←
Limestone Digital1,792 posts
Limestone Digital avatar

Limestone Digital

@LimestoneHQ

AI T̶r̶a̶n̶s̶f̶o̶r̶m̶a̶t̶i̶o̶n̶ Implementation Partner with a decade of engineering heritage. 🟪 limestonedigital.com

◷ Joined May 2026▧ 174 Photos and videos
Limestone Digital avatar
Limestone Digital@LimestoneHQ·Jul 17

Some patterns we're seeing across our engineering clients that tell you where AI adoption is actually heading: 1. Companies are done spending blind on AI. A few months ago, most were approving unlimited API budgets for OpenAI and Anthropic with no tracking on where the tokens went. Today they're asking us to build governance layers and monitor every dollar. One client had three teams running up five-figure token bills before a single production agent shipped. Now they want dashboards that show exactly what's AI-generated, what it costs, and whether it's actually faster than adding headcount. That means more diagnostic work for us and meaningful savings for them. 2. Companies are accepting the real timeline. A few months ago, leadership expected to go AI-native over a few workshops and a handful of tool subscriptions. Now they're calling us because the codebase is seven years old and can't absorb automation without a full foundation phase first. Their developers are using AI at 60-70% adoption, but nobody's tracking what's generated vs. human-written. QA hasn't been touched. There's no governance. Going AI-native means rethinking org charts, handoffs, and process ownership across every department. Not a software purchase. 3. Companies are walking away from the big firms. The Accentures and McKinseys are all pitching AI transformations. Our clients keep showing up saying the same thing: "We paid seven figures for a strategy deck and nothing shipped." This is why firms like ours have a market. We operate at the engineering layer while having the business sense to map a process end to end, then build the agents to automate it. A year ago we had to explain this on every first call. Now prospects open the conversation with it. In the last 90 days we've taken calls with PE firms, law firms, healthcare companies, VPN providers, construction platforms, and retail tech businesses. Some are direct competitors with each other. Nearly all came through referral. AI SaaS won't transform your company. Giving every developer a Copilot subscription and calling it a strategy won't either. The only path to real operational ROI is combining a team that learns your business processes on the ground with engineering capacity to automate every manual workflow worth automating. We built Limestone Digital around this thesis ten years ago. AI made it urgent. If you're looking to transform engineering delivery or automate business processes with production-grade agents, visit limestonedigital.com.

Limestone Digital avatar
Limestone Digital@LimestoneHQ·8h

Want to put AI to work in your operations next? Get our free guide: how to pick your first workflow, measure the results and scale what works, with 3 case studies from healthcare, insurance and commercial lending. thefoundation.limestonedigit…

Limestone Digital avatar
Limestone Digital@LimestoneHQ·8h

Run it with claude --permission-mode plan. When the plan comes up, pick "Yes, manually approve edits" and approve only the ones you agree with. Claude Code enforces a permission rule on every matching call. A line in CLAUDE.md is context the model can skip. Put your one-way doors in the rules, and back the ones that must hold with the sandbox or a hook.

Limestone Digital avatar
Limestone Digital@LimestoneHQ·8h

<task> Write the approval policy for the agents that work in this repo: what they do without asking, and what needs a person first. Change nothing until I approve. </task> <inventory> List every action an agent here can take: built-in tools, MCP servers, scripts, CLI commands, cloud and database access, and any ticket, email or chat integration. Read the permission settings, MCP config, CI config and scripts to find them. Cite the file for each. </inventory> <classify> Two-way door: it can be undone before anyone outside the company notices. Local code and branches, test data, dev environments. Routine updates to colleagues in channels I name stay two-way. One-way door: it can't be undone, or it reaches someone outside the company. Production data, deploys, patient or financial records, claim or payment submissions, messages to customers, patients or vendors, force-pushing shared branches, deleting what the agent didn't create, access and secret changes. If you're unsure, mark it one-way and say why. </classify> <write> 1. A short policy section for CLAUDE.md or AGENTS.md: "Do without asking:" and "Ask first, naming the exact action:". Write "ask before" rules instead of bare "never" rules. 2. For every one-way door Claude Code's permissions can catch, including MCP tools by name, an ask or deny rule for the project settings, shown as a diff. 3. One-way doors a rule can't catch reliably, such as a script or npm command that deploys or sends email inside it, or a push to a branch CI deploys, listed for a person to gate at the tool, the API or CI. </write> <judge> Start a fresh subagent (not a fork) as the judge and give it the table and the diffs. It re-reads the cited files and checks that every one-way door has a rule or a listed gate, that each Bash rule matches the command as the scripts write it, and that no two-way door asks for approval. </judge> <output> A table: action, door, why, how it's gated. Then the diffs. Stop and wait for my yes. </output>

Limestone Digital avatar
Limestone Digital@LimestoneHQ·8h

Two-way doors are anything you can undo before anyone outside the company notices: code on a local branch, test data, dev environments. The agent walks through and shows you after. One-way doors can't be undone, or they reach someone outside the company: production data, deploys, patient or financial records, claim and payment submissions, emails to customers or vendors, force-pushing a shared branch, deleting what the agent didn't create. Anthropic's Fable 5.2 guide says to name those actions one by one, and to write "ask for approval before deleting files" instead of "never delete files." The model treats "never" as absolute.

