get slower before they get faster when adopting AI tools. Expect a ramp-up. Iteration beats perfection.
Step 1: Pick One Task
Don't try to AI-ify your entire job. Pick the single task you do most often that has a clear input and output.
Examples by role:
- Developers: Code review, writing tests, refactoring, PR descriptions
- Data people: SQL generation, pipeline design, data exploration
- DevOps/SRE: Runbook creation, log analysis, incident summaries
- Managers: Status updates, 1:1 prep, capacity planning drafts
- Tech writers: Doc outlines, API reference drafts, tutorial structure
- Support/Solutions: Response drafts, FAQ creation, solution proposals
Pick one. Make it concrete: "Every PR I open, I use AI to draft the description from the diff."
Step 2: Define the Sequence
Write it down. Literally. Bonus: treat prompts as source code — version them, reuse them, refine them.
- Trigger: What starts this? (e.g., "I've written code and am about to open a PR")
- Input: What do you feed the AI? (e.g., diff + ticket link)
- Prompt/tool: What do you ask for? (e.g., "Summarize changes, list risks, suggest reviewers")
- Output: What do you get? (e.g., draft PR description)
- Review & ship: What do you add/change before submitting? Add human-in-the-loop checkpoints for safety-critical outputs.
If you can't write it, it's not a workflow yet.
Step 3: Choose the Tool(s)
Shift from raw chat to tool-augmented systems where possible — AI that runs code, indexes your codebase, and produces verifiable results. Match the tool to the task:
| Task Type | Tool Examples |
|---|---|
| Code in IDE, codebase-aware | Cursor, Copilot, Claude Code, Windsurf |
| Quick Q&A, docs, summaries | ChatGPT, Claude (browser) |
| Multi-file code changes | Cursor Agent, Claude Code |
| Logs, traces, debugging | Claude, ChatGPT (paste + ask) |
| Spreadsheets, planning | ChatGPT, Claude |
You don't need 10 tools. Start with 1–2. Add when you hit a limit.
Step 4: Iterate for a Week
Run your workflow for 5–10 instances. Note:
- What worked?
- What broke?
- What did you have to fix every time?
Tweak the prompt. Add a step. Remove one. Workflows improve with use. Fun fact: in some agentic setups (AI that takes multi-step actions autonomously), task time can be cut in half — but only after you've ironed out the kinks. A 200-engineer org can see ~$4.5M/year in productivity gains without extra hiring, but only if workflows are repeatable, not ad-hoc.
Step 5: Add a Second Task
Once the first workflow feels automatic, add another. Connect them if they chain (e.g., "after incident, run log analysis → then draft postmortem"). Avoid over-automation without governance — in regulated domains, every AI move needs to be trackable.
Sample Workflows by Role
Developer (PR flow): Trigger: PR ready. Input: diff + ticket. Tool: Cursor/Copilot. Output: description draft. Review: add business context, verify.
Manager (status update): Trigger: Friday EOD. Input: notes from the week. Tool: Claude. Output: draft update. Review: add nuance, remove anything sensitive.
Support engineer (case response): Trigger: New ticket with known pattern. Input: ticket + KB article. Tool: ChatGPT. Output: draft response. Review: personalize, verify accuracy.
# My PR Workflow (save this as a template)
## Trigger
I've written code and I'm about to open a PR.
## Steps
1. Select the changed files in Cursor
2. Open Cursor Agent (Cmd+L) and paste:
"Summarize these changes for a PR description.
Include: what changed, why, risks, and testing notes.
Format: GitHub PR template with ## sections."
3. Review the draft — add business context, verify accuracy
4. Copy into GitHub PR
5. Ask Cursor: "Suggest 2-3 reviewers based on
git blame for these files"
## Time saved: ~15-20 min per PR
## Quality: Better descriptions = faster reviewsQuick Check
What's the first step to building a personal AI workflow?
Do This Next
- Write your first workflow for one task. Use the 5-step structure. Run it at least 5 times this week (research suggests 5–10 instances reveal what breaks).
- Version one prompt. Save it as a template — prompts as source code. Reuse and refine it over the next 2 weeks.