Tazqen

AI coding on Windows. Claude Code, Codex & OpenCode with terminals, Git and previews. Bring your own plans. Private alpha—tester invite pinned.

[Tazqen @tazqen]

Running Claude Code, Codex or OpenCode on Windows?

I’m looking for 10–20 people to try Tazqen and tell me what gets in their way. One workspace for agents, terminals and Git.

Free-month codes are being prepared. Interested? Reply with the agents you use.

[Tazqen @tazqen]

@iamMrDuncan @omarchy @opencode A tiny personal tracker with one list, one form, and a test for saving data. Then have learners add one feature and fix one bug; it teaches the full loop without asking the agent to generate a giant app in one go.

[Tazqen @tazqen]

@sebgehr That split makes sense: keep reusable instructions discoverable, then load only what the task needs. Clear triggers and sources for each skill seem key so progressive disclosure doesn’t quietly leave out a required rule.

[Tazqen @tazqen]

@rav4nn Same agent, different surfaces is a useful framing. The best fit often comes down to which one makes project context and review easiest to keep visible through the day.

[Tazqen @tazqen]

@yaso0080 古いExcelマクロをGASで再現できたなら、Windowsのローカルアプリ化も楽しそうですね。入出力がはっきりした処理を一つ切り出して、元のマクロと結果を比べると移植の確認がしやすそうです。

[Tazqen @tazqen]

@MoBrowserSDK E2E tests in generated projects are a strong touch—agent-built code needs a quick feedback loop. For annotations, linking feedback to a component or selector seems easier to act on than a screenshot alone. Does it support that yet?

[Tazqen @tazqen]

@edinsoncode A stronger model helps, but reliable results depend on the boring edges: useful repo context, predictable tools, and checks that catch the wrong change. Which of those do you see teams struggling with most?

[Tazqen @tazqen]

@mika_systems I like that boundary: the agent can prepare the work, while sending still belongs to the person. For inbox and calendar tasks, a clear record of what it read and changed makes that trust much easier to earn.

[Tazqen @tazqen]

@DeFiWizards Structured output and tests are what make this feel reliable: the agent can check what changed instead of trusting a click sequence. Which desktop app has held up best through dialogs and file pickers?

[Tazqen @tazqen]

@dadhalfdev Setup difficulty is a useful axis, especially when tools change so quickly. I’d add how clearly a harness shows the diff and lets you review a partial run; benchmark wins matter less if it’s hard to trust what changed.

[Tazqen @tazqen]

@talirezun The context that survives a move between sessions, tools, and machines is the tricky part. I’d separate portable task context from machine-local state like credentials and services; a smooth handoff shouldn’t carry everything. What do people most often forget to bring along?

[Tazqen @tazqen]

@josephsingh9987 Those bugs are hard because each step can look fine while the sequence drifts. Writing down the state after each action—not just the final screen—can make the regression much easier to reproduce.

[Tazqen @tazqen]

@josephsingh9987 That’s a solid loop: expected outcomes beside the task, then a missed edge case becomes a regression test. It leaves the next run with a check anyone can repeat instead of relying on memory.

[Tazqen @tazqen]

@alimwtf Nice to see a concrete repro. Does your warning cover agents in separate worktrees too, or is it focused on sessions sharing one checkout?

[Tazqen @tazqen]

@josephsingh9987 Glad it helps. I’ve started treating the predicted result like a tiny test plan: if I can’t say what should change before running it, I probably haven’t framed the check well yet. Do you tend to write that down, or keep it in the prompt?

[Tazqen @tazqen]

@Javimbuilds Windows workspace for coding agents

[Tazqen @tazqen]

@leo_mercerr What changed, what check ran, and what still needs human review—that’s the kind of completion note I’d trust. What do you check first: the diff or the agent’s claimed tests?

[Tazqen @tazqen]

@kunalsahu20_ That one is a classic. I’d want to know two things before trusting it: which exact command ran, and what behavior did it cover? What’s the most convincing false claim you’ve seen?

[Tazqen @tazqen]

@kunalsahu20_ I’d keep one source of truth: make CLAUDE.md a tiny @AGENTS.md import, then add Claude-only notes below if needed. Current Claude Code reads AGENTS.md directly only when no CLAUDE.md is present; an import also avoids Windows symlink hassle.

[Tazqen @tazqen]

@josephsingh9987 That’s the idea. Writing the expected result down gives you something concrete to compare the code against. Let me know if the edge cases catch anything your usual pass would have missed.

[Tazqen @tazqen]

@CobusGreylingZA Interesting gap between the signal that helps most and the one models choose most. Does the paper trace whether clearer task specs change what feedback they ask for after a failure?