Instructing Devin effectively

Cloud Devin burns the most ACUs when the first message is vague. Official guidance is blunt: be as specific as possible — provide the same detail you would give a coworker with a written spec.

Hub: Tokenminning in Devin. Metering: Cloud ACUs, Token usage.

What makes a prompt work

From Instructing Devin Effectively , strong prompts combine:

Element	Why it saves ACUs
Concrete scope	Repo, files, subsystem named — less exploration
Task breakdown	Numbered sub-tasks with logical order — fewer dead ends
Success criteria	Tests, CI, status codes, UI checks — Devin self-validates instead of guessing
Pointers to patterns	Filenames, doc links, example modules — reuses existing design

Do make design decisions for Devin:

Optimize the getOrderDetails query in orderService.js by adding a composite index on order_id and product_id in order_items. Replace the correlated subquery with a JOIN to products. Run npm test after each iteration.

Don’t leave architecture open:

Improve our database performance.

Pre-task checklist

Before starting a cloud session, from When to use Devin :

Task definition

Clear start and end
Explicit success criteria (tests pass, CI green, matches existing pattern)
For complex work: scope with Ask Devin first — investigate codebase and generate a high-context prompt before implementation

Available context

Example files or modules to mirror
Links to API docs, design specs, or Figma via MCP
Relevant MCP integrations (Sentry, Datadog, databases) connected before the session

Success validation

Test suites, lint, build steps Devin can run autonomously
Browser QA instructions: spin up app, navigate flows, screenshot at breakpoints
Devin Review with Auto-Fix for closed-loop PR iteration

Task size

Rule of thumb: if you could finish in ~3 hours, Devin can likely handle it
Larger work: split into sub-tasks — one session per sub-task, run in parallel when independent
Target XS–M session size in Session Insights (L/XL = unhealthy)

Session size as a guardrail

Size	ACU (self-serve)	Messages	Action
XS–M	≤ 10	≤ 10	Healthy — replicate prompt pattern
L	≤ 20	≤ 20	Review Issue Timeline; split next time
XL	> 20	> 20	Break task apart; fix environment gaps

High ACU with few user messages → autonomous retries (check environment). Many messages with low ACU → underspecified initial prompt (use Improved Prompt from Session Insights).

Validation and checkpoints

Tell Devin how to prove correctness before opening a PR:


When modifying the API, confirm the endpoint returns 200 and includes fields id, name, status.
Run npm test after each checkpoint.
Ensure CircleCI passes before pushing.
Pass ESLint/Prettier before any commit.

For UI work, ask Devin to test in its browser: run npm run dev, navigate to the page, record the flow, compare against design specs.

Playbooks vs Knowledge

Mechanism	Use for	Token impact
Playbooks	Repeatable step-by-step procedures (CI triage, release checklist)	Reusable prompt — less retyping
Knowledge	Org conventions, deployment steps, common fixes	Retrieved when relevant — trim misleading items flagged in Session Insights

Playbooks = task procedures. Knowledge = persistent context across sessions. Do not duplicate the same prose in chat, Knowledge, and AGENTS.md.

After the session

Open Session Insights — read Improved Prompt and Action Items
Update Knowledge if Misleading Knowledge was flagged
Save effective prompts as Playbooks for the team
Fix recurring machine-setup issues so future sessions start clean

Devin hub
Cloud ACUs
Devin CLI — local work and /handoff
Rules — Desktop context cost
Good vs bad instructions
Prompt templates cheat sheet