Skill Issue

Bad prompts waste good AI. Skill Issue catches missing context and sharpens your prompts before they reach Codex or Claude. Less back-and-forth. More building.

Comment

Inspiration

Bad prompts waste good AI. A missing constraint or an unclear goal can send a coding agent in the wrong direction, leaving you to explain again and untangle the result.

We wanted to catch that problem before the agent starts working. Skill Issue brings assistance into the moment you’re writing the instructions, helping you turn an unfinished idea into a clearer request.

The name is playful. The frustration it addresses is familiar: knowing what you want, but struggling to give AI enough context to deliver it.

What it does

Skill Issue improves prompts as you write them. It suggests inline completions and asks targeted, multiple-choice questions when an important detail is missing.

Suggestions draw on a searchable knowledge base of skills, example prompts, and agent workflows. You can preview the changes, accept or dismiss them, and keep control of your instructions.

You can also save named prompt versions, compare them, restore an earlier draft, and hand the refined prompt to Codex through our plugins.

How we built it

We built the companion editor with Next.js, React, and Tailwind, backed by a FastAPI and Pydantic service.

The enhancement pipeline uses Gemini to identify relevant topics, generate embeddings, assess retrieved knowledge, and produce either a completion or a clarifying question.

Our data layer combines PostgreSQL, pgvector, and TimescaleDB on Tiger Data for knowledge retrieval, prompt history, and telemetry. We also built a Snowflake analytics pipeline to study suggestion outcomes using derived features.

A shared MCP server connects the backend to Codex, giving the agent access to refined prompts, saved versions, and knowledge search.

Challenges we ran into

Keeping suggestions relevant while someone is typing required careful coordination. A response can arrive after the prompt has changed, so we track revisions and discard outdated suggestions.

Applying edits safely was another challenge. Question options can change several parts of a prompt, and those edits need to remain accurate across Unicode text and changing drafts.

Latency remains a major challenge. Retrieving knowledge, judging relevance, and generating suggestions improves grounding, but adds waiting time. We instrumented the pipeline to understand where those seconds go.

Integration constraints also shaped the product: we built a companion editor and explicit plugin handoff around the interfaces available.

Accomplishments that we're proud of

We connected the full journey: live model suggestions, knowledge retrieval, editable previews, persistent history, and working agent handoffs.

We verified real autocomplete and clarifying-question responses, then saved, restored, and retrieved prompts through installed plugins.

We’re especially proud of preserving user control throughout that flow. Suggestions require acceptance, changes can be inspected, and prompt contents stay out of telemetry logs.

What we learned

A useful prompt improvement can be a single question. Adding more instructions only helps when those instructions resolve something that matters.

We also learned that speed is part of suggestion quality. Even a relevant completion loses value when it arrives after the writer has moved on.

Finally, prompt history deserves attention: being able to compare and recover an earlier version makes experimentation easier.

What's next for Skill Issue

Our first priority is faster autocomplete through reusable retrieval results, smaller generation requests, and cancellation of obsolete work.

We also want richer project context and better evaluation of whether suggestions reduce corrections and rework.

The goal is simple: help people express what they mean before an agent builds something else.

Built With

Share this project:

Updates

Submission history