Patterns 4 & 5: Socratic prompting and let it read the real thing

The first three prompting patterns — outcome-based, structured-task, and phased execution — all assume you know enough about the task to give Cowork a clear brief. The last two patterns handle the situations where you don't, and where the most valuable input you can provide isn't words at all — it's the actual files.

Pattern 4: Socratic prompting

When a task is complex or unfamiliar, flip the dynamic. Instead of briefing Cowork, let it interview you.

The template:

"I want to [TASK] so that [SUCCESS CRITERIA]. Before you execute, ask me the questions you'd need answered to do this well."

What happens next: Cowork surfaces its assumptions. It asks you the things it would need to know to produce the right output — the format you want, the audience you're writing for, the constraints you haven't mentioned, the edge cases it's unsure how to handle. You answer them. Then it proposes a plan based on your answers.

This inverts the usual sequence. Instead of you trying to anticipate every variable and stuff it into the initial prompt, Cowork identifies the variables and asks you about the ones that matter. The result is a more complete brief, assembled in half the time it would take you to write it from scratch.

When to use Socratic prompting: first-time workflows where you haven't yet built context files for the task. Cowork is effectively interviewing you to learn what your CLAUDE.md doesn't yet contain. Every question it asks is something you should eventually add to your context files — so the interview is also a diagnostic. Once you've done the workflow a few times and your context files are solid, you'll need this pattern less. Claude will already know the answers.

When not to use it: routine tasks with established context. If your CLAUDE.md already describes your voice, your output formats, and your constraints, adding a Socratic step just slows things down. Reserve it for genuinely ambiguous work.

Pattern 5: Let it read the real thing

This is Cowork's most underused capability — and arguably its biggest differentiator from Chat. Cowork reads. Don't summarize documents in your prompt. Point it at the actual files.

The instinct to summarize is understandable: you want to make the task easier for Cowork. But the summary is always a lossy compression of the original. You decide what matters. You might be wrong. The original document contains the full signal.

Some examples of what this looks like in practice:

Writing a deck? Don't write out the bullet points you want on each slide. Let Cowork read your prep notes, your briefing document, your research folder.

Writing an article? Don't describe your voice to Cowork. Let it read your last five articles and learn your voice from the source.

Building a report from multiple sources? Don't summarize each source in your prompt. Let Cowork read each source file directly and draw its own synthesis.

Generating slides that match a brand template? Don't describe the colors and fonts. Let it read the actual template file.

The prompt pattern is simple:

"Read everything in context/ first, then [task]."

Or more specifically:

"Before you start, read context/voice-examples.md, raw/meeting-transcript-2026-03.txt, and raw/competitor-analysis.pdf. Then write the article."

Why this works: Claude's context window holds up to 200,000 tokens — roughly 500 pages of text. A 10,000-word briefing document, five past articles, and a research summary all fit comfortably inside a single session. The information density of real source files is almost always higher than any summary you could write of them. When you let Cowork read the real thing, it draws on the full signal. When you summarize, it draws on your interpretation of the signal.

The two patterns in this lesson are inverses of each other. Socratic prompting is for when you need to put more into the brief. Reading the real thing is for when you have more than you could ever fit in a brief — and the answer is to stop summarizing and just point.