Writing
One question per post, answered with something we measured. Every number names the test that keeps it true, the same rule the rest of the site follows.
CLAUDE.md vs AGENTS.md: what Claude Code loads, in 11 setups and 7 versions
Does Claude Code read AGENTS.md? We checked 11 file setups on 7 versions. It does now, but one CLAUDE.md anywhere above you switches it off.
Is git merge=union safe? 11 merges, 10 silent, 5 wrong
We ran merge=union through 11 merges in plain git. 10 exited 0 with no warning, and 5 of those left a file that was wrong. Here is which shapes break.
Does beads store issues in git? Not on your branches any more
Search still says beads keeps issues in a JSONL file in git. In bd 1.3.1 they live in Dolt. We ran five cases to see what that does to branches.
A Markdown task tracker in your git repo for AI agents: where status lives
Backlog.md, Pillar, Tasker, Yaks, Cairn. We merged two agents' edits to one task under each storage shape. Where status lives decides what conflicts.
Why your custom git merge driver did not run
We put one logging merge driver through 10 merges in plain git. It ran in 6. One miss left a conflicted file with no markers, holding only our side.
Jira, Linear or ClickUp for coding agents? Mostly the same choice
For an AI coding agent, Jira, Linear and ClickUp are one design — a hosted tracker behind an MCP server. What each gives an agent, and the other design.
I told my coding agent to search before grep. 0 of 10 did
A line in CLAUDE.md did not change what agents did first. Putting the records next to the prompt did. Ten questions, one repository, and every caveat.
An append-only task log still has to be merged by something
Two git task trackers both use an append-only log. We ran the same 24 merges through each. Which one is safer depends on where your merges happen.
Grep finds the right document. Then the agent has to read the others
We asked five real questions of one repository's docs. Search found every answer, in a pile of 42 to 64 matching files. What a link saves.
Why task files conflict when two agents edit them
The usual answer is git worktrees. We replayed 8,396 merges to find how often it actually happens, which type it is, and what worktrees defer rather than fix.
kadence on npm: the registry and the search page disagree
Searching for the npm package finds a distributed-hash-table framework. Installing it gets you a CLI. Here is what is actually published under the name.
kadence task tracker: what it does, and four times not to use it
Four different products called kadence track tasks. This is the git one — how it stores work, and the cases where one of the other three is the better answer.
kadence vs cadence: which one you actually want
Type it with a C and you get a baby-name poll. Six software projects share the sound, one of them is a real overlap, and this is how to tell them apart.
Which kadence did you find?
Eight products share this name and two more share its sound. A developer's map of all of them, and the one word that separates the CLI from the rest.
AGENTS.md or a decisions log: the costs nobody names
Everyone says use both. That is correct and useless — it does not say what each one costs, or which of the three failure modes you are choosing.
Backlog.md alternatives: the one difference that matters
Eight git-native task trackers, one real design split — a file per task, or an append-only log. What each costs when two agents touch the same task.
Your agent forgets everything. Four ways teams fix it
AGENTS.md, a memory service, a decisions folder, an append-only journal — what each one actually costs an agent to read, measured.
What an AI agent's context actually costs
Everyone measures the context window. Almost nobody measures what fills it before the agent reads a word — so we measured our own, and gave a number back.