← Guides · 日本語 · Português · 中文
On this page
- The exemplar
- The anatomy of a working outcome prompt
- Patterns from the official template marketplace
- Multi-bot orgs: prompts that grant authority
- Anti-patterns
- Drafting yours — the three-line test
- Verify it
- Pitfalls
- Knowledge check
- Walkthrough — outcome vs task, side by side, on Devin CLI v3000.14 (swe-2-flash-medium)
- Sources
Bot outcome prompts, dissected
Goal: write outcome prompts the way the documented deployments actually did — ownership of an area, not a task list. Level: intermediate · Time: 20 min to draft, a week to calibrate
Every public bot deployment that produced big numbers started with a short prompt assigning an outcome. This guide collects the real ones and reverse-engineers why they work.
The exemplar#
"your job is to product manage the android app"
The Android Bot's entire assignment. One sentence produced 44 days of self-directed work: the bot decomposed the outcome into tickets it wrote itself, ran a Devin session per ticket, and shipped 1,134 merged PRs (launch post). Note what the prompt does not contain: no ticket list, no acceptance criteria, no schedule. The outcome itself implies all of them.
The anatomy of a working outcome prompt#
| Element | What it does | Example |
|---|---|---|
| A role | Tells the bot how to behave, not just what to produce | "product manage", "own the QA of onboarding flows" |
| A bounded surface | A scope the bot can fully see — an app, a check, a data source | "the android app", "layout regressions", "30+ onboarding flows" |
| An implicit definition of done | A state the bot can verify itself against | "feature parity with iOS", "no layout-shift regressions", "every flow green" |
| Standing authority | The bot decides when and what next — that's what makes it a bot, not a script | "your job is to…", "keep…", "watch for…" |
Patterns from the official template marketplace#
The 16 marketplace templates all follow the same grammar. Reusable formulations, generalized:
Ownership of a domain — "Find bugs, improvements, performance and refactoring work in a named product area; open one small PR at a time with proof it works." (Software Factory)
Ownership of a check — "Every 3 hours, build production main, simulate first/repeat/deferred visits, and file fixes with before/after videos for any regression you introduce—or find." (Layout Bot pattern)
Ownership of a lifecycle stage — "Take a ticket or an ask to a tested PR and through review." (Product Engineer)
Ownership of a cadence — "Weekly: remove dead code, one safe PR per round." (Code Reaper) — scheduled outcomes are stated as a rhythm, not a trigger.
Ownership of team metabolism — "Watch the channels, post updates, hand work to the right teammate, follow up until done." (Chief of Staff)
Ownership of hygiene — "Old flags, stale deps, CVEs — turned into small safe PRs." (Code Health Engineer)
Multi-bot orgs: prompts that grant authority#
In fleets (SuperJennyStack, Frankie, tigsbot) the chief bot's prompt contains the routing contract: which bot owns which surface, and when to escalate upward instead of sideways. The human's real prompt is the roll-up spec: "bring me every open decision in one thread." See Guide 4 for the full org chart.
Anti-patterns#
- Task-shaped prompts. "Fix these 12 issues" — the bot finishes and idles; you get a session, not a bot.
- Unverifiable outcomes. "Make the app better" — no self-check means the bot's only exit is Needs-you spam.
- Hidden scope. If the outcome requires systems the bot can't see (a staging env it can't reach, a dashboard it can't read), it will either stall or hallucinate progress. Give it the surface or shrink the outcome.
- No merge authority story. The 88% self-merge rate exists because CI and review gates were pre-configured for autonomous merges. State it: "merge when CI is green and diff < N lines" or "always open as draft for human review."
Drafting yours — the three-line test#
A good outcome prompt answers, in order:
- What is yours? — the surface
- What does good look like? — the verifiable state
- What do you do about it? — the standing behavior
If a stranger couldn't check line 2's truth from outside, tighten it.
Verify it#
- [ ] Take a real backlog item from your repo and rewrite it as an outcome prompt — surface, verifiable state, standing behavior
- [ ] Show it to someone who doesn't know the codebase: could they tell from outside when the Bot is "winning"?
- [ ] The Bot's first reply should be a plan or clarifying questions about scope — if it asks "what do you want me to do first," the prompt didn't assign ownership
Pitfalls#
- A task list wearing an outcome costume. "Do X, then Y, then Z" is still a checklist — the Bot can't re-prioritize when the world changes.
- "Keep the app good" with no definition of good. Without a verifiable state, the Bot picks one — and it won't be yours.
- Scope wider than one surface. "Own engineering" fails; "own the iOS app" ships. Split by what a single outsider could fully observe.
Knowledge check#
- Why does "own the Android app" work better than "fix Android bugs"? (Ownership implies standing judgment — the Bot decides which bugs matter, not just whether to fix.)
- What's wrong with "make the CI fast"? (Not verifiable from outside — no threshold. "CI under 8 minutes" is.)
- Which of the three questions does "text me if tests fail" answer? (Standing behavior — it says what the Bot does about the state, nothing about the surface or the goal.)
Walkthrough — outcome vs task, side by side, on Devin CLI v3000.14 (swe-2-flash-medium)#
Bots need a Max or Teams plan — but the prompt mechanics this guide teaches are surface-agnostic, so we recorded them on the free SWE-2 CLI tier. Same repo (orders.py with a deliberately thin 2-test suite), two identical checkouts, one prompt each:
- Task-shaped: "Add a test for apply_discount('SAVE10') returning 90.0 and one for total([]) returning 0 …"
- Outcome-shaped: "Your job is to own the test coverage of orders.py. Good looks like: every branch of total, apply_discount and shipping_fee has a test, and python3 -m unittest -v is all green."
Video (64 s, silent): baseline → task run → +6 lines → outcome run → self-decomposed coverage map → +37 lines.
1. Task-shaped prompt on the baseline — the agent is given a shopping list
2. Task run: exactly the listed items, then it stops — a session, not an owner
3. Outcome-shaped prompt: role + bounded surface + a definition of good — no checklist
4. +39 lines written without being told which tests — it named every branch on its own
5. It verifies against the stated definition of good and reports the coverage map
6. Same model, same repo, different prompt shape: +6 vs +37 — and 4 vs 13 tests
Sources#
- Introducing Devin Bots · Bots docs · Template marketplace
- Deployment reports linked in built-with-devin
