Ayush

Senior Backend & Cloud Engineer | Shipping GenAI into Production | Creator of unwreck (Ctrl+Z for AI coding agents)

[Ayush @ayushdoesdev]

Anthropic's own data: junior tech hires dropped from 27% to 9% in a year.

Meanwhile, software job postings are up 15%.

So AI isn't killing engineering jobs.

It's killing the entry door.

If nobody hires juniors today, who becomes the senior engineer in 2030?

[Ayush @ayushdoesdev]

Vibe coding gets you to v1 in a weekend.

Then production asks:
• What happens at 10x traffic?
• Why is the API key in the frontend?
• Where are the backups?

AI writes the code.
Someone still owns the 3 AM page.

Who's on call for your vibe-coded app? 💀

[Ayush @ayushdoesdev]

AI coding agents are great until one runs `rm -rf` on files Git never tracked.

unwreck is Ctrl+Z for Claude Code, Codex & Gemini CLI.
Local snapshots of your whole working tree (untracked files included). One command rolls back the last turn.

⌘ cargo install unwreck

No cloud. No .git changes.

Check it out: https://t.co/thHcXe8aAV

[Ayush @ayushdoesdev]

Dependency bumps: the scariest PRs with the smallest diffs.

One version number changes.

Thousands of lines of someone else's code changed under you.

That's why I'm building Verspect: see what an upgrade actually changes before you merge it.

How do you vet upgrades today?

[Ayush @ayushdoesdev]

Added an index. Query still slow.

Usual suspects:

• Composite index, wrong column order
• A function on the column: WHERE LOWER(email) = ...
• Leading wildcard: LIKE '%abc'
• Stale stats → bad query plan

Run EXPLAIN before you guess.

Which one got you?

[Ayush @ayushdoesdev]

Exactly. Observability turns ‘we can roll back’ into ‘we’ll know we need to within minutes.’

One practical add we use: pair the 15-min check with a canary or feature flag so the blast radius stays small even if the alert is imperfect.

Makes the two-way door actually two-way. Great thread.

[Ayush @ayushdoesdev]

AI code review can turn into a slot machine.

Pull → new nitpicks. Pull again → more.

What keeps it sane:
• Define "done" upfront: tests pass
• Fix only critical findings
• Review with a different model than the author
• Cap it at 3 rounds

What's your stop rule?

[Ayush @ayushdoesdev]

@iamvishal16_ios Glad that resonates, untracked files are the exact case I built it for.

If you try it, a ⭐ and any feedback on the repo would help a lot while it's still early:

https://t.co/kzYqZxAFsU

[Ayush @ayushdoesdev]

1/
AI coding agents are great, until one runs rm -rf on files git doesn't track.

So I built unwreck: Ctrl+Z for AI coding agents 🦀

unwreck undo and everything's back.

Open source, single Rust binary.
https://t.co/kzYqZxAFsU

[Ayush @ayushdoesdev]

PM: why is signup so slow?

Me: we send the welcome email, resize the avatar, and sync the CRM before responding

PM: and?

Me: moved it all to a queue. Respond first, work later.

PM: so you just... procrastinate?

Me: we call it async architecture 💀

[Ayush @ayushdoesdev]

@shade_ofc Cool… actually owning the invalidation problem instead of just documenting it.

How does the auto-caching + partition scoping decide what to invalidate on a write, and can you still force a fresh read when you need it?

[Ayush @ayushdoesdev]

Me: API is slow, let's add a cache

Week 1: responses are instant 🚀

Week 2: "why does my profile still show my old name?"

Week 3: me writing cache invalidation logic at 2 AM

There are only two hard things in CS and I'm failing both at once 😭

[Ayush @ayushdoesdev]

“Love the sticky-note framing. One small upgrade we’ve been testing: after the ‘pain + rollback’ check, ask ‘How will we know it broke within 15 minutes?’ If the answer is weak (no metrics, alerts, or easy verification), treat even a two-way door as a one-way until observability is added. Keeps the speed while reducing silent failures.”

This builds directly on the two-way/one-way idea without contradicting it, adds a practical filter most teams overlook with AI PRs, and stays concise and positive.

[Ayush @ayushdoesdev]

The failing test gate is solid. One upgrade that works even better for me: force the model to first output the exact sequence of side effects (payment captured → DB timeout → webhook redelivery) as a numbered list before any code, then generate the test that asserts the desired end state after recovery. If the list has any ambiguity, the test is useless.

[Ayush @ayushdoesdev]

Exactly. The senior engineer filter forces the missing context (deadlines, blast radius, reversibility) that models never see.

Quick version: after AI flags something, ask “If this ships and breaks, how painful and how fast can we fix it?”

Low pain + easy rollback → ship.
Data/money/auth → fix.

[Ayush @ayushdoesdev]

@AlexBedeleu Same exception here. For auth/payments/migrations I also want the AI to explicitly show the failure path + rollback, not just say “looks good.” Style comments get ignored after round 1.
What’s your practical way of forcing that targeted failure case from the model?

[Ayush @ayushdoesdev]

@dovrikson This is the cleanest filter I’ve seen. Once you force the model to name a specific input + the exact line + what the user actually experiences, most of the “nicer abstraction” comments evaporate. The remaining ones are usually worth doing.

[Ayush @ayushdoesdev]

@BenjianDai This is the right filter. We use almost the same rule, but add one more question: ‘Would a senior engineer have blocked this on a human review?’ If the answer is no, we ship. AI is excellent at surfacing real risks, but it has no sense of when something is good enough.

[Ayush @ayushdoesdev]

Good q. Untracked files are covered. Each snapshot copies the whole working tree into a store outside the repo, so git status doesn't matter. Catch: a file is only recoverable if it existed at a snapshot (every prompt and before each tool call). Adding a proper "what's recoverable" section to the README today 👍

[Ayush @ayushdoesdev]

Try it:
- cargo install unwreck
- unwreck hooks --write

It's v0.1.0, so feedback, bugs, and ⭐ are all welcome 🙏

Has an agent ever wrecked your repo?

[Ayush @ayushdoesdev]

It's built so it won't make things worse:

- never overwrites edits you made yourself after the agent's turn
- snapshots live outside your repo, so git clean can't wipe them
- about 27 ms per snapshot on a 10k-file repo when nothing changed
- macOS, Linux, Windows