Making AI-assisted code changes reviewable with Agent Kit

Khiêm Nguyễn — Thu, 12 Nov 2026 06:40:04 +0000

AI-assisted changes need a reviewable record of what was requested, what was approved, and what was actually checked. Agent Kit packages one way to keep that record inside a project.

Disclosure: I maintain Agent Kit. This article explains its documented workflow with an illustrative example. It is not a benchmark or a report of verified behavior across every supported client.

Start with a bounded change

Consider a small request: add Copy buttons to two onboarding prompts. A useful plan identifies exactly what each button copies, how success and failure appear, and how someone using a keyboard reaches the controls. It also identifies the checks needed to accept the result.

For this example, the scope might be:

Copy the full prompt without changing its text.
Show success only after the clipboard operation succeeds.
Show a recoverable error when clipboard access fails.
Support keyboard activation and a narrow viewport.

These are illustrative acceptance criteria, not claims that this article implements or tests the feature.

Make approval concrete

Agent Kit uses the sequence Plan → Approve → Build → Verify → Hand off. Before implementation, a WHAT table identifies the intended work, its purpose, location, status, and the reason for anything excluded.

The owner completes the approved scope without asking for approval again for each file. A material change in scope needs a revised plan. This makes approval specific enough to compare with the eventual diff.

In the Copy-button example, changing authentication or redesigning the whole onboarding page would require a separate decision.

Keep implementation and evidence separate

At handoff, use the same work items to report the actual outcome. A completed source edit is one fact. A successful clipboard check is another. Human acceptance and publication are separate decisions.

For the example, useful evidence includes an exact comparison between the prompt and clipboard contents, keyboard activation, the narrow layout, and behavior when clipboard access is denied. A screenshot can show the layout, but cannot prove clipboard contents.

If a check was not run, say so and explain what remains uncertain. A structured report makes the evidence available for inspection; the report itself does not independently verify the claim.

Preserve the context for the next session

The kit separates durable instructions, stable project context, intended outcomes, and one current-state summary. Detailed task records retain decisions and evidence when useful. A new session can inspect those records rather than relying on an earlier conversation being available.

Records still need maintenance. An old approval does not authorize a new task, and an old passing check does not establish that the current revision works.

Use focused guidance

Agent Kit includes 17 technical skills and optional exploration, implementation, and review roles. Select guidance relevant to the task rather than loading everything. Skill contribution tables make that selection visible; their percentages are subjective estimates, not measured time savings. A multi-agent team is optional.

Project-local instructions and adapters cover Codex, Claude Code, Gemini CLI, and Cursor. Grok web/app has a manual reference. These are source capabilities, with client-specific verification limits documented in the repository. I have not personally used Claude Code, so I am not presenting a Claude Code runtime experience here.

Try one small task

Use the quick-start prompts in a disposable project. Ask for one bounded change, review the plan, and compare the final diff and checks with the approved scope. Keep existing business rules, stack, and permissions when integrating into an established codebase.

Instructions guide the assistant. Backend authorization, sandbox controls, tests, and human review still need their own mechanisms.

The template is MIT-licensed: github.com/khiemnd777/agent-kit. Feedback on which records help during a fresh session—and which add unnecessary work—would be useful.

Prepared with AI assistance and checked against the repository’s documented workflow.