Developer Uses OpenSpec to Build Offline Ship Inspection App for Friend

·1 views

A platform engineer built a desktop application to help a technical superintendent streamline the process of completing formal vessel inspection reports, which previously took over four hours each. The app, built using Wails, Go, Svelte, and SQLite, guides users through a 400-question checklist, attaches notes and photos, tracks progress, and exports a PDF — all offline on a Windows laptop. The developer used OpenSpec, a specification workflow tool, to keep plans and decisions documented alongside the code rather than buried in AI chat histories. This approach allowed work to pause for weeks and resume seamlessly across different AI agents and models. With the new app, the superintendent was able to produce a complete inspection report in under two hours.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer recreates tonkotsu ramen as interactive CSS art in a single HTML file

A developer submitted a CSS art piece depicting a tonkotsu ramen bowl for DEV Community's Frontend Challenge - Comfort Food Edition. The project is built as a single self-contained HTML file using CSS as the primary tool, with minimal JavaScript added only to randomize steam animation rhythm and enable a subtle cursor-driven parallax effect. Visual details such as the narutomaki spiral, chashu roll, and milky broth surface are rendered entirely through advanced CSS techniques including conic gradients, radial gradients, blend modes, and layered box shadows. The scene extends beyond the bowl to include a wooden table, bamboo trivet, chopsticks, and a stamped order ticket, evoking the atmosphere of a late-night ramen shop. The developer noted a stacking-context bug caused by a blend-mode layer rendering above the scene, which was resolved using isolation: isolate to contain internal blend modes correctly.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Better Tool Descriptions, Not More Tools, Fix Most AI Agent Failures

A software developer analyzing AI agent failures found that roughly 40% of errors stemmed from incorrect tool selection rather than flawed reasoning by the model. The core issue is that tool descriptions typically explain what a tool accepts but not when to use it versus a similar alternative. The author proposes a structured description format that includes purpose, use-case conditions, explicit do-not-use scenarios with redirects to correct tools, side effects, and reversibility. When selection errors occur, the recommended fix is updating individual tool descriptions rather than patching the system prompt, as prompt-level fixes are global, conflict-prone, and break across model upgrades. The piece also notes that pruning irrelevant tool schemas before each agent step can meaningfully reduce token costs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Match AI Tools to Task Failure Modes, Not Benchmark Rankings

A practical framework for selecting AI tools argues that the key question is not which model scores highest on leaderboards, but what kind of failure a given task cannot tolerate. Different work carries different risks: a robotic tone ruins marketing copy, a miscalculation corrupts a finance sheet, and a hallucinated fact undermines research. The framework maps task types to tool families — coding models with repo access for code, execution-enabled models for data analysis, retrieval-backed tools for current facts, and general chat models for open-ended writing. General-purpose models are appropriate defaults for soft, subjective tasks but become unreliable when a task has a mechanical ground truth, such as arithmetic or live information. Rather than trusting averaged public benchmarks, the author recommends running a small personal bake-off using real tasks to identify which tool best fits one's actual workflow.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Engineers Replace Manual Voice Clip Selection with Automated Three-Stage Filter System

Developers of a voice conversion app faced inconsistent and unscalable manual selection of 4-to-12-second audio anchors from voice actor recordings. To address this, they built a three-stage automated pipeline that filters audio clips for single-speaker content, natural pronunciation, and recording quality. The system begins with 48 equally spaced candidate windows per file, narrows them using pitch and energy analysis, then applies speaker embedding and SQUIM-based quality scoring. As a result, the minimum PESQ score of the anchor set rose significantly from 1.24 to 2.31, and the team compiled a library of over 70 usable anchors. The approach was originally documented in Japanese on the team's engineering blog.

0 comments Read more at DEV Community