Curiosive
TR
Writing

Notes from the build.

What we have learned shipping AI-assisted software that has to survive in production.

  1. 8 min read

    Membership platform access: a checklist for Nordic teams

    A practical membership access checklist for English-working Nordic teams: separate identity, subscription state, permissions and CMS publication, then test a scheduled change.

  2. 9 min read

    Browser extension updates: a release checklist for existing users

    Plan browser extension updates for existing users: stored-data migrations, background restarts, permission changes and per-browser release evidence.

  3. 8 min read

    Multilingual website launches in Europe: a buyer’s acceptance checklist

    Scope a European multilingual website around one complete buyer journey: approved content, stable URLs, missing translations and a repeatable cached release.

  4. 6 min read

    AI-assisted design to code: preserve client intent and reuse real components

    An AI-generated screen should implement the approved design within the real application. Map components, expose missing decisions and review visual and interaction fidelity.

  5. 6 min read

    Client credentials in AI coding-agent history: a practical agency workflow

    AI coding-agent history can retain incidental credential copies. A recommended agency workflow connects prevention, coverage review and accurate cleanup boundaries.

  6. 6 min read

    AI-assisted agency progress reports: show accepted work and the next decision

    Give clients a clear view of accepted work, evidence and next decisions. A recommended reporting workflow for AI-assisted website and software delivery.

  7. 6 min read

    AI-generated client portals: a practical production handover review

    Turn a client portal demonstration into a defined operating agreement. A recommended agency review covers access, saved information, integrations and ownership.

  8. 6 min read

    From client brief to tested AI feature: a practical agency workflow

    Turn a broad AI idea into one reviewable feature. A recommended agency workflow connects the client task, implementation boundary, failure cases and acceptance evidence.

  9. 6 min read

    AI-assisted agency delivery: how Curiosive published a complete website content package

    A walkthrough of Curiosive’s own editorial delivery: AI prepared content and visuals, then source review, integration checks and release verification established the published result.

  10. 6 min read

    turbopuffer v3: why AI search needs more than vector similarity

    turbopuffer is redesigning storage beyond a vector-primary layout. AI search features need semantic intent, exact terms and authorized evidence evaluated together.

  11. 6 min read

    Gemini 4 Argon: evaluate long-running AI engineering tasks

    Argon targets extended AI work. Client engineering still needs scoped tasks, durable checkpoints, bounded execution and verified changes before release.

  12. 6 min read

    Muse Gadgets: voice AI hardware and the commercial boundary

    Muse’s hardware SDKs reveal useful voice-interface patterns. Its personal, non-commercial token program also shows why client products need a separate operating contract.

  13. 6 min read

    AstaBrief: build AI research reports readers can verify

    AstaBrief specializes in cited research reports. The client opportunity is an evidence workflow with inspectable sources, scope-preserving claims and reviewed outputs.

  14. 7 min read

    Cloudflare Clef: design a reviewable AI document and image intake

    Clef returns typed decisions and probabilities. A useful intake feature still needs clear labels, compliant inputs, validated policy and a reviewable next step.

  15. 6 min read

    Pi 1.0: connect AI coding agents to client tools with clear boundaries

    Pi 1.0 adds native MCP and scriptable tool composition. Useful development workflows connect approved context, bounded actions and evidence a reviewer can inspect.

  16. 9 min read

    Multilingual AI agents: fluent answers still need verified sources

    A recent English/Farsi agent study highlights a client integration problem: a fluent answer can hide weak retrieval. Evaluate the full evidence path in each language.

  17. 8 min read

    FLUX 3 Image: build a controllable AI asset workflow for your product

    FLUX 3 Image adds spatial controls and targeted edits. The useful client feature is a reviewable asset workflow with clear layout constraints and publishing boundaries.

  18. 9 min read

    SaaS onboarding usability: test the first useful task, not the tour

    An interface can feel obvious to its makers and still leave customers guessing. Observe the first useful task and improve the barriers the evidence reveals.

