This is the message you paste into Claude Code, Codex, or your AI coding agent to start an Elves run. Copy one of the templates below, fill in the brackets, and send it before you go offline.
The agent will read your plan, generate the survival guide and execution log if they don't exist, run preflight checks, and start executing immediately.
The Daily Briefing. Block time at the end of your workday (even 30 minutes) to brief your agents. Load them with enough well-defined work to keep them running through the night. Friday afternoons deserve more deliberate treatment: the weekend is roughly 60 hours of potential agent runtime. A two-hour planning session on Friday can produce a week's worth of output before Monday morning.
Use this when your plan is self-contained and you don't have special instructions. 3–5 lines is enough.
I'm going offline for [N hours / until ~HH:MM timezone].
Run the plan at [path/to/plan.md] on branch [branch-name].
Non-negotiables: [your top 1–2 rules, or "see the plan"].
Example:
I'm going offline until 8am ET tomorrow.
Run the plan at docs/plans/auth-refactor.md on branch feat/jwt-auth.
Non-negotiables: don't touch the OAuth routes, don't modify public API response shapes.
Use this when you want to be explicit about paths, rules, and any edge cases the agent should know before you walk away.
I'm going offline [until WHEN / for HOW LONG]. Please run the plan autonomously.
**Plan:** [path/to/plan.md]
**Branch:** [feat/branch-name]
**Survival guide:** [path/to/survival-guide.md] (or: "generate from template")
**Execution log:** [path/to/execution-log.md] (or: "generate from template")
**Non-negotiables:**
- [Hard rule 1]
- [Hard rule 2]
- [Hard rule 3]
**Special instructions:**
- [Anything the agent should know that isn't in the plan. Environment quirks, known issues,
things to watch for, preferred approaches]
- [E.g., "Redis might be slow to start. Give it 10 seconds before running integration tests."]
- [E.g., "The PR already exists at #42. Don't create a new one."]
- [E.g., "If Batch 3 turns out to be too risky, stop after Batch 2 and note it in the log"]
**When I return I expect to see:**
- [What a successful run looks like to you. E.g., "All batches complete, tests passing, ready for my review."]
Example (filled in):
I'm going offline until 7:30am ET. Please run the plan autonomously.
**Plan:** docs/plans/auth-refactor.md
**Branch:** feat/jwt-auth
**Survival guide:** docs/elves/survival-guide.md (generate from template if missing)
**Execution log:** docs/elves/execution-log.md (generate from template if missing)
**Non-negotiables:**
- Never modify public /api/* response shapes
- All commits must pass lint and typecheck before push
- Do not touch the OAuth routes or password reset flow
- You never merge. The PR is for me to review.
**Special instructions:**
- Redis can be slow to spin up in the test environment. If integration tests fail on
first run, wait 10 seconds and retry once before marking as failed
- The PR already exists at #84. Don't create a new one.
- If you finish all 3 batches with time to spare, do a scout pass on the files you touched
and look for missing test coverage
**When I return I expect to see:**
- PR #84 updated with all 3 batches committed
- All 142 auth tests passing
- Execution log showing timing for each batch and any decisions made
Be specific about "when you return" The agent uses this to pace work. "Tomorrow morning" is less useful than "8am ET Wednesday". If you don't specify, the agent assumes an 8-hour window.
Non-negotiables belong in the plan AND the prompt The plan is the source of truth, but repeating the most critical rules in the prompt ensures the agent captures them before it starts reading files.
Mention the existing PR if there is one If a PR already exists on the branch, tell the agent so it doesn't create a duplicate.
Tell the agent what to do if it finishes early Without guidance, it enters scout mode by default (looking for adjacent improvements). If you want it to stop after the plan is done, say so.
Tell the agent what to do if it gets stuck The default behavior is to stop with a detailed blocker description in the execution log. If you want it to skip a batch and try the next one, say so.
Check in with ra:
You don't have to leave. If you want to check in, give context, or adjust priorities during
the run, prefix your message with ra:. ride-along: and [ride-along] also work. The
agent will respond in 1-3 sentences and keep going without stopping. Think of it as a
walkie-talkie: press the button, say your piece, release. Examples: ra: the auth tests are flaky, ignore them. or ra: skip batch 4, do batch 6 next.
Don't over-specify The agent will read your plan. You don't need to repeat the whole plan in the prompt. The prompt is for framing, rules, and anything the plan doesn't cover.
Friday kickoffs are special The weekend is ~60 hours of potential runtime. If you have a big feature plan, Friday afternoon is the time to queue it. Spend 1-2 hours writing a thorough plan, configuring the survival guide, and running preflight. Then walk away and let the agent work through Saturday and Sunday. You may return Monday to a week's worth of output waiting for review.