#Agent prompts

Every surface that names a piece of work can hand it to a coding agent. An Agent Prompts menu appears on epics, quests and feedback, and on the Quests page, the feedback inbox and the blights inbox, and a click copies a ready prompt for Claude Code, Codex or anything else that reads a paste.

It starts off. The switch is Settings ▸ Quests ▸ Features ▸ Agent prompts, one click, once, per project. Until you turn it on the menus are not there, and that is deliberate: a project that does not hand work to agents should not carry a menu that does.

Nothing is ever sent anywhere. The prompt is rendered in your browser and put on your clipboard, and Lore neither transmits it nor keeps a copy of what you pasted.

#The prompts

Review asks an agent to sharpen a plan before anyone works it: take the decisions the spec left open, tighten the vague quests, and say plainly what it cannot answer alone.

Work on it on an epic hands the whole epic over, quest by quest: accept, work, verify, commit and complete each quest in turn. It sets no status. The first quest the agent accepts is what moves a ready epic to in progress, and the last one it closes is what completes it. It is not offered on a draft epic: that one's quests refuse to be accepted, and whether its spec is done is your call, not the agent's. Copying it changes nothing about the epic.

Work on it on a quest is the narrow version: one quest, one branch, one commit, and anything discovered beyond it is a comment rather than extra scope.

Work on it on a feedback item starts before there is a quest: it accepts the report, creates the quest linked to it, does the work, and ends by telling the reporter what shipped.

The last three name a surface rather than one item, so they carry no {{number}}, {{id}}, {{reference}} or {{title}}: there is nothing to number. Each one is a loop that re-reads its list every round, because it changes the list it reads. Triage the inbox accepts, rejects or asks about every pending report until none is left, and Triage the blights files a quest for each real defect and resolves the noise.

Work the loose quests hands over every quest in todo that belongs to no epic, one at a time, under the same rules as an epic's Work on it: one worktree and one branch for the run, each quest accepted before its first commit and named in every commit, completed when it lands, and one merge at the end once CI is green. It reads the list again after each quest, so a quest filed during the run is picked up rather than missed.

#Writing your own

Each prompt has a built-in default, and Settings ▸ Quests ▸ Agent prompts is where you replace it. Reset to default deletes your version rather than copying today's text into the box, so a prompt you have reset keeps following the built-in one as it improves.

The templates are written in English on purpose. Their words are the names of the tools an agent calls, so translating them would break them. The labels and toasts around them are localized; the payload is not.

#The seven placeholders

{{project}} and {{slug}} are not interchangeable, and this is the one thing worth getting right. An MCP call's project_name matches the project's title, never its slug. A project titled Kanban v2 has the slug kanban-v2, and a prompt that hands the slug to project_name answers "not found". Use {{project}} wherever you name the project to a tool, and {{slug}} only where you are building a URL by hand.

A placeholder Lore does not know is left in the text exactly as you wrote it, so a typo shows up in what you paste rather than blanking a line. And a placeholder that happens to appear inside a value is never expanded: a quest titled Fix {{url}} handling keeps its own title.

#A worked example

Every project's verify commands and branch rules are different, and that is the edit worth making. An epic Work on it template for a project that verifies with one command and does not let an agent merge:

Work epic {{reference}} "{{title}}" of the Lore project "{{project}}" to
completion, quest by quest.

The epic: {{url}}

Read it with `epic_get` (project_name "{{project}}", number {{number}}) and
its quests with `quest_list` (`epic: {{id}}`, `detail: "full"`).

Work in a git worktree, on a branch named after the epic. Never on main.

For each quest: `quest_accept`, do the work, tick each objective with
`quest_objective_set`, then run `make verify` and fix everything red before
moving on. One commit per quest. `quest_complete` with a short note.

When every quest is done, run `make verify` once more on the whole branch,
push the branch, and open a pull request. Do NOT merge to main: this project
merges by review. Then file an outcome folio under the epic with
`folio_create` and `epic_number` {{number}}.

That is the shape most edits take: the same skeleton, with the project's own commands in it and its own rule about who is allowed to merge.

#Who can change them

Reading the templates is open to any project member; changing one is the project owner's. The prompts are stored per project, not per member: one project, one set.