  19. 8 min read

    Web framework upgrades: a SvelteKit 3 checklist for existing products

    SvelteKit 3 is a prompt to plan an upgrade around the deployed product: compatibility, behavior and recovery, beyond a successful migration command.

  20. 9 min read

    AI monitoring tools: better leads start with visible evidence

    Lessons from the Inquirer’s Scrape story for internal monitoring tools: define relevance, preserve source evidence and measure the work reviewers actually use.

  21. 8 min read

    Database portability: prove the path out before a vendor change

    The Supabase and Turso announcement is a prompt to review dependencies. Here is how to prove a usable restore and a proportionate path out.

  22. 9 min read

    Local LLM deployment: hardware fit, data boundaries and operating cost

    DwarfStar puts local inference on the agenda. Here is how to evaluate one workload before buying hardware or turning a workstation into a shared service.

  23. 8 min read

    Apple Wallet pass integration: build the lifecycle behind the design

    A practical guide to the system behind a Wallet pass: entitlement, issuance, updates, scanning and a customer journey that survives failures.

  24. 4 min read

    Technical SEO for Developers: Myths and What Actually Matters

    Valid HTML, the keywords meta tag and a slightly better Core Web Vitals score do not rank a site. Here is what developers should get right in technical SEO.

  25. 3 min read

    Build Order for AI-Built Web Apps: Auth, Payments, Then AI

    One giant prompt makes a demo. A staged build order, with accounts first, payments second and AI last, makes a web product that survives real users.

  26. 3 min read

    Breaking Large Features into Slices AI Coding Agents Can Finish

    Hand an AI coding agent a big feature in one go and quality slips. Vertical slices, ticket graphs and a clear split between planning and execution keep it on track.

  27. 9 min read

    Production AI architecture: routing, cost controls and human review

    A practical guide to moving AI beyond the demo: model routing, spending limits, reliable fallbacks, evaluations and human approval.

  28. 3 min read

    Why AI Code Review Should Look Beyond the Diff

    Reviewers that only judge the changed lines miss the structural drift that makes a codebase harder to change. Here is how to widen the scope without drowning in noise.

  29. 3 min read

    Verification Loops for AI-Written Web Code

    Reading every AI-generated diff does not scale. Move rigour into the spec before the build and into automated verification after it.

  30. 2 min read

    Staging AI App Builds: Which Prompt to Use at Each Phase

    AI-assisted builds rarely fail on the first screen. They fail when real logic and structural change arrive. Here is how to match the prompt to the phase.

  31. 3 min read

    Design Before Delegation: Where Humans Matter With AI Agents

    When agents write the code, senior effort moves to technical design, task breakdown and the instructions agents follow. Here is how to spend it well.

  32. 3 min read

    Next.js Server Actions or API Routes: How to Choose

    Server actions remove a lot of plumbing from forms, but they are not the right tool for every mutation. Here is how we decide, and what to get right when we use them.

  33. 3 min read

    Keeping Your Team Fluent in Code That AI Writes

    When AI agents draft the code, the hard work moves to specs, supervision and shared knowledge. Here is how product teams keep control of a system they no longer type.

  34. 3 min read

    Quality Gates for AI Coding Agents: Rules Are Not Enough

    Instructions are advice, and agents forget advice. Checkpoints and enforced gates give AI-assisted builds a process that cannot be talked past.

  35. 3 min read

    What AI Code Review Bots Miss in Real Web Apps

    Automated reviewers catch textbook bugs but miss ownership checks, growth limits and framework idioms. Here is how to use them without trusting them blindly.

  36. 3 min read

    Why AI Coding Agents Can Degrade a Codebase One Feature at a Time

    Reviewing changes made by AI coding agents takes more than a passing test suite. Practical questions for planning, structure and refactoring.

  37. 3 min read

    What AI Coding Benchmarks Miss About Maintaining Code

    Most benchmarks reward solving one isolated task. Real products are extended feature by feature, and that is where AI-written code quietly degrades.

