Eric BONNET

building a small saas on the side. ai backend, real numbers, the occasional fail.

[Eric BONNET @1bonneteric]

@mattpocockuk The "label an issue, get an implementation" loop cut my own dev grind hard, I'm shipping ~3 small PRs/day now vs 1 before, mostly just reviewing Claude's diffs.

[Eric BONNET @1bonneteric]

@mattpocockuk The "fog of war" framing hits home. I skipped the grilling step on my last build, fanned out into 3 prototypes at once, and burned a week before realizing two were dead ends. Front-loading the frontier would've saved me that.

[Eric BONNET @1bonneteric]

@gregisenberg Bootstrapped doesn't mean cheap to run though. My side project hit $3k MRR but my LLM bill was eating 30% of it until I started routing through TokenReply to hot-swap to cheaper models per task.

[Eric BONNET @1bonneteric]

@mattpocockuk "fighting the good fight against slop" hits home. switched my backend last month and my regen rate dropped from ~30% to under 10%, cut my bill from $140 to $52/mo just from fewer retries.

[Eric BONNET @1bonneteric]

@v0 "Precision of a design tool" is what I keep wishing for in code-gen too. My LLM spits out clean components but I still burn ~3 hrs/week nudging spacing by hand.

[Eric BONNET @1bonneteric]

@mattpocockuk The skill-invocable tier is the one I keep wanting. Half my agent skills are internal plumbing that should never show up to the model as a callable option, but I still want to chain them from other skills.

[Eric BONNET @1bonneteric]

@mckaywrigley Still have my v1 prototype too, ran ~$340/mo in inference before I moved to TokenReply and hot-swapped down to Gemini for the cheap calls. Bill's at $90/mo now and the compute story for these teams keeps getting wilder.

[Eric BONNET @1bonneteric]

@gregisenberg On #4: started local to save cash but the GPU babysitting killed my nights. Moved to cloud and route through TokenReply so I can hot-swap when one model gets pricey. My bill went from ~$140 to ~$60/mo doing that.

[Eric BONNET @1bonneteric]

@mattpocockuk "Go back to sleep, Codex" hit different after I left a batch job running overnight and woke up to a $40 Claude bill for what should've been a $6 task. AFK coding is great until it isn't.

[Eric BONNET @1bonneteric]

@mattpocockuk The "over-indexes" part bit me hard. My agent picked up an auto-memory saying "user prefers terse code" and started stripping error handling everywhere. Killed auto-memories, kept a hand-edited rules file.

[Eric BONNET @1bonneteric]

@gregisenberg The "can't even code" part is the unlock most people miss. I shipped my whole app routing through TokenReply, swapping between Claude and GPT depending on the task, and never touched a model API directly.

[Eric BONNET @1bonneteric]

@v0 Reusable skills you just attach beats re-pasting the same context every gen. Did similar on my backend, pulled my prompt presets out of code and into TokenReply so I can hot-swap models without rewriting them each time.

[Eric BONNET @1bonneteric]

@mattpocockuk Mapping the design tree before prototyping is exactly where I was burning cash. I now sketch shape with a cheap model then hot-swap to Claude for the high-fidelity branch via TokenReply. Dropped my prototyping bill from ~$60 to ~$20/wk.

[Eric BONNET @1bonneteric]

@mckaywrigley "magic model smell" is real, you don't notice it til it's gone. I had a model deprecated mid-project once, scrambled to re-route everything through TokenReply just to keep my app sane.

[Eric BONNET @1bonneteric]

@v0 The batching elements into one prompt part is what I needed. Sending feedback one comment at a time was killing my iteration speed.

[Eric BONNET @1bonneteric]

@v0 Builds like this nudged me to add a live match-stats agent last month. Routing the play-by-play summaries through TokenReply so I can swap models per query cut my bill from $40 to $14/mo.

[Eric BONNET @1bonneteric]

@v0 Shipped a debug log that printed full prompts to my console... including a user's API key they pasted in. Caught it after 2 days. Now I scrub before anything hits my TokenReply calls.

[Eric BONNET @1bonneteric]

@v0 Been routing my prompt-builder feature through Claude for the last month, code-gen quality jumped enough that I dropped my retry logic. Nice to have it inside v0 too.

[Eric BONNET @1bonneteric]

@v0 Opus on Max is the same reason I keep Claude as my default for code-gen tasks. I route everything through TokenReply so I can A/B it against Gemini, and Claude still wins ~70% of my eval runs for refactors.

[Eric BONNET @1bonneteric]

@v0 "no setup needed" hits different — switched my checkout flow from juggling three SDKs to one integration last month, cut my deploy time from 20min to 4.