Chris Edgington

Since claude code moved to 1M token context - I have never hit the limit. Today I have hit it probably 10 times - yet I'm doing the same things I always do. Must be a newly introduced bug?

@c0nst @barbinbrad LOL - that's what I was going to say, either of them named Devin 😅

@barbinbrad both ai agents?

we're about to make two new job offers for carbon. we weren't sure about either at first. then they worked like crazy with open source contributions and gave us no choice. now we're confident that we've got A players. open source ftw.

@KentonVarda Yeah, almost as bad as "vibe coding" 😅

I find I can't bring myself to use the term "clanker". I mean, I know logically it's fine. AIs aren't people. I don't judge anyone else for saying it, in fact I think it's funny when they do. But when it comes to me saying it, something in my brain says "no".

// AGENTS.md Never, ever, under any circumstances, ever, not once, no matter what, try to start the fucking dev server, it’s already fucking running.

@wesbos @syntaxfm Now I'm working with two custom ERP vendors to integrate support for both products. One is open-source ... carbon.ms

@EdgingtonC @syntaxfm Do you work in the label/POS industry with these projects?

What are you working on? Send me your project. OSS, Paid, whatever. We're doing a @syntaxfm Syntax Highlight and we will review and/or roast your projects

@wesbos @syntaxfm I then got hired to help a company build their own custom internal ERP - which revealed the need for document templates, document generation, etc., in on-demand formats (pdf, zpl, png, html, etc.) - which is why I built BinderyPress.

@wesbos @syntaxfm I'm just a custom software consultant (been coding for 30+ years) ... got hired to build some ERP integrations a couple years ago encountered significant pain of printing from browser client and from cloud ERP to local printers. That's what led me to build ProxyBox Zero.

Oh claude, at least your are consistently untruthful:

@kellabyte IME - structured logging often just ends up being unstructured logging stuffed into JSON - often doubling the space-per-message. Just give me raw logs - a lot can be done with those.

Am I the only one who feels conflicted about structured logging? Whenever I look at logs in half of the observability tools it’s hard to view on the screen and JSON sucks. But advocating for an unstructured logging format these days would get you murdered I feel in most orgs.

Which non-password login mechanism do you prefer:

For a SAAS platform that does not provide passwords - would you rather receive a magic-link that needs clicked on (browsed) to finish auth or an OTP email that you enter into a waiting login box?

These are rules I apply in my own processes: Have a documented architecture that the agent must follow — not just code style, but actual structural rules for where logic goes and why. Never blindly trust agents. ____ I actually put an ARCHITECTURE.md in each project - it includes references to hand-written code that follows patterns I want followed. So when I have my review agent go over changes - I say "make sure all the changes meet the ARCHITECTURE.md requirements". And I've been burned too many times just trusting the agent when I was in a hurry. "Is that code ready to go?" "Yep" deploy ... its broken.

@EdgingtonC Got two more so we can round it out?

The things that have worked the best for me to keep Claude etc from writing complete trash code. (This assumes you’re using the top models at medium to high reasoning, and paying $200+/mo for a good plan, not $20.) 1. Excellent test suite that the agent has to run and fix if anything is broken, and write its own tests. By far the best way to improve outcomes. Also include linting, type checking, compiling, other static analysis tools and validations, and even access to a debugger if I can make it happen. 2. Excellent docs covering systems, code style, testing strategies, and more that I hand-wrote initially and that the agent has to keep up to date with every commit / PR. 3. An opinionated and carefully curated code base with well-named functions/classes/filenames, small files, extremely flat folder structure, and an AGENTS md that indexes and describes each concisely. Don’t let the intoxicating speed let this get out of hand. You’ll pay for it. 4. Review agents, using codex to review Claude and vice versa. I have Claude spawn codex reviews via CLI and it works super well. Also add in review checklists that it has to use before it’s done. 5. Well-written specifications that I hand-write and take my time on. 6. Review every line of every change that it makes and update docs, tests, or how I write specifications to ensure problems never happen again. 7. Run the agents at night so I am forced to improve everything above this one in order to not wake up to slop. 8. Be willing to hand-write features and bug fixes from time to time to make sure you stay in tune with the code base.

@mattpocockuk Yes, but it is GOOD that describing requirements is slow. It forces us to think. And we cannot think faster.

The main fatigue I'm getting with AI is communication fatigue Implementation is now crazy fast, but describing requirements is slow And pushing it faster leaves me knackered

@waynesutton @Fayeezashaikh Hmmm - went to that account - no DM option?

Building something cool with Convex and want to get featured? Send @fayeezashaikh a DM.

The Convex community didn't sleep this week.

@Rasmic Same. Tried three times. I'm not wasting time trying again until the grok coding agent comes out.

I keep going back to Opus

I usually have 4-5 claude code sessions going ... I follow the same routine. I spend extensive time in plan mode. After a plan is done, I have two custom agents that review and help revise the plans. Then I let claude run with the plans. Once done - I have a few prompts I use everytime to "finish" an implementation before ask the same two custom agents to review the code. Then when all that is done - I read all the code. After rounds of feedback / revisions ... then time to test.

what does your AI coding setup look like right now? single agent sessions? parallel worktrees? custom commands? something else entirely? curious how people are structuring their workflows beyond "open chat, type prompt."

I'd also love a real "cloudlfare tunnels for iot" management and control plane. My product "proxybox zero" relies on tunnels. As it continues to grow - tunnel management is getting to be a serious pain. I'm building my own dashboards - but it would be great if this was a first-class thing from cloudflare.

Just over week until @Cloudflare next Innovation Week! What should we announce?

@dok2001 @Cloudflare I would love "projects" - where I could group my project-relative CF resources. I've got so many workers / pages configured - such a pain to navigate.