Amplifying Engineering Skills with Claude Code

At a recent technical meetup, Solution Street consultant Ryan Gehl delivered a practical and wide-ranging presentation titled Amplifying Engineering Skills with Claude Code. The session focused on how developers can use Claude Code not simply as an autocomplete tool, but as a true engineering collaborator capable of accelerating coding, technical research, experimentation, and even lightweight product creation.

Ryan has been personally using Claude Code heavily since around October 2025 on Anthropic’s $100/month plan and described it as “the best $100” he spends each month. According to Ryan, the tool has “unlocked” his imagination and enabled him to pursue problems and experiments he otherwise would not have attempted.

The presentation was organized around several core themes: coding workflows, technical research, and future-looking ideas around autonomous agents and deployable applications.

Understanding Claude Code

Ryan described Claude Code in very simple terms: “a model with access to tools in a loop.” Rather than framing it as a complicated AI platform, he encouraged attendees to think of it as an AI model capable of repeatedly using tools such as browsers, file systems, terminals, APIs, and scripts to complete tasks.

He compared tool usage to everyday developer actions such as opening a browser or searching files on a desktop. In his view, Claude Code simply automates those same actions in a continuous reasoning loop.

Ryan emphasized that the interface itself matters less than the workflow. While modern AI coding assistants are commonly embedded inside IDEs like Cursor or GitHub Copilot, Claude Code brings developers back to a terminal-centered workflow, something many engineers are already comfortable with.

Typical Claude Code Workflow

Ryan walked through a standard interaction pattern he uses with Claude Code. His process generally begins by opening a terminal and asking Claude a carefully structured question. He noted that many prompts are substantial enough that developers often draft them in separate text editors before pasting them into the terminal.

One of his strongest recommendations was to use “plan mode” extensively. In plan mode, Claude investigates the task, launches sub-agents, performs research, and returns with clarifying questions before making changes.

Ryan explained that earlier versions of the workflow required repeated iterations on plans, sometimes three to five rounds before arriving at a useful implementation strategy. Over time, however, he found that either the models improved significantly or he became better at providing context. Today, he often reaches a workable plan after only one or two iterations.

Throughout the talk, Ryan repeatedly emphasized the importance of context and specificity. Clear instructions, structured prompts, and detailed examples consistently produced better outcomes.

Letting Claude “See What You See”

One of the most practical sections of the meetup focused on authenticated web access and API usage. Ryan explained that Claude Code can often access public resources directly, but authenticated systems present a challenge. His solution was to use browser developer tools and the “Copy as cURL” feature.

By copying a request as a cURL command, developers provide Claude with not only the URL but also authentication headers and session tokens. This allows Claude to access APIs and authenticated systems exactly as the developer sees them in the browser. Ryan demonstrated this using a Home Chef account, showing how Claude could retrieve recipe data from authenticated API calls.

Audience members discussed alternatives such as Playwright, which Ryan acknowledged as another effective option. Playwright enables Claude to automate browsers, log into systems, and run UI-driven workflows and tests.

Security concerns naturally arose during the discussion. Ryan explained that Anthropic allows users to disable model training on their data, and he also pointed out that many session tokens expire relatively quickly. Still, the discussion highlighted the need for caution when exposing credentials or sensitive information to AI systems.

Claude as an Engineering Teammate

A recurring theme throughout the meetup was that Claude should be viewed as a collaborator rather than just a code generator.

Ryan explained that Claude already understands many standard engineering principles and methodologies. Developers simply need to instruct it appropriately. For example, if asked to follow “red-green TDD,” Claude will automatically begin by writing failing tests before implementing functionality.

Similarly, Claude understands principles such as DRY (“Don’t Repeat Yourself”) and can perform structured code reviews. However, Ryan also cautioned that AI-driven code reviews can become endless loops where the model continually finds additional “issues” in an effort to satisfy the user.

He encouraged attendees to imagine working with “an awesome teammate” and ask themselves what tasks they would delegate. In practice, this meant assigning repetitive experimentation, optimization work, and exploratory coding to Claude while the developer focused on higher-level direction.

