Skip to content

BERD TASK · 0013 · DO THE WORK

Run a working session

The daily task: one conversation with one agent that finishes something. A chat in Berd can make and edit files, run things and leave something behind. The outcome is a piece of work you have judged, in a session you can find again.

Steps at a glance

  1. Start a chat where the work has a home
  2. Bring context in with mentions and attachments
  3. Give strong intent and let the agent choose
  4. Queue follow-ups, and steer when it matters
  5. Watch context and cost, and expect compaction
  6. Stop, or fork, when it goes wrong
  7. Judge the output before it leaves the session

1. Start a chat where the work has a home

If the work belongs to an engagement, start the chat inside that project so it gets the project's description and folders as standing context. A chat with no project is a general chat: fine for a question, wrong for anything you will need next week, because its files land in the general artifact folder rather than beside the engagement. Choose project, agent and model from the toolbar before the first message; an agent cannot be invoked without a provider and model, so a silent start usually means one is missing. See Choose a harness for a session.

2. Bring context in with mentions and attachments

The composer's own hint says it: @ for agents and files, / for skills. The mention picker offers agents, references, paths, skills and files, and can reach the project root, your home folder or the filesystem root. Attach files or folders from the toolbar or by dropping them in.

[SCREENSHOT: T13-01 — the mention picker]

Do this instead of pasting background into the message. A mentioned file is the file; a pasted summary is your memory of it.

3. Give strong intent and let the agent choose

Say what you want and what shape the result should take, as you would to a colleague who has been in the room. Then let the agent pick the skills and, where it sees fit, delegate — the transcript shows the subagent's name and task when known. Morgan, Berd's design lead, on Block's On The Block (September 2026): give "a strong intent" and let the agent choose which skill to use.

4. Queue follow-ups, and steer when it matters

While the agent works, keep sending. Every message is queued in order and dispatched when the session is ready — including messages sent while the chat is still starting, which since v0.6.3 are no longer stranded. A queued message can be edited or dismissed without disturbing the others.

[SCREENSHOT: T13-02 — queued messages with edit/dismiss]

Steering pushes the first queued message into the running turn instead of waiting for it to end. The v0.6.1 release notes name the gesture — Enter with an empty composer while the agent is working — and a steer is refused while you have draft text or are editing a queued message. It needs a Goose backend that supports it; if Berd says steering is not available in this backend, the message waits in the queue. A steered message is marked Steered in the transcript.

5. Watch context and cost, and expect compaction

The toolbar shows a context ring with the percentage of the window used, the token count against the limit, and — when the provider reports it and the setting is on — a running session cost. Cost is an estimate and does not show for every provider; the ring is the reliable signal.

[SCREENSHOT: T13-03 — context usage and cost]

As the window fills, Berd compacts: older context is summarised to free space. You can trigger it from the toolbar; the auto-compaction threshold lives in Settings. One notice says compaction happened; another says the compacted conversation could not be verified because refreshed history did not arrive — your messages are still shown, and reloading the session is the fix.

[SCREENSHOT: T13-04 — compaction notice]

6. Stop, or fork, when it goes wrong

Stop generation from the toolbar when the agent has gone the wrong way; do not talk it down mid-turn. To try a different path without losing the good part, fork the session from a message — everything up to that point comes with you. Switching model mid-chat is allowed and sometimes fails; Berd says which model the chat is still using. One documented dead end: a chat whose earlier reasoning history a Claude model no longer accepts has to move to a non-Claude model or start again.

[SCREENSHOT: T13-05 — fork from here]

7. Judge the output before it leaves the session

Nothing is validated by having been produced in a chat. Read the result for what happened, not for how it reads. This is where the discipline of the method and the convenience of the tool quietly diverge — the same rule as in Buzz: Agents as colleagues, not bots.

Watch out

Compaction summarises. A number, a quote or a decision you need verbatim belongs in a file, not in the part of the conversation that will be summarised away.

Practitioner decision

Two rhythms, both supported by the sources. Queue-and-steer: send follow-ups as they occur, edit them in the queue, steer only when a correction must land now. One turn at a time: wait, read, then send; never steer. Recommendation: queue-and-steer for drafting work, one-turn-at-a-time for anything that writes to a client repository, because a steered correction mid-turn is hard to reconstruct afterwards. Confirm after A4, where the steer gesture is checked hands-on.

Last checked: Berd v0.6.4, 2026-09-11 — read from source and release notes, not yet tested in the app.

Learn more

Reference

Moment
apply
Takes
untimed
Steps
7
Verified
Berd v0.6.4 · 2026-09-11 · source-read
NEVER BESTUCK AGAINCLICK ME

Guide built with Berd, Claude and Codex · verified against Berd v0.6.4, 2026-09-11 · read from source, not yet tested in the app · Buzz Guide · About · not affiliated with Block