  38. 3 min read

    Why AI Coding Speed Gains Stall Before Reaching Production

    Teams ship more code with AI assistants, yet delivery gains stay modest. Larger pull requests, lower change confidence and weak foundations explain why.

  39. 3 min read

    Five Working Habits That Make AI Coding Agents Pay Off

    Teams that gain the most from AI coding agents change how they work. Five habits separate real delivery gains from a tool sprinkled on old process.

  40. 3 min read

    Why Passing Tests Is Not Enough to Merge AI-Written Code

    AI agents write code far faster than people can review it. Here is how web teams can rebuild review as a system instead of a habit.

  41. 3 min read

    Why AI Coding Agents Erode Code Maintainability Over Time

    Coding agents are rewarded for passing tests, not for good design. Here is why codebases degrade over time and how planning first keeps review manageable.

  42. 3 min read

    Internal Tools Built With AI: Access Control Comes First

    A polished dashboard is easy to generate. Real access control, persistent data and clear platform ownership are what make an internal tool safe to trust.

  43. 3 min read

    Guardrails for AI Coding Agents: Roles, Hooks and Isolation

    Long agent sessions drift, and prompts alone cannot stop a risky command. Structure can: scoped plans, narrow agent roles, hooks, isolated worktrees and reversible steps.

  44. 3 min read

    Interview First, Build Second: Requirements for AI Coding Agents

    Most AI-assisted builds go wrong before any code exists. Interviewing the requirements, sharing a glossary and planning in milestones keeps agents on track.

  45. 3 min read

    Securing AI Agents in Web Products: Least Privilege First

    AI agents act through tools, so a single hijacked prompt can do real damage. Here is how product teams can limit an agent's reach before it ships.

  46. 3 min read

    How to Write an AGENTS.md That Improves AI Coding Output

    AI coding output looks right until review. A short repository instruction file, detailed task briefs and built-in verification make it fit your conventions.

  47. 3 min read

    Why Multi-Agent Coding Needs Separate Builders and Verifiers

    Human attention, not model intelligence, limits AI-assisted delivery. Here is how role separation, validation contracts and written handoffs keep agent teams reliable.

  48. 3 min read

    Decision Quality Is the New Code Quality in AI-Assisted Builds

    AI makes clean code cheap. What separates good products now is the quality of the engineering decisions around it, and the evidence that proves they work.

  49. 3 min read

    Why Typed Code and Strong Fundamentals Matter More with AI

    AI tools draft code quickly, but types and core engineering knowledge decide whether that code is safe to ship. Here is how we think about the split.

  50. 3 min read

    How to Measure the Real Impact of AI Coding Tools

    Acceptance rates and lines of code make AI tooling look productive without proving it. Here is a better way for product teams to measure what AI coding actually delivers.

  51. 3 min read

    Designing an AI-Assisted Delivery Pipeline for Web Products

    Coding agents work best inside a defined delivery loop. Here is how product teams can place human checkpoints, verification and feedback in that loop.

  52. 3 min read

    Context, Guardrails and Tests: Running AI Coding Agents Well

    AI makes code cheap to write, but reliable delivery depends on context, guardrails and verification. Here is how web teams manage all three.

  53. 3 min read

    Scheduled and Event-Driven AI Agents for Software Teams

    Most AI coding agents wait for a prompt. Here is how web teams can put them to work on recurring chores, with clear triggers, the right context and human oversight.

  54. 3 min read

    A Plan-First Workflow for Working With AI Coding Agents

    Most agent frustration comes from skipping the plan. A repeatable explore, plan, code and commit loop keeps AI-written changes reviewable and under human control.

  55. 3 min read

    How to Read AI-Generated Code Before It Reaches Production

    Writing code is getting cheaper. Understanding it is the scarce skill. Six habits for reading unfamiliar, AI-written code quickly and catching what the model missed.

  56. 3 min read

    Review Depth by Blast Radius: Feature Gates for AI-Written Code

    AI agents now produce more code than one engineer can read line by line. Scaling review effort to blast radius, with feature gates as a safety net, keeps releases fast and safe.

  57. 3 min read

    Agentic Engineering: Where AI Coding Agents Can Be Trusted

    AI coding agents now get code right first time, which is exactly when teams stop checking. Here is how to decide what to delegate and what to keep.

  58. 3 min read

    Giving AI Coding Agents the Right Context, Not More of It

    Teams often fix a confused AI coding agent by adding more context. The better fix is structured context: skills, live data access, retrieval and memory.

  59. 3 min read

    From AI Demo to Production: The Engineering Work in Between

    A working demo and a production ready AI feature are built from different disciplines. This is the engineering work that separates the two.

  60. 3 min read

    What Makes AI Coding Agents Production Ready

    The model behind an AI coding agent is the easy part. What decides whether it works in production is the harness, the layered architecture and the review process around it.

  61. 3 min read

    What AI Coding Agents Get Wrong That Tests Do Not Catch

    AI coding agents rarely write broken syntax anymore. The defects that matter now are silent assumptions and concurrency bugs no test suite surfaces.

  62. 3 min read

    The Operating Model Behind Reliable AI-Assisted Coding

    Faster AI code generation does not by itself produce reliable software. The discipline around context, verification and parallel review decides the outcome.

  63. 3 min read

    What Next.js Caching Teaches About Briefing AI Coding Agents

    Next.js 16's cache components and AI coding agent context both come down to the same discipline: deciding on purpose what a system is allowed to assume.

  64. 3 min read

    Turning Code Review Into a Self-Improving Feedback Loop

    As AI agents write more code, reviewing every line stops scaling. The fix: review automation that learns from a team's own history and gets checked the way tests check code.

  65. 3 min read

    How Web Teams Keep Control When Running Multiple AI Coding Agents

    Running multiple AI coding agents at once only works with scoped tasks, isolated work and independent review, not a single unmanaged prompt.

  66. 3 min read

    Why AI-Generated Code Needs Outcome-Based Review

    As AI writes more of the code in a typical project, review has to shift from checking syntax to validating that the result actually meets the business requirement.

  67. 3 min read

    Why Fast AI-Generated Code Still Needs Production Engineering

    AI coding tools ship changes fast, but speed without architectural review creates a new kind of technical debt that costs far more to unwind.

  68. 3 min read

    The Line Between an AI App Builder and a Production App

    AI app builders can take an idea from description to a live prototype in minutes, but turning it into a product people can trust needs a different kind of work.

  69. 3 min read

    Why AI-Generated Code Needs Tolerances, Not Just Trust

    AI coding tools are non-deterministic, which changes what counts as correct in production code: why directing, reviewing and refactoring output matters more than trusting it.

  70. 3 min read

    What Tokens, Context Windows And Hallucinations Mean For Code

    Understanding tokens, context windows and hallucinations helps product teams catch the real risks in AI-assisted code before they reach production.

  71. 3 min read

    Spec-Driven Development: A Better Path Than Vibe Coding

    AI can generate a polished front end in minutes, but a working product needs a specification, a review step and senior engineering judgement behind it.

  72. 3 min read

    A Risk-Based Framework for Trusting AI-Generated Code

    Most teams either ban AI-written code outright or let it into everything. A tiered approach, matched to review infrastructure and task scope, works better.

  73. 3 min read

    Code Review Is the Real Bottleneck in AI Assisted Development

    As AI agents write more code, review becomes the constraint that decides whether a product is safe to ship. Here is what a layered process looks like.

  74. 3 min read

    Why Faster AI Coding Doesn't Speed Up Software Delivery

    AI can write code fast, but speed at the coding step rarely shortens delivery. The gains come from disciplined review, not from typing faster.

  75. 5 min read

    Vibe Coding Risks: How Senior Developers Use AI Strategically

    Vibe coding creates production fragility with code no one fully understands. Strategic AI use for boilerplate and relentless review preserves debugging skills and system intuition.