Limestone Digital avatar
Limestone Digital@LimestoneHQ·8h

BREAKING: Claude Code can now write your agent's approval policy with Amazon's door rule: walk through two-way doors, stop at one-way doors. In Anthropic's test, one broad approval rule made Fable 5.2 hold back about half its routine requests to colleagues. The prompt:

Limestone Digital avatar
⟳ Limestone Digital retweeted
Limestone Digital@LimestoneHQ·Nov 17

BREAKING: Claude Code can now go through 60 days of your coding agent's mistakes and turn each repeat one into a lint rule, a test or a type, instead of another line in CLAUDE.md. The prompt is below:

Limestone Digital avatar
Limestone Digital@LimestoneHQ·10h

Want to put AI to work in your operations next? Get our free guide: how to pick your first workflow, measure the results and scale what works, with 3 case studies from healthcare, insurance and commercial lending. thefoundation.limestonedigit…

Limestone Digital avatar
Limestone Digital@LimestoneHQ·10h

Use the sentence on code your team owns and wants improved. Leave it out where the job is the smallest possible change, like a hotfix or code your team doesn't own. The judge step matters because the model writes the summary of its own diff. Make a second agent read the diff cold.

Limestone Digital avatar
Limestone Digital@LimestoneHQ·10h

<task> You are the owner and long-term maintainer of this codebase, and its design is up to you. Improve the structure of [module or folder] on a new branch. Never push and never merge. I'll open the PR. </task> <before> - Run the full test suite and save the result. - List the public functions, endpoints and CLI commands in this module, with their signatures. - If the tests don't cover what this module does today, write characterization tests that pin the current behavior first, and commit them on their own. </before> <work> - Merge duplicated code, delete dead code and fix nearby problems you find. - Any change to existing behavior goes in its own commit, with the reason in the message. - Run the tests before each commit. If a characterization test fails, revert the change, or put it in its own commit that updates that test and says why. </work> <check> When you're done, start a fresh subagent (not a fork) as the judge. Give it only the branch name, the public list from <before> and your PR summary. It runs git diff main...HEAD itself and lists every change the summary leaves out and every change to public behavior, with file and line. Put its list under your PR summary, unedited. </check> <output> The PR summary, the judge's list, tests before and after, and the commits that change behavior. Then stop. </output>

Limestone Digital avatar
Limestone Digital@LimestoneHQ·10h

The sentence, from Anthropic's Fable 5.2 prompting guide: "You are the owner and long-term maintainer of this codebase, and its design is up to you." Without it, the model tends to add to whatever structure is there, even when that duplicates code, and reports nearby problems instead of fixing them. With it, Anthropic saw three side effects: → 70 to 80% more output tokens → Changes nobody asked for, many of them new tests and docs, and in some runs a change to how an existing feature behaves → PR summaries that left out small changes A "no drive-by refactors" rule doesn't cancel it. The model still merged duplicated code in more than half of runs.

Limestone Digital avatar
Limestone Digital@LimestoneHQ·10h

BREAKING: In Anthropic's tests, one sentence made Claude Fable 5.2 merge duplicated code in about 75% of runs, up from under 10%. It also raised output tokens 70 to 80% and changed how existing features behave in some runs. The sentence, and a prompt that catches it:

Mark Ajzenstadt avatar
⟳ Limestone Digital retweeted
Mark Ajzenstadt@mardehaym·16h

I spent a lot of this year helping PE-backed companies put AI to work. Five things stood out. 1. Every fund wants a playbook it can use across the portfolio. I get it. A fund can’t build something from scratch for every company it owns. And the same few use cases do come up again and again. The catch is that each company already runs on its own legacy systems. You can take the same idea to the next company and still have to work out how to make it run there. 2. Buying everyone a license changes very little on its own. We talk to a lot of companies that have paid for AI tools across a whole department. People still do the work much the same way, and leadership is still waiting to see a difference in the P&L. 3. Start in engineering and watch where the work backs up. At Limestone, agents write about 98% of our code. Once we could write code that quickly, it started piling up in review. Then the bottleneck moved outside engineering. The team was ready to build, but somebody still had to decide what the software should do. 4. The same approach works outside engineering. Pick one workflow, give it a month, then decide whether to keep going. We built a research agent for a PE-backed healthcare billing software company. About six weeks in, their CPTO was testing it. At the first deployment, the agent answered 59 of 60 test questions correctly. When challenged on a wrong answer, it acknowledged the mistake and corrected it. The customer expanded the engagement. 5. The hardest part is changing how people work. When an agent takes over part of a job, somebody still owns the outcome. People need to know what they’re responsible for now, and they need to trust the agent enough to use it. In my experience, that takes more work than the technology itself. If I were running a fund next quarter, I’d pick one workflow at one company and write down how it performs today. Give the team a month, then measure what changed before asking the rest of the portfolio to roll it out.

Limestone Digital avatar
⟳ Limestone Digital retweeted
Limestone Digital@LimestoneHQ·19h

Don't tell Fable 5.2 "continue." Don't tell Fable 5.2 "never delete files." Don't tell Fable 5.2 "answer directly without deliberating." Anthropic's own guide says to replace all three. The model's thinking bills as output at $40 per million tokens. 6 lines to use instead: