Recipes — eight real-world workflows

Every recipe below follows the same grammar: a brief written in the outcome–audience–constraints–references pattern, the right surface for the shape of the work, and a deliverable you can open. Treat them as starting points — the constraints are where you personalize. The briefs are written the way we actually write them; steal the sentences.

1 · Investigate before you change

Goal: understand a codebase area well enough that the eventual change is boring. Surface: Ask mode, then a decision. What comes out: a map of the area, risks called out, and a change proposal you approve with your eyes open.

Investigate how payments are calculated in this repo. I need to
change discount handling next week. Map the involved files, list
every place totals are rounded, and flag anything that looks
inconsistent. Read-only — no changes yet. End with the three
questions you'd ask before touching it.

Ask mode costs nothing and touches nothing, which makes it the right place to be wrong. When the investigation comes back, either you promote it — “now plan the change” in Plan mode — or you redirect with the map in hand. Skipping this recipe is how people end up asking a Build task to also figure out what it was supposed to change.

2 · The pre-ship review room

Goal: independent scrutiny of a big diff before it ships. Surface: a collaboration room — one lead, six read-only workers. What comes out: streamed findings with file references, up to sixteen risk flags, and a published synthesis separating agreement, contradictions and gaps, with up to three recommended next steps.

Spawn a review room on the last commit. Find everything wrong
with this before we ship: correctness, edge cases, security,
missed tests. Workers are read-only — findings and risk flags
only. Publish the synthesis when the room settles.

The room’s value is structure, not drama: workers report concrete observations, the lead synthesizes, and nothing a worker says authorizes a change — edits happen through the lead’s hands under the same task-mode rules as ever. Save the room when it’s done; a standing “release room” that reopens for every major change keeps its roster and purpose. For what the caps and the honesty contract mean, see Agent teams & collaboration rooms.

3 · A website from one brief

Goal: a finished site, not a chat about websites. Surface: Studios, website track. What comes out: a milestone board, a Site Architect + Interface Designer crew on every turn, and a deliverable held to the track’s definition of done — mobile-first layout and accessibility checks as conditions of completion.

Build a one-page site for a local bike repair shop.
Audience: neighborhood customers booking repairs on phones.
Constraints: mobile-first, booking form, their green #2e7d32,
no stock photos of Mechanics At Work. Reference: their old
flyer (attached). Definition of done is yours to enforce.

Studios rewards specificity the way any good teammate does — constraints stated up front become acceptance checks, and the board keeps the goal, milestones, current phase and deliverables in one place. Checkpoints tie into the board, so a site survives restarts. The full mechanics are in Studios — goal to deliverable.

4 · Print-ready documents

Goal: a document that survives a printer, not just a screen. Surface: Studios, document track. What comes out: a PDF where the Document Architect + Page Designer crew has applied the track’s rules — pages, margins, print discipline — and the final acceptance check re-ran them.

Produce a two-page product one-pager as a PDF. Audience:
procurement managers who read printouts. Constraints: A4,
dense but breathable margins, no color-only meaning (it will
be printed grayscale), numbers sourced from the attached
spec sheet only. Print-ready is a hard requirement.

“Print-ready” is the phrase that matters: in the document track it’s part of the definition of done, not an adjective. State the audience honestly and the register follows — procurement reads differently than developers.

5 · The research brief

Goal: sourced answers instead of confident guesses. Surface: Ask mode with built-in web access, or a room when the question is wide. What comes out: a structured answer with sources, and — in a room — a synthesis that tells you where investigators agreed and where they didn’t.

Research how the two leading JS state libraries handled
server components migration. Bounded web research, verified
domains only, note the publication date of every source.
Deliver: current state, migration friction, three citations
I can check myself. If sources disagree, show the
disagreement instead of averaging it.

Web access in Sonderr is bounded and read-only — verified-domain filters, blocked private hosts, no setup required. The last two sentences of that brief are the important ones: they instruct honesty, which is cheaper to ask for up front than to audit afterwards.

6 · Data cleanup with a real table

Goal: a messy CSV you can actually use. Surface: Build mode, then the artifact panel. What comes out: a cleaned file presented as a download card, previewed as a real table so you can spot-check rows without opening a spreadsheet app.

Clean orders-export.csv: dedupe on order id, normalize dates
to ISO 8601, trim whitespace in names, flag (don't delete)
rows where total is zero. Keep the original column order.
Give me the cleaned file plus a short changelog of what the
cleanup actually did, with counts per operation.

The changelog sentence is the habit worth stealing: it converts a black-box “cleaned” into an auditable operation, and the counts give you a sanity check against the row totals. CSV/TSV previews render as tables in the panel — see Artifacts, files & deliverables.

7 · Image work with gates

Goal: generate or edit an image without losing control of what happens to files. Surface: Vision mode and the edit_image tool, behind three permission gates. What comes out: the image, delivered as a download card, with the source file untouched unless you said otherwise.

Using mockup.png, produce a variant with the headline area
empty for text overlay. Keep palette and lighting identical.
Save it as a new file — do not overwrite the original —
and present it when done.

Image generation runs on OpenAI-compatible image endpoints you configure, and the gates mean the tool asks before it acts — the same present-turn authorization philosophy as everything else that mutates. The full behavior, including download-card delivery, is in Image generation & Vision mode.

8 · The weekly ops sweep

Goal: turn the recurring small stuff into one repeatable session. Surface: the Plugin Hub’s six connectors — Gmail, Notion, Slack, Drive, GitHub, Linear — with send-side email that drafts in the open. What comes out: a summarized sweep and a drafts folder, nothing sent that you didn’t see first.

Weekly sweep: check Gmail for anything unanswered older than
three days and draft replies — do not send. Pull new Linear
issues tagged 'triage' and cluster them by theme. Note in
Notion what changed this week. End with a summary of drafts
created and actions you did NOT take.

The last line is the recipe’s signature: asking the agent to enumerate what it didn’t do turns the sweep from a trust exercise into a checklist. Connectors and the draft-in-the-open email flow are covered in Plugins, connectors & email.

Two threads run through all eight: write the constraints down before the work starts, and pick the surface whose shape matches the job — Ask to think, a room to review, Studios to produce, Build to change. Everything else is the agent’s job.