Prompt anatomy — what good Claude Cowork prompts share — The Cowork Bible — Tinkso

TL;DR. Every good Cowork prompt names five things: inputs, output, transformation, constraints, plan request. Internalise that structure and you can stop copying from libraries — you can write your own. The 30-minute workshop on this template lifts a team's output quality more than a quarter of trial and error.

The five pieces of a working prompt#

Every good Cowork prompt names:

The inputs — exact folder, exact filenames or patterns.
The output — exact folder, exact format, naming convention.
The transformation — what to do, in plain language.
The constraints — voice, length, quality bar, edge cases.
The plan request — "plan before acting" for any non-trivial work.

Treat each one as a question the prompt answers, not a thing to mechanically include.

The minimum viable prompt#

Four lines, all five pieces present:

Read /inbox/notes-this-week.md.
Summarize for an executive in &lt;200 words.
Tone: matter-of-fact, no marketing language.
Save to /output/exec-summary-[date].docx.

This is enough for a one-off task. Read what, do what, in what voice, save where.

The robust prompt template#

For repeated work, layer the constraints:

Goal: produce a one-page exec summary of the week's work.

Inputs:
- /inbox/notes-this-week.md
- /inbox/last-week-summary.docx (for tone reference)

Output:
- /output/exec-summary-[YYYY-MM-DD].docx
- Length: 200 words, ±10%
- Voice: matter-of-fact, British English, no em-dashes

Method:
1. Plan the structure first; show me before drafting.
2. Use last week's summary as a tone anchor only — do not copy structure.
3. Surface anything you cannot verify from the notes as "[CONFIRM]".

Quality bar:
- Every claim traces to a note in /inbox/notes-this-week.md
- No marketing adjectives ("transformative," "exciting," "innovative")
- The summary works for a CFO who has 60 seconds.

The robust template is what becomes a saved prompt and eventually a skill.

Common failure modes and the fix#

Failure	The prompt fix
Too verbose	Cap the word count
Lost the voice	Reference 3 examples, not just one
Made stuff up	Require source citations or `[CONFIRM]` tags
Did the wrong thing	Require a plan step before acting
Inconsistent across runs	Save the prompt in `CLAUDE.md`; don't retype

Most failures fall into one of these five buckets. The fix is in the prompt, not in the model.

The plan-mode habit#

For any operation touching more than five files, ask for a plan.
Read the plan. Push back in plain English.
Approve when the plan matches your intent.

This habit is worth 100× what it costs in time. It is the difference between "Cowork is unreliable" (without it) and "Cowork is part of the workflow" (with it).

From prompt to skill#

A prompt becomes a saved prompt when you use it twice.
A saved prompt becomes a skill when three or more operators use it regularly.
Save frequently used prompts in your workspace's CLAUDE.md under a ## Saved prompts section. See CLAUDE.md and memory.

The graduation matters because it changes who maintains the prompt: an individual operator owns a saved prompt; a function owns a skill.

Tinkso's take#

Prompt anatomy is the cheapest training we deliver. A 30-minute workshop on this template lifts a team's output quality more than a quarter of trial and error. The prompts in this library are illustrations; the structure is the point. Once an operator can score their own prompts against the five pieces, they no longer need the library.

Try this#

Take a prompt you have already used. Score it against the five pieces. The lowest-scored piece is the upgrade target. Most operators discover their constraints section is the weakest — fix that and the next ten runs will land cleaner.