Performance Optimization Example

Ryan shared a detailed example involving API optimization work on a product with approximately twenty endpoints. The APIs needed performance improvements, so he instructed Claude to:

1. Discover all API endpoints
2. Build standalone test files rather than modifying production code
3. Experiment with multiple optimization strategies
4. Benchmark the results
5. Iterate further based on findings

He found Claude especially effective at generating multiple experimental implementations and testing combinations of approaches. In some cases, endpoint performance improved dramatically, such as reducing execution time from thirty seconds to twelve seconds.

An especially valuable insight was Claude’s ability to transfer optimization ideas across endpoints within the same repository. Once it learned patterns that improved one endpoint, it reused those concepts elsewhere.

Managing Context Windows

Another major discussion topic centered around context management.

Ryan described himself as “hyper vigilant” about monitoring context usage. He observed that once a session exceeded roughly 40-50% of the available context window, response quality often degraded noticeably.

Rather than relying heavily on context compaction, Ryan said he now typically starts fresh sessions and asks Claude to summarize what should be carried forward into the next conversation.

Audience members shared similar experiences, noting that smaller, cleaner coding sessions often produce more reliable outcomes than long-running conversational threads.

Ryan compared overloaded context windows to onboarding an intern by first forcing them to read the entire U.S. tax code before asking a practical question. Too much irrelevant context can dilute focus and reduce performance.

Claude Skills

The meetup also covered “Claude Skills,” which Ryan described as structured markdown-based workflows and instructions that can be dynamically loaded into Claude sessions.

Skills can encode:

• Coding conventions
• Project-specific rules
• Testing workflows
• Troubleshooting procedures
• Deployment instructions

Ryan shared examples where skills defined date-handling conventions, evaluation frameworks, and troubleshooting procedures for chatbot systems that generated S3 session data and CloudWatch logs.

Instead of manually searching logs and session histories, Claude could automatically retrieve and analyze the relevant data using instructions embedded in the skill definition.

He also highlighted Anthropic’s own “skill creator skill,” which helps developers generate new skills automatically from successful workflows.

Technical Research Workflows

A particularly compelling portion of the meetup focused on using Claude for technical research.

Ryan described a scenario where a product manager requested reporting functionality for a dashboarding platform and provided twenty manually-created customer reports as examples.

Using Claude, Ryan:

• Downloaded the files
• Extracted and analyzed the report structures
• Generated markdown summaries
• Compared report requirements against the platform’s APIs and data sources
• Evaluated implementation feasibility
• Produced actual sample reports using real data

Claude introduced concepts such as “data readiness level” and created structured analyses showing which report types could be generated from existing APIs.

Ryan estimated that work which might previously have required a team several weeks to complete took approximately twenty-two minutes from start to finish.

Autonomous Research and Deployable Applications

The final portion of the meetup shifted toward more experimental territory.

Ryan discussed combining Claude Code with tools such as Fly.io, OpenClaw, GitHub workflows, and lightweight Linux environments called “sprites.”

These stateful sandbox environments can rapidly start, sleep when idle, and wake in milliseconds. Ryan used them to create autonomous agents and deployable web applications.

Examples included:

• A tennis court availability assistant
• Physics simulations for ultimate frisbee throws
• Bridge probability analysis tools
• Public API exploration systems
• Airport visualization apps
• Germany travel planners
• Lunch recommendation tools
• Sailboat racing and tide analysis applications

Many of these were initiated directly from Ryan’s phone using Claude Code for Web. He explained that he could ask a question, wait fifteen or twenty minutes, and receive an email containing a deployed web application generated from the request.

Rather than giving Claude a traditional coding task, Ryan was giving it a research problem while using software engineering infrastructure such as repositories, workflows, and deployment pipelines to manage the results.

Final Takeaways

The meetup showcased Claude Code not merely as an AI assistant, but as a force multiplier for software engineering and technical exploration.

Ryan consistently emphasized experimentation, iterative workflows, context management, and structured prompting. His examples demonstrated how developers can offload repetitive work, accelerate discovery, automate analysis, and rapidly prototype ideas that previously required substantial engineering effort.

Perhaps the strongest message of the session was that the limiting factor is no longer simply implementation speed. Instead, the challenge is learning how to effectively collaborate with AI systems that can reason, experiment, research, and build alongside developers in increasingly sophisticated ways.

Solution Street is a Software Engineering and AI Consulting company. Reach out today to learn how AI can streamline your engineering process and deliver measurable efficiency gains!