Claude Cowork output quality issues

TL;DR. Output quality is rarely a model problem; it is almost always a specification problem. Run the five-rung diagnostic ladder before reaching for Opus. Most issues resolve in the first three rungs — prompt specification, CLAUDE.md coverage, source-input quality.

The diagnostic ladder#

Run through these in order. Most issues resolve in the first three rungs without changing model.

Rung 1: Is the prompt specifying enough?#

Check against Prompt anatomy:

• Does it name input + output + format?
• Does it set length and voice?
• Does it cap word count or set a quality bar?

If no to any of these, fix the prompt before changing anything else.

Rung 2: Is CLAUDE.md doing its job?#

• Is the workspace's CLAUDE.md present?
• Does it have a house-style section?
• Does it include three examples of good output?
• Has it been updated in the last 30 days?

If no to any, fix CLAUDE.md.

Rung 3: Is the input clear?#

• Are the source files readable — not corrupt PDFs, not low-res scans?
• Is the source content actually answering the question being asked?
• Is the source folder cluttered with off-topic files Cowork is also reading?

If no, fix the input — re-scan, prune, restructure.

Rung 4: Is the model right for the task?#

• For agentic and structured work — Sonnet 4.6 is right.
• For deep reasoning (legal, regulatory, complex synthesis) — try Opus 4.6.
• Switching to Opus for routine work is rarely the answer.

See Models and context.

Rung 5: Is iteration happening?#

• Did you accept the first draft?
• Did you give Cowork specific feedback ("shorten the executive summary by 30%; remove the passive voice")?
• Iteration usually doubles output quality between draft one and draft two.

The teams that complain most about quality are the teams that ship draft one. The teams that ship draft two-or-three rarely have quality complaints.

Common quality complaints and fixes#

The voice problem (specifically)#

• Three good and three bad examples in CLAUDE.md beat any explicit voice description.
• Refresh the examples quarterly.
• Cowork's drift toward marketing language is real; counter with explicit stop-words.

Tinkso's take#

Output quality is rarely a model problem; it is almost always a specification problem. We lose count of the times a client says "Cowork isn't smart enough" and the actual fix is in the prompt. Keep Prompt anatomy open in another tab when reviewing a poor output — the answer is usually two rungs up the ladder from where the operator was looking.