Module 02 -- Build it, run it, refine it

Prompting Best Practices for Cowork

You already know how to prompt a chat. This is what is different about prompting Cowork, anchored in real PE tasks. About an hour lab.

LabTBC, week of May 25

What we cover

Context in Cowork persists inside a Project (the right-panel Project instructions plus auto-accumulated Project memory). Treat a chat tab like a stateless conversation; treat a Cowork Project like a long-running workspace for one engagement. Lab anchor: an IC memo draft that builds across 3 turns inside one Project instead of one mega-prompt.
Multi-turn workflows: break a complex task into a Skill plus follow-up prompts. Do not try to one-shot a 10-step problem. Anchor: a board pack synthesis where the first turn pulls metrics, the second turn drafts the narrative, the third turn surfaces 5 questions for the CEO.
Output quality patterns: when to ask for a draft, when to ask for a polish, when to ask for a structured output (table, JSON, numbered list). Anchor: a market sizing where the first ask is a TAM number with assumptions in a table; the second ask is a 2-paragraph narrative built off that table.
Editing beats rewriting: "change paragraph 3 to be more direct" beats "redo the whole thing." Costs less. Reads better. Anchor: take a Cowork-drafted IC memo and run 3 targeted edits instead of a full regenerate.
Brief Cowork once, well: name the audience, name the format, name the constraint up front, not in iteration. Anchor: "Draft a quarterly portco update for the IC, 2 pages max, lead with 3 things going right and 2 things to watch."
Three side-by-side chat-vs-Cowork prompt examples on real PE tasks: IC memo draft, portco news scan, board pack synthesis. For each task, the same prompt and the same model -- different surfaces. The gap shows where Cowork wins (multi-turn refinement inside a Project, persistent Project context, MCP-backed retrieval) and where chat is still the right tool (one-shot factual questions, throwaway analysis, ad-hoc explorations).
Three common prompting traps that work in chat but break in Cowork. Vague follow-ups ("redo that" -- Cowork holds state across the conversation, so 'that' now has more to redo than you expected; specify exactly what to change). No audience naming (Project instructions cover some of this baseline, but explicit beats inferred every time -- name the audience inside the prompt). Format ambiguity (Cowork will infer a format if you do not name one; chat will guess. Name the format you want -- table, JSON, numbered list, prose).

Why it matters

You have spent a year prompting Claude and ChatGPT in chat tabs. Most of what you learned carries; some of it does not. This lab is the delta -- the 5 moves plus the worked examples plus the trap list that make Cowork prompting feel different from chat prompting.

Hands-on moment

Pick a prompt you used in chat recently (or one of the samples in the deck). Re-run it in Cowork using the 5 patterns above. Compare the outputs side by side.

Peer moment (3 min)

“Swap chat-vs-Cowork output pairs with a neighbor. Where did Cowork make a bigger difference -- the synthesis tasks or the structured-output tasks?”

Workbook steps

Context in Cowork persists inside a Project (the right-panel Project instructions plus auto-accumulated Project memory). Treat a chat tab like a stateless conversation; treat a Cowork Project like a long-running workspace for one engagement. Lab anchor: an IC memo draft that builds across 3 turns inside one Project instead of one mega-prompt.

Multi-turn workflows: break a complex task into a Skill plus follow-up prompts. Do not try to one-shot a 10-step problem. Anchor: a board pack synthesis where the first turn pulls metrics, the second turn drafts the narrative, the third turn surfaces 5 questions for the CEO.

Output quality patterns: when to ask for a draft, when to ask for a polish, when to ask for a structured output (table, JSON, numbered list). Anchor: a market sizing where the first ask is a TAM number with assumptions in a table; the second ask is a 2-paragraph narrative built off that table.

Editing beats rewriting: "change paragraph 3 to be more direct" beats "redo the whole thing." Costs less. Reads better. Anchor: take a Cowork-drafted IC memo and run 3 targeted edits instead of a full regenerate.

Brief Cowork once, well: name the audience, name the format, name the constraint up front, not in iteration. Anchor: "Draft a quarterly portco update for the IC, 2 pages max, lead with 3 things going right and 2 things to watch."

Three side-by-side chat-vs-Cowork prompt examples on real PE tasks: IC memo draft, portco news scan, board pack synthesis. For each task, the same prompt and the same model -- different surfaces. The gap shows where Cowork wins (multi-turn refinement inside a Project, persistent Project context, MCP-backed retrieval) and where chat is still the right tool (one-shot factual questions, throwaway analysis, ad-hoc explorations).

Three common prompting traps that work in chat but break in Cowork. Vague follow-ups ("redo that" -- Cowork holds state across the conversation, so 'that' now has more to redo than you expected; specify exactly what to change). No audience naming (Project instructions cover some of this baseline, but explicit beats inferred every time -- name the audience inside the prompt). Format ambiguity (Cowork will infer a format if you do not name one; chat will guess. Name the format you want -- table, JSON, numbered list, prose).

Workbook bundle

Print or open these alongside the session. They are the working surface for the steps above.

checklistOutput quality reference

Related artifacts

checklistOutput quality reference../../course-content/module-2-prompting-best-practices/output-quality-checklist.md

Source files live alongside this site under clients/greenridge-growth/; paths above are relative to the syllabus-site root.