Putting it together · for software engineers
Building AI workflows in software engineering
One request can produce a function. A feature needs more than that: understanding the existing code, a plan, the change, tests, a review, and a pull request description. Asking for all of it at once tends to produce a large diff that is hard to review and wrong in places.
Putting it together means splitting recurring engineering work into stages. A human check sits at the points where a wrong turn is expensive. The result is a process a teammate can run again.
Where it comes up
A feature from ticket to pull request
Read the relevant code, write a short plan, implement in small commits, generate and check tests, then draft the pull request description. You approve the plan before any code is written.
A dependency upgrade across many services
Use AI to summarize breaking changes from the changelog, then apply the change to one service and review it fully. Only after that service passes its tests do you apply the same change to the others.
An incident write-up
Collect the timeline from logs, alerts, and chat history first. Have AI draft the write-up in the team’s template from that material. Write the root cause and follow-up actions yourself after checking the evidence.
What it looks like at each band
E1 · Foundational
Uses AI for one step at a time and moves results between tools by hand. Each task starts from a new, blank request. Results vary from one task to the next.
E2 · Proficient
Runs recurring work as a sequence of requests and reuses the good ones. Keeps changes small enough to review. Reviews at fixed points, such as after the plan and before the commit.
E3 · Distinguished
Designs workflows with clear handoffs between AI steps, tests, and human review. Chooses different tools for different stages, such as an agent for mechanical edits and a chat for design questions. Documents the workflow so a teammate can run it.
E4 · Exceptional
Builds the team’s AI-assisted development process and tracks where errors enter it. Adds automated checks where review keeps catching the same problem. Removes steps that no longer save time. Other engineers adopt these workflows because they produce fewer regressions.
Common mistakes
One request for the whole feature
A single large request produces a large diff with many decisions buried in it. Split the work so each change is small enough to review in one sitting.
No checkpoint before implementation
If the plan misreads the requirement, every later step inherits the error. Review the plan, or the list of files to change, before any code is written.
Letting an agent merge unreviewed changes
An automated step that commits or deploys without a human reading the change will eventually ship a defect. Keep a review gate before anything reaches the main branch.
A worked example
Replacing a deprecated logging library across a service with about forty files that use it.
Ask an agent to replace the old logging library everywhere and open one pull request with all the changes.
The diff touches forty files at once. Some log levels and structured fields may be mapped wrongly, and the reviewer cannot check every line with care. Errors reach production in the logs that matter most during an incident.
Step 1: ask for a mapping table from old calls to new calls, and check it against both libraries’ documentation. Step 2: convert one module and review it in full. Step 3: run the agent on the remaining modules in batches of a few files, using the checked mapping. Step 4: run the tests and compare sample log output before and after. Step 5: open small pull requests a reviewer can read.
The mapping is checked once and reused. The first module proves the approach before scale. Each step has one job, and the human checks sit where a mistake would spread.
The full study guide for this skill, for every role, is Putting it together. The other five skills for software engineers are on the AI skills for software engineers page. A one-page summary is in the AI workflow checklist.
See where you stand
The free check grades putting it together and the other five skills in about fifteen minutes. Nobody sees the grade but you.
Then build the skills and prove it. The three steps for software engineers list the courses and templates that fit the work.