Cowork vs Code
Choose chat collaboration or file-based work deliberately.
At Car-Mart Technology Group: Cowork (Claude Desktop) is for docs, specs, research, design, meeting notes — chat-style interface with persistent Projects. Code (Claude Code terminal) is for code, scripts, debugging, refactoring — CLI interface with file-system access. The decision question is not 'is this hard' — it is 'where does the work live' (chat scrollback vs your filesystem). The full long-form read with the 3-question decision tree, the Car-Mart-shaped vendor-spec example, the 30-min runway per lane, the what-it-looks-like-under-the-hood table, and the common questions lives at hool.dev/aitraining/cowork-vs-code (the companion doc, ~10 min).
The bridge is the work. Most Car-Mart work spans both surfaces. A bug fix in Code becomes a doc update in Cowork. A spec draft in Cowork becomes a feature in Code. The senior backend engineer on the api-refactor and the data analyst on the q3-dashboard move between them every day. The skill is the switch — notice the moment, and switch. Promotion from Cowork to Code is cheap (one sentence: 'save that to foo.py'). Demotion is not: a Code session that should have been Cowork wastes minutes spinning up a project, picking a folder, and confirming paths.
Three Car-Mart-shaped traps. People who only use Cowork: lose hours of work that was 'right there in the chat' (the 200-line script, the spec draft, the SQL query they finally got right). People who only use Code: spend 10 minutes setting up a project to ask a 10-second question. People who pick wrong: do both. The default is Cowork. Promote to Code when the work is growing past what you can scroll back to, you've said 'wait, I need to undo that' once already, you're about to copy-paste the same answer into a second file, a teammate will need to see the work, or you'll need to run this again next week.
I'm a [role] at Car-Mart Technology Group (seniority, primary language, one repo or Project in scope: api-refactor / q3-dashboard / runbook-update / dealer-portal / other). I'm starting [task]. Decide: should I use Cowork (Claude Desktop) or Code (Claude Code)? Give the reasoning (which of the 3 decision-tree questions tipped it), the first command or click to start (e.g., 'open Desktop, new chat' vs 'open Code, point at /path/to/repo'), and the expected output shape (one-shot text answer vs multi-file diff). End with the one Car-Mart TG mistake I should avoid on this kind of task.In your lane
Dev lane: T2 Cowork — paste a 5-line stack trace, ask 'what does this mean' before opening the project. T3 Code — point at the api-refactor folder, ask for 3 fixes ranked by safety, apply the safest one, ask Claude to write a regression test, commit. T4 Code + Cowork — use Cowork to draft the PR description and Code to make the change, save the prompt pattern to api-refactor's prompts/ folder. Concrete artifact: open Code → point at /api-refactor → new chat → paste the failing test → 'propose 3 fixes ranked by safety, apply #1, write a regression test' → commit → open Cowork → Project: api-refactor → 'summarize the PR diff in our PR template' → save both prompts to prompts/ folder.
Data lane: T2 Cowork — paste a SQL query, ask for explanation, then list the 3 edge cases it doesn't cover. T3 Cowork + Project — open the q3-dashboard Project, paste a stakeholder summary draft, ask Claude to rewrite it tables-first with KPIs from the data dict. T4 Cowork + Code — use Cowork to draft the stakeholder summary and Code to write the SQL-with-comments file, save the SQL+summary pattern to q3-dashboard's prompts/ folder. Concrete artifact: open Cowork → Project: q3-dashboard → new chat → paste the SQL → 'explain this query, then list the 3 edge cases it does not cover, then format the explanation as a stakeholder summary with 5 KPIs from the data dict' → save the prompt to prompts/ folder.
Ops/Security lane: T2 Cowork — paste an alert payload, ask for severity + hypothesis + next action. T3 Code — point at the runbook-update repo, drop in the new runbook, ask Claude to compare it to the template, flag deltas in escalation path or on-call rotation. T4 Code + Cowork — use Code to make the runbook diff and Cowork to write the #oncall postmortem summary, save the incident-triage pattern to runbook-update's prompts/ folder. Concrete artifact: open Code → point at /runbook-update → new chat → drop in the new runbook → 'compare to the runbook template, flag deltas in escalation path or on-call rotation, write the next runbook in the queue using the team's template' → commit → open Cowork → Project: runbook-update → 'draft the #oncall postmortem summary, lead with severity, list the contributing factors' → save both prompts to prompts/ folder.
Pick one task from your day. Apply the 3-question decision tree (durable? multi-file? shareable?). Open the right surface. Do the work. Save the prompt pattern to your Project's prompts/ folder. Then post the path to: (a) your team's Slack #ai-training, (b) the 90-day plan export, (c) the workshop's hool.dev/aitraining/cowork-vs-code companion doc as a reply with your 'I picked Cowork because…' or 'I picked Code because…' one-liner. The directors and CTO Josh are in the room to see what your team can ship today.