Make your first code change with an agent
Connect a model, choose a working directory, describe one small change, then inspect the diff and run checks. Explore testing and interface-design Skills as your workflow grows.
For people with an existing project who want help fixing an issue or making a small change.
A practical order to follow
Prepare a working tool
Confirm a model connection in Lumoswitch, select a supported agent, and review before applying its configuration. Open a project copy or a committed working directory in the agent.
Choose one small change
Specify the desired behavior, writable scope, and completion criteria. Start with a title change or a reproducible bug before attempting a large rewrite.
Inspect before continuing
Review the diff and run the project's existing checks or tests. Ask the agent to identify anything it cannot verify, then move to the next small task.
Read this project and find the home-page title. Change it to ‘My first project’. Only edit the relevant page; leave dependencies and APIs alone. List changed files, explain how to check locally, and identify anything you could not verify.
Look for a clear file diff and actual check results. Confirm the page meets the request before keeping the change.
Tutorials and common questions
- How do I ask an AI agent to make a first small project change?
Choose the project folder first, then describe one small change. Use Plan mode for multi-step work and inspect the resulting diff and checks.
- How do I check code changes made by an agent?
Inspect the patch, then confirm which tests and reviews actually ran. Do not treat an unrun check as a pass.
- Where should I start with an AI desktop app?
Install the desktop app, confirm a usable model connection and an agent, then review before applying configuration changes.
- Which AI agent should I install: desktop or CLI?
Choose a desktop app for a graphical workflow and a CLI for terminals or scripts; configuration management depends on adapter support.
- How do I apply a model connection to an agent?
Review and confirm a supported native configuration before writing it; reopen the agent to load the new settings.
- How should I troubleshoot a model connection error in my agent?
Confirm the active configuration path and model ID first, then check keys, provider access, limits, and request records.
Explore Skills for this task
Review the purpose and prerequisites before choosing. These are source reading notes; community availability and client compatibility are shown on their respective pages.
Add useful tests to a Python project
python-testing-patternsUse pytest, fixtures, and isolated dependencies to cover ordinary inputs, failures, and boundaries. Useful for regression tests after a bug fix.
Prerequisites, example, and checks
- Before you start
- Provide a runnable Python project and test environment. Specify writable directories and external services that must not be called.
- Example task
- Add pytest coverage for order totals, discounts, empty orders, and invalid inputs. Only edit tests, do not call real payment services, and run the tests with a failure summary.
- Check the result
- Check that tests expose meaningful mistakes rather than mirror the implementation. Inspect test output and confirm no production database or paid API was called.
Find why a Python program is slow
python-performance-optimizationMeasure runtime and memory before changing code. Useful for slow scripts, data pipelines, and API diagnostics; no fixed speedup is promised.
Prerequisites, example, and checks
- Before you start
- Provide repeatable input, a run command, and a performance goal. Work in a test environment instead of load-testing production.
- Example task
- This CSV cleaning script is slow on 100,000 rows. Measure and locate the bottleneck, propose a minimal change, preserve output, and compare using identical input.
- Check the result
- Repeat timing, memory, and correctness checks with the same environment and data. Separate network variance from code improvements.
Catch interface mistakes earlier with TypeScript
typescript-advanced-typesUse generics, conditional types, and mapped types to express interface constraints. Helpful for utilities, component props, and API clients without needless complexity.
Prerequisites, example, and checks
- Before you start
- Provide existing definitions, call examples, and the TypeScript version. Identify interfaces that may change and preserve business behavior.
- Example task
- Type this API client so each endpoint returns the corresponding result. Preserve runtime behavior, avoid any, and show calls that should pass or fail type checking.
- Check the result
- Run the project type check and verify invalid examples fail. Static types do not replace runtime validation of external input.
Design clear API and module interfaces
api-and-interface-designDefine inputs, outputs, errors, and compatibility before implementation. Useful for new APIs, frontend/backend collaboration, and changes to public interfaces.
Prerequisites, example, and checks
- Before you start
- Provide real use cases and behavior relied on by existing callers. Specify fields and error formats that must remain compatible.
- Example task
- Design a paginated tasks endpoint with request parameters, responses, errors, and edge cases. Preserve existing callers. Present a design and examples before changing the server.
- Check the result
- Review pagination, empty results, permission errors, and old callers. Passing tests alone does not prove a public behavior change is compatible.
Request a focused code review before delivery
requesting-code-reviewFor completed features or significant changes, define scope and expected behavior so review can focus on correctness, regressions, and requirements. Useful before submitting or merging.
Prerequisites, example, and checks
- Before you start
- Provide the change scope, base and current revisions, task requirements, and known constraints. List commands already run and their outcomes.
- Example task
- Review this permission change against the requirements: check unauthorized access, compatibility with existing callers, and error handling; report reproducible issues by severity and file location.
- Check the result
- Review findings should cite concrete evidence and impact rather than style preferences. Recheck the scope after important fixes.
Plan and build a maintainable MCP service
mcp-builderDesign clear MCP tools and resources around user tasks, then refine the implementation with examples and checks. Useful for connecting an existing API or data source to an agent.
Prerequisites, example, and checks
- Before you start
- Describe users, available APIs, authentication, deployment environment, and permitted data. Include versions of existing services or SDKs.
- Example task
- Design an MCP service for an internal ticket API. Propose a minimal tool set, inputs, outputs, and errors, then implement read-only lookup with local verification steps.
- Check the result
- Review tool descriptions, parameter constraints, and actionable errors; verify authentication and data boundaries cannot be bypassed.
Turn a repeatable workflow into a reusable Skill
skill-creatorDefine when the Skill should trigger and what it should deliver, then organize concise instructions, references, and optional scripts. Useful for stable multi-step workflows.
Prerequisites, example, and checks
- Before you start
- Prepare real task examples, expected outputs, tool dependencies, and nearby cases that should not trigger it. Choose where the files belong.
- Example task
- Turn a weekly release-note process into a Skill: define triggers, inputs, steps, and an output template, then create representative prompts and look for gaps or excessive scope.
- Check the result
- Ensure trigger text matches the workflow, linked resources exist, examples cover the main path, and no secrets or unauthorized actions are included.
Turn a correction into a testable working lesson
self-improving-agentExtract candidate lessons from failures, corrections, or repeated problems; validate before applying durable guidance. This does not train the model or provide unlimited memory.
Prerequisites, example, and checks
- Before you start
- Provide concrete evidence and permitted guidance files. Redact secrets, personal data, and raw conversation contents.
- Example task
- Review the failure caused by the incorrect directory assumption. Draft one candidate rule and a validation method without changing durable settings; propose applying it only after validation.
- Check the result
- Distinguish candidate, validated, and applied states. Check representative behavior and rollback options. No update is a valid outcome when evidence is insufficient.