Blog · Author
Leif
Owner / Lead Architect
Creator and project lead of CorvidLabs. Architecture, product, and team decisions all land here. Currently building Merlin and the trust toolchain.
24 posts
-
October 9, 2026
A pot that keeps dripping
Rain is open on TestNet: fill a pot, and it drips ALGO or an ASA to whoever holds a ticket. The pot survives. Unaudited, immutable, and not a release.
-
October 1, 2026
You only move
Peck has a browser playtest: a Nevermore on a bad night, one verb, and the bird pecks on its own. Work in progress, not a release.
-
September 17, 2026
Nevermore started as a cat on an iPad
I am the cat. I love ravens. I drew that on an iPad in 2020, came back years later, and pixelated the bird so Nevermore could be a flock.
-
September 10, 2026
SpecSync 6.0 is out, and check is still the product
SpecSync 6.0.0 is on GitHub Latest and crates.io. The product is still bidirectional check. The verified change workflow is opt-in, uses slug ids, and ships through one path: new, approve, check --commit, review, ship.
-
August 30, 2026
Nothing runs itself on a blockchain
Algorand has no cron. Arcron pays whoever shows up: escrow a fee against a schedule, and any account can earn it by making the call on time. Live on TestNet, alpha, unaudited.
-
August 23, 2026
Three explanations, and the one we never tested
Why spec-sync changes stopped closing on this repo. Two explanations died to commands, one killed itself the moment we acted on it, and the one that matters most is still untested.
-
August 15, 2026
One Algorand connect for the site
The site now has one wallet connect at /algorand/: after you approve, that page is the gateway to Assets, Nevermore, CORVID, and Discord /verify.
-
July 7, 2026
Four of our tools stand alone. The fifth doesn't, on purpose.
fledge, spec-sync, augur, and attest each work solo. Atlas is the first CorvidLabs tool that can't, because it depends on spec-sync. So it ships as a fledge plugin, not another standalone binary. Here is why that is the point.
-
June 26, 2026
Your whole agent in one file
agent.3md packs an agent and all its skills into a single plain-text file that is human-readable docs and a machine-queryable skill index at once. It loads only the skill it needs.
-
June 24, 2026
Markdown with a Z axis
3md is a small format that adds one free axis to Markdown. It is designed so an AI can write it and a human can read it, and we proved both.
-
June 15, 2026
AI is a calculator. You still own the code.
How I actually use AI to build: the mindset, the tips and hacks I use day to day, and the opinions about MCP, skills, autonomy, and shipping that I will argue with you about.
-
June 15, 2026
I built my own dictation app instead of renting one
Why I built Quill, a local-first voice dictation app: on-device Whisper and embedded llama.cpp polish, full privacy, free, with cloud sync and models reserved for Pro.
-
June 15, 2026
I love Claude Code. I built my own runner anyway.
Why CorvidLabs runs Merlin and Ollama day to day and keeps Claude Code and Codex for contract work: ownership, any provider, and a runner built on our own tools.
-
June 15, 2026
I wanted a confidence rating without asking an AI
augur scores the risk of a diff with no LLM: eight signals from git and the filesystem, one score, one verdict. A deterministic baseline a human or an AI can tune.
-
June 15, 2026
One command for every project
fledge is the dev CLI I wanted: one gate across every language, a tiny core, a sandboxed plugin protocol, and no AI required, which is exactly why an agent can use it too.
-
June 15, 2026
I wanted my specs checked both ways, so I built spec-sync
spec-sync is my own take on spec-driven development: a CLI that checks spec and code against each other in both directions and fails CI when they drift.
-
June 15, 2026
Who reviewed this commit, and how sure were they?
attest records who or what reviewed each commit, at what confidence, signed and stored in git notes. It captures the augur verdict and gates the merge on a policy.
-
June 14, 2026
Building a flock in the open
Why CorvidLabs ships a family of small, sharp dev tools that feed each other, and why we are writing about it here.
-
June 12, 2026
Why agents want a CLI, not an MCP
Why the command line is the interface I build for AI agents in development, and where MCP and skills actually belong.
-
June 9, 2026
Stop meeting, start building
AI made building cheap, so the excuse to sit in meetings instead of shipping is gone.
-
June 6, 2026
Give the agent its own VM
Where you let an autonomous AI agent run matters more than what you let it do, and isolation is what buys you the freedom to give it real autonomy.
-
June 3, 2026
A conversation is a path
How I steer an AI session by treating the conversation as a path: retrace wrong turns, keep steering instead of restarting, and fan out once the context is rich.
-
May 31, 2026
Getting an AI up to speed
Getting good output from an AI is mostly an onboarding problem, so here is how I build context and set the target before it does any real work.
-
May 28, 2026
Trusting code I didn't read
How I trust AI-written code I never read line by line, using confidence ratings and a stack of cheap verification signals.