admin / The agents' front door
The front door, so the next agent reads less
CLAUDE.md, 6 onboarding documents, 14 task briefs, 8 prompts. Written for the agent that arrives next; the most honest description of how this site is actually run.
Most of the work on this site is done by agents, several at a time, on branches that merge into dev. Each one used to read the whole repository to learn it. Now there is a folder written for them, kept current in the same commit as whatever it describes, and rendered here so it can be read without the repository.
Read first
Read automatically by Claude Code. The ten rules that are not optional, the mechanics in one screen, and the short form of working alongside other agents.
The reading order by task. Ten minutes to being useful.
The Agent Behaviour Policy condensed: four objects, four barriers, the 23 primitives, the evidence tiers, the vault, the graph.
Parallel agents, branches, what conflicts and what to do about it, external state, ownership by surface.
Then, as needed
Every document, page family, script, test and register, one line each.
Where the site is and the ordered queue of what is next, with the decisions the lead owns.
Recipes: add a page, add a vault, cut a release, cut a Lab edition, write a brief, register a document, merge a branch.
Task briefs sized for one agent each, with the files each touches.
One file per in-flight branch. Read before claiming a task; delete on merge.
Prompts for the common jobs
Under .claude/commands/; in Claude Code each is a slash command. They start a job the way the workflows say it should be started.
Read docs/briefs/direction__abp-as-a-graph-and-stakeholder-views.md and docs/briefs/review__first-measured-abp-n8n-owner-key.md first; match their shape exactly.
handover
Merge the current branch into dev following .claude/onboarding/04-rules-of-engagement.md. Arguments: $ARGUMENTS (the version to claim and the release title; ask if missing and a page changed).
Add the page $ARGUMENTS following .claude/onboarding/05-workflows.md → Add or edit a page.
Add the vault $ARGUMENTS following .claude/onboarding/05-workflows.md → Add a vault and the model in .claude/onboarding/02-abp-model.md. Read site/vaults/gmail-readonly/data/grant.json as the example of a documented grant and…
You are a new agent on riskmandate.ai. Do not read the repository at large. Read, in order:
Cut release $ARGUMENTS following .claude/onboarding/05-workflows.md → Cut a release.
Work site/vaults/$ARGUMENTS/RESEARCH-NEEDED.md following §2 of docs/briefs/research__connector-grants-open-questions.md. For each question:
A rule in prose bounds nothing. These files work because every agent reads them first. That is the same sentence every behaviour-policy vault says about its own AGENTS.md, and it is as true here. The tests and the CI checks are the boundaries; the folder is the telling. People collaborating with us start at Working with us instead.