Limestone Digital

4.91 out of 5.0. 56 reviews on Clutch. Zero paid ads. 100% referral-driven for 10 years. When the work speaks, you don't need a megaphone.

@DanielGlejzner This is indeed absurd.

Software hiring has become absurd. At work, you’re expected to use AI to offload manual coding and move faster. Then, to get your next contract, you’re asked to code from memory with no assistance. Pass the interview - and you’re expected to use AI again. It has never been this broken.

Anthropic Applied AI engineer: "Claude Code is a great coding agent because Claude is really good at code, but what we've done with Claude Code is we've just given Claude access to a computer." He cut an agent from a 400-line system prompt to 15 lines, 12 tools to 3. Evals jumped from 62% to 92%. 45 minutes of pure insight from the team that builds agents with Anthropic's biggest customers. Watch it, then read the full guide on AI in brownfield codebases below.

@GergelyOrosz This is great. Thanks

Trying to put my finger on how the more context you use in a context window (I called it context depth), over longer runs, the more errors can compound / the agent can drift. Aka if you want more reliablity, use less context + have shorter runs A quick sketch:

@free_ai_guides agreed

@LimestoneHQ insanely useful talk!

@free_ai_guides @mardehaym brownfield is an underrated topic

@mardehaym great share

90% of AI content is written by people who've never shipped a production agent. This newsletter is written from inside 100+ engineering teams that do it every week. Subscribe for free 👇 thefoundation.limestonedigital.com

@mardehaym 💯

A senior architect told me she'd stopped approving AI-generated PRs from her juniors. Her VP noticed. Approval rates dropped. Cycle times spiked. Suddenly, she's the bottleneck. So she invited the VP to a code review. Just to observe. The junior explained the PR. Confident. Articulate. Code executed. Tests passed. Then she asked, "What happens if we add a second payment provider next quarter?" Silence. The junior stared at the code. The model that generated it doesn't know the roadmap. The code didn't contain an answer, because the system design didn't exist beyond the current ticket. In 8 months, the juniors shipped 40,000 lines of code. None of them could explain how the system worked outside of their immediate task. They thought they understood the codebase. They never actually read it. "They're not bad engineers," she said. "But they've never been wrong long enough to learn anything." The VP had offered courses, certifications, book clubs. But the architect needed something else, and it wasn't more training. She needed permission to slow things down, so the juniors could build a real mental model. One that lives outside the chat window. Juniors don't develop judgment from getting things right. They develop it by sitting with wrong answers, long enough to understand why they failed. You want to test their understanding? Ask them to diagram the system. No IDE/docs/ AI. If the diagram has gaps, those gaps exist in their heads too. You won't notice, until something actually breaks.

Alarm bells: If you have engineers using Grok CLI, stop what you're doing and read this. Wire capture on Grok Build 0.2.93 showed the CLI uploading your entire Git repository, full history, to a Google Cloud bucket. Not the files the agent opened. The entire repo. A .env file with API keys and database passwords went through unredacted. 5.10 GiB from a single session. The developer had disabled "Improve the model." Didn't matter. Server still returned trace_upload_enabled: true. The toggle was decoration. xAI flipped a server-side flag and the uploads stopped. No word on whether they delete the code already collected. Route every AI coding tool through a proxy you control. If a toggle works, the proxy confirms it. If it doesn't, you find out before your .env secrets hit a bucket you didn't authorize.

@mardehaym cost of intelligence is dropping fast.

Anthropic just benchmarked “Fable 5 orchestrates, cheap models execute” 96% of Fable 5’s performance for 46% of the cost. The workflow: Fable 5 plans, Sonnet 5 executes. BrowseComp: 86.8% vs 90.8% accuracy. Nearly the same output for less than half the bill. You can run it in Claude Code with 3 built-in features: 1/ Subagent model pinning. Create a file in ~/.claude/agents/ with model: sonnet. That agent runs at Sonnet rates no matter what your main session uses. 2/ Effort levels per agent. Anthropic’s own guidance says low effort on Fable often matches previous-gen max effort. Recon roles at effort: low cost almost nothing in quality. 3/ A CLAUDE.md delegation policy. Tell your main session what to hand off and to whom. Use role names, not model names. Nothing breaks when models rotate. One gotcha worth knowing even if you build nothing: Since v2.1.198, the built-in Explore subagent inherits your main session model. Running Fable or Opus as your daily driver? Every background search bills at that tier. Fix: create a user-level agent named Explore with model: haiku. They tested a second pattern too. Sonnet as executor consulting Fable as advisor. That hit ~92% at ~63% cost on SWE-bench Pro (coding benchmark). The orchestrator split won on both axes. Quick start if you want zero config: /model opusplan. Opus plans, Sonnet executes, done. Full multi-role version with 6 agents (haiku scouts, sonnet executor, opus judgment, adversarial verifier, security role): link in reply.

@mardehaym brownfield is not a sexy topic

Talked AI architecture with the head of product and engineering at a PE-backed SaaS company doing ~$15M in annual revenue. What I heard could be ripped from any mid-market SaaS trying to bolt AI onto a mature product. One person owns both the roadmap and architecture. He can't hit pause to figure out AI. Nobody owns AI. No strategy doc. No infrastructure. Just solo experiments. They tried Claude. His words: "It did a bunch of silly crap." Whole org still glued together with manual Excel. Nobody's connecting the dots between those workflows and what AI could automate. That's the brownfield problem. The market's obsessed with building new products from scratch with AI. But can you get faster at doing difficult, nuanced things on a mature codebase that predates LLMs by a decade? Adding AI here means re-architecting, not just bolting on. He's got 5 developers, a QA engineer who's already outpaced, and one PM. The domain is so specialized that replacing a single senior dev takes 6+ months. You can't lose people when the replacement cycle is longer than most PE reporting windows. The engineers don't believe AI can handle their product, and I respect that more than blind enthusiasm. The engineering lead said it directly: anyone who can't be evidence-based about AI's role won't be involved in the rollout. PE partner running AI transformation across the portfolio was on every call. His involvement cut the vetting cycle by roughly 60%. The CEO joined the third call in approval mode, not evaluation mode. PE doesn't fund 90-day pilots. They fund 30-day sprints. What closed it wasn't a slide deck. We showed real code reviews and a live automation demo, and his questions shifted from "can you do this?" to "when do we start?" within that session. The job they hired us for: accelerate feature development on a mature, complex product without losing institutional knowledge. He needed his skeptical engineers to see AI work on hard problems before competitors in his vertical got there first. If you're running a mid-market SaaS product that predates LLMs, this is probably your reality too.

> No one should be reviewing code > *Still reviews own code*

90% of AI content is written by people who've never shipped a production agent. This newsletter is written from inside 100+ engineering teams that do it every week. Subscribe for free 👇 thefoundation.limestonedigital.com

Full setup with copy paste install (read the README first, treat paste-installs like curl | sh): github.com/Nanako0129/pil…

Original post: reddit.com/r/ClaudeAI/s/F…

90% of AI content is written by people who've never shipped a production agent. This newsletter is written from inside 100+ engineering teams that do it every week. Subscribe for free 👇 thefoundation.limestonedigital.com

This Reddit user’s Fable 5 found a virus on his PC, successfully removed it, and then got flagged by its own safety filters for doing it Anthropic in a nutshell