Blog
Mert Özen's blog posts on software development, backend, frontend, and DevOps.
Privacy and Data Controls in Claude: Model Training, Retention, KVKK and Organization Settings
Turning off "Help improve our AI models" in Claude affects only new conversations; data in a started training run and in trained models is not taken back, and a conversation given feedback can be kept for up to 5 years. Consumer and commercial plan rules, retention periods, incognito and connector data, Enterprise retention settings, where data is stored, the KVKK question and a masking checklist for Bookstore support e-mails.
Agentic Work with Claude: Cowork Tasks, Claude in Chrome and Safe Permission Settings
The way to run an agentic task in Claude safely is to start with Manually approve on new sites and tools, turn off tools that can write and use the browser in a separate profile without sensitive accounts. Cowork tasks that run in the cloud and steps that need Desktop, scheduled tasks, the three permission modes, actions Claude in Chrome blocks in every mode, computer use and Bookstore's price comparison task.
Claude Skills and Plugins: Teaching Claude How Your Team Works
Claude picks a skill by looking only at its description; if the description does not say what the job is and which requests it is for, the skill never triggers. The SKILL.md structure and progressive disclosure, uploading to claude.ai and syncing to Claude Code, field limits, team sharing, plugins, the risk of untrusted skills and Bookstore's release notes skill.
Claude Connectors and MCP: Connecting Google Drive, GitHub, Slack and Your Own MCP Server
Because custom connector connections in Claude come from Anthropic's cloud rather than your computer, your MCP server must be reachable from the public internet over HTTPS on port 443; a localhost address does not work. The Connectors Directory, the Google Drive, GitHub and Slack connectors, the Always allow, Needs approval and Blocked permissions, and connecting the Bookstore MCP server with read tools only.
Research and Web Search with Claude: Cited Research and Verification
The way to use a Claude Research report safely is to write a primary-source rule and a row-by-row checkable table into the brief, then ask for each row's basis in the source's own words. Web search without a toggle in the unified experience, Research plans and limit cost, connectors used without approval, and Bookstore's .NET 11 migration research.
Claude Artifacts and File Creation: Excel, PowerPoint, Word, PDF and Interactive Visuals
The way to use an Excel report produced by Claude safely is to write the definition into the brief, ask for formulas and add a checks sheet that calculates the total two ways. Artifacts, available on every plan since 16 September 2026, temporary in-conversation visuals, creating xlsx, pptx, docx and pdf files in a sandbox, sharing options by plan, and Bookstore's monthly sales report.
Claude Memory, Chat Search and Incognito: What It Remembers and How to Manage It
Deleting a conversation in Claude does not delete the memory entries generated from it; you have to delete them separately on the Manage screen under Settings > Your Claude. Topic-based memory in 2026, the sensitive topics setting, project memory versus account memory, chat search, the limits of incognito, and working safely with customer data on the Bookstore team.
Claude Projects: A Persistent Workspace with Project Knowledge, Instructions and RAG
When Claude seems not to "see" a file in a large project, the reason is usually RAG: on paid plans, once knowledge approaches the context window, Claude reads the relevant passages it finds by search rather than every file. Project instructions, knowledge limits, project memory, sharing, and file naming and questioning rules for Bookstore's technical documentation project.
Getting Good Work from Claude: Instructions for Claude, Task Briefs and Verification Habits
The way to get usable work from Claude is to put rules that hold in every conversation into Instructions for Claude, job-specific context into a project, and the task itself into a brief made of goal, context, constraints, output format and an ambiguity rule. The personalization layers, how Claude picking its own tools affects the brief, a bad and a good brief for a Bookstore campaign e-mail, and three verification questions to ask before using an answer.
Choosing a Model and Effort in Claude: Which of Opus, Sonnet, Haiku and Fable for Which Job?
The first step to protect your Claude limit is adjusting effort, not the model: Anthropic recommends starting most work with Opus 5.5 and changing effort first when results fall short or the limit melts. The November 2026 model family (Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5), the model picker and effort levels, how Fable counts on each plan, and a comparison method using Bookstore's return policy draft.
Getting Started with the Claude Apps: Web, Desktop, Mobile, Plans and Usage Limits
In Claude, the limit is used up mostly by long conversations, large files and multi-step tasks that create files; usage is measured in a five-hour session and a weekly limit and tracked in Settings > Usage. Web, Desktop and mobile surfaces, the merge of chat and Cowork, the Free, Pro, Max, Ultra, Team and Enterprise plans as of 15 November 2026, and choosing a plan for the Bookstore team.
Improving Prompts by Measuring: Test Sets, Rule-Based Checks, LLM-as-Judge, Microsoft.Extensions.AI.Evaluation and Regression in CI with promptfoo
Measuring a prompt needs a golden set with reference answers and human labels, rule checks that run on every pull request, and an LLM-as-judge whose agreement with people has been measured. On Bookstore's return policy questions a short-answer rule cut a small model's accuracy from 8/10 to 5/10; rules and two judges agreed with the person on at most 23 of 30 answers, and a pairwise judge depended on order or gave no decision in six of ten pairs.