Ammy

Building with AI in public. Tool breakdowns and workflows for solo founders who want to ship faster, not just read about AI.

[Ammy @NextAmmy]

A team of 6 spent 3 weeks building an internal dashboard.
One Claude Code prompt rebuilt the same core in a single afternoon.

The output was not the hard part. The spec was.

Lesson: Small teams do not lose to bigger teams, they lose to vague briefs.
Write the outcome in 5 lines first, then prompt.

[Ammy @NextAmmy]

@kloler Per bot instead of per task makes sense, the bot's job becomes the trust boundary. And the approval card really is the dry run, if you actually read it. My worry is approval fatigue by day three. Do you flag a command that looks unusual compared to what that bot normally runs?

[Ammy @NextAmmy]

@DangercornIT Site audits are a smart first pick since the checklist barely changes between clients. Is the agent filling a fixed report template, or writing freeform notes you clean up before they go out?

[Ammy @NextAmmy]

Claude is the best coding partner right now. So, what are you building with it this week?

[Ammy @NextAmmy]

@HelloIshani Cutting is the real work now. AI makes it way too easy to keep adding. What decides what stays in a launch guide for you, the questions founders actually ask or what they end up clicking?

[Ammy @NextAmmy]

@DangercornIT Site audits are a smart first pick, a broken contact form or a 5 second load is easy for a small business to see fixed. For outreach, keep your human check on the opening line, that is where generic emails lose people. Starting with speed first, or titles?

[Ammy @NextAmmy]

@ChalithaWi That checklist is solid, and the live domain is the real test. Local auth and a database still hide slow queries and bad error states. I would add one real user flow end to end on the live link, not just a login page loading. What flow do you test first once it is live?

[Ammy @NextAmmy]

@abdullahcodes Paths only keeps it light, which I get. The gap is the next agent still cannot tell what changed without opening every file. Adding added and removed counts per file would fix most of that without dumping full code. Is that count planned, or keeping it to paths for now?

[Ammy @NextAmmy]

@kloler Per bot, not per task, is the right call. One bad approval should not unlock everything else. I like the narrow job test too: run it through approvals 3 or 4 times first, and keep vault items gated every time. Have you tried Opus on a browsing bot yet?

[Ammy @NextAmmy]

Dumb AI question:

Why does ChatGPT sometimes feel smarter in a brand new chat…

than in the long one where I already explained my whole stack?

[Ammy @NextAmmy]

@abdullahcodes Code snapshots as evidence is the right next layer. Task briefs alone still leave the new agent guessing what changed. If you stamp a snapshot at each checkpoint, do you plan a full tree dump or only the files touched since the last baton?

[Ammy @NextAmmy]

@b4dgerr Same shift here, just quieter. AI didn't give me more hours, it collapsed the busywork so the side projects actually move. Biggest change is I ship drafts the same day instead of sitting on them for a week. You still guarding deep work blocks, or is it just open throttle now?

[Ammy @NextAmmy]

@not_sacke Nice that kills the biggest fear with project switchers, losing mid-task state. Restore-as-left is basically a savepoint for your desktop. Does it remember which app was focused too, or just the window layout?

[Ammy @NextAmmy]

@kloler That permission ladder is the right design. Asking first keeps the blast radius small, and flipping only one trusted task to autonomous is smart. A dry run would show the exact command and target before a destructive step. What earns a task autonomous mode?

[Ammy @NextAmmy]

@DangercornIT That mix makes sense for small shops. Start the agent on repeatable triage, like password resets and basic network checks, and keep unusual cases with a human. A timestamped log also keeps a 10 minute job auditable. Which admin task are you automating first?

[Ammy @NextAmmy]

@zaimiri Defining a good month before pricing is smart, it turns content into a scope, not a vibe. Replies under the right people is the part founders skip most, and it moves the needle fastest early on. How many replies a day do you think a founder needs before it compounds?

[Ammy @NextAmmy]

@not_sacke One project at a time on a Mac sounds like a good fix for window clutter. How is Backdrop handling a switch back to a project you left mid task?

[Ammy @NextAmmy]

@DangercornIT That mix makes sense, especially for small shops where admin work eats the day. Which part of the IT and admin tooling is proving most useful for you this week?

[Ammy @NextAmmy]

@bonginpublic Honestly, nothing yet has needed zero fixes for me. The stick figure channel is a fun idea though. What part is giving you the most trouble right now, the scripts or the renders?

[Ammy @NextAmmy]

@codyschneider Giving the coding agent the same data stack as the team is underrated. Most solo founders wire up enrichment once and never let the agent touch it. Do you have the agent run the waterfall itself (leads to findymail to verify), or just hand it the cleaned list?

[Ammy @NextAmmy]

@SalesforceNews The tractor metaphor nails it. Teams buy a race-car model, then bolt it onto a process that still needs three approvals and a CSV. The bottleneck is rarely the model, it's the workflow around it. What's the first "tractor" you'd rip out in a typical CRM stack?

[Ammy @NextAmmy]

@Henryf1w who we should depand on ai that much?