使い方
- インストールコマンドをコピーし、自分の環境へ導入する前にソースを確認します。
- 必要なツール、認証情報、対応環境を原文で確認します。
- 入力と成果物を指定し、出力を資料と照合します。
npx skills add https://github.com/dozaldevs/dozaldevs-skills --skill hire-ai-employee開始前の確認
- 導入にはモデル契約、API枠、ファイルへのアクセス許可は含まれません。
- スター数や導入数は活動の指標であり、独立した品質・安全性試験ではありません。
見出し、例、コード、表、リンク、参照画像を含む原文を表示しています。
Hire an AI Employee
This sets up a new AI employee (a Dozal) by chatting with the Hiring Dozal, the platform's own setup assistant. It asks questions, looks things up (which apps are connected, which Slack channel you mean), shows a finished Dozal, and you hire it.
Prerequisite: loadoperate-ai-employeesfirst. It installs thedozaldevsCLI and explains the key. The chat needs a key at Admin level or above; a lower key answers403 FORBIDDEN.
Concepts and response shapes: ../operate-ai-employees/references/customer-api.md. For one command's arguments, run dozaldevs <command> --help. This skill keeps only how the chat behaves. Answers are short; --detailed gives every field. --detailed needs dozaldevs 0.5.0 or newer (an older CLI reads the flag as needing a value).
What a Dozal can do
Say what the job needs in your own words; the chat picks the matching abilities. Every ability, with when a job needs it, is in references/abilities.md (generated from the platform's registry). An organization's own list, with only the apps it connected, is dozaldevs list-dozal-abilities.
How the chat works
The chat lives on the server. You open it once, then send one message at a time. Each send answers 202 straight away and the reply is written later (up to about a minute and a half when the first machine has to start). So every turn is: send, wait, read.
- Open. Keep the chat id.
CHAT=$(dozaldevs start-hiring-chat --data kind=create --field conversation.id)- Send. Read the current
turn_seqfirst, and send it back with the message.
SEQ=$(dozaldevs get-hiring-chat "$CHAT" --field turn_seq)
dozaldevs send-hiring-message "$CHAT" --data message="$MESSAGE" --data turn_seq:="$SEQ"- Wait. Run this, and run it again while it prints
true. It stops as soon as the turn ends, or after 80 seconds. The turn worked whenturn_seqwent up by one. If not, the lastmessagesentry witherror:truesays why.
dozaldevs get-hiring-chat "$CHAT" --while turn_in_progress=true --every 3 --timeout 80 --field turn_in_progress- Read. Run
dozaldevs get-hiring-chat "$CHAT". Show the human the new assistant entries inmessagesand every entry inopen_questions. - Answer. Send the next message with the current
turn_seq. Button presses go inanswers, keyed by the question'sid:--data answers.q1.option=<value>, or--data answers.q2.text="..."where typing is allowed.messagemay be empty when you only press buttons.
dozaldevs send-hiring-message "$CHAT" --data message="" --data turn_seq:="$SEQ" --data answers.q1.option="$OPTION"- Stop when ready:
presentedistrue,open_questionsis empty anddraftis set.
A long, specific first message (when it runs, what it produces, where results go, which apps) usually gets straight to a finished Dozal. "Make me a bot" always gets questions. Keep going until it's ready; stopping after the first question is the most common mistake.
The questions the platform asks itself
The assistant asks in words, and every setting it proposes shows on the change card, where the human keeps or leaves out each part. Never keep "just send it" for them, even to finish faster.
Three settings come as button questions about the held change (protocol app_account, listen_channel, frequent_schedule): which account of an app the Dozal uses, which Slack channel it listens in for a tag (or decline), and whether to keep a schedule that runs more than 24 times a day. Only the human's press counts; the change can't be saved until each is answered. Send the press with no words: words, or "change how often", go to the assistant and turn the change down.
When a send is refused
| Code | What to do |
|---|---|
409 TURN_IN_PROGRESS | Wait for the turn to finish, then send. |
409 DRAFT_CONTINUED_ELSEWHERE | Someone moved the chat on. Read it back, show what changed, resend once with the new turn_seq if the message still makes sense. |
503 HIRING_BUSY | No setup machine yet. Wait a minute and send again. |
409 CONTEXT_FULL | The chat got long. Restart it (below) for a fresh one on the same draft. |
409 CONVERSATION_CLOSED | It already finished (details.closed_reason). Open a new one. |
503 HIRING_UNAVAILABLE | Setup by chat is off right now. Tell the human and try later. |
dozaldevs restart-hiring-chat "$CHAT"Hiring the Dozal
Nothing saves on its own: the assistant holds a change to check, named by the chat's held_change (task_id, wait_id, action_id). Read it part by part:
dozaldevs get-wait-dozal-change "$TASK" "$WAIT" "$ACTION"Show each part's before and after, and the platform's own adjustments (platform, with reason). Approve all of it, or edit with the part ids to keep (steps, name_and_notes, apps_and_tools, tool:<path>, ...), then submit:
dozaldevs decide-wait-action "$TASK" "$WAIT" --data actionId="$ACTION" --data decision=approve
dozaldevs decide-wait-action "$TASK" "$WAIT" --data actionId="$ACTION" --data decision=edit --data 'keep_sections:=["steps"]'
dozaldevs submit-wait-decisions "$TASK" "$WAIT" --data mode=submitThe assistant saves it and reports in the chat, refusals included (a connection still to make). A chat message while a change waits turns it down; the assistant proposes it again if still wanted. hire-from-hiring-chat answers 409 USE_THE_CHECK_CARD with the held_change.
To throw the setup away:
dozaldevs discard-hiring-chat "$CHAT"The hired Dozal's id is archetype_id on the chat. Its slug is the stable name you run it by:
DOZAL=$(dozaldevs get-hiring-chat "$CHAT" --field archetype_id)
dozaldevs list-dozals --id "$DOZAL" --field items[0].slugAfter hiring
Tell the human it's live and offer to run it: "It's set up! Want me to run it now?" Then load run-ai-employee.