Most writing about AI tools is either sales material or anxiety. This is what several hours a day of actual use looks like, including where I do not let it near the work.
I use Claude for several hours most working days, and I want to describe that honestly rather than enthusiastically, because most writing about AI tools is either sales material or anxiety.
What it genuinely does for me
Four things, in order of how much time they save.
Turning rough notes into a first draft. I know what needs saying; assembling it into a clean paragraph is the slow part. I paste my notes, state who it is for and how long it should be, and edit what comes back. The draft is never what I send, and it is a far better starting point than a blank page.
Thinking a decision through out loud. Not asking for the answer. Describing a situation properly and being asked questions about it. Half the value is that writing the situation down clearly usually reveals what I actually think, which is an old technique that happens to work well here.
Pressure-testing something before it goes out. Asking what a reader might misread, what is missing, what a sceptical person would push back on. This is where it earns its place most reliably, because it is genuinely difficult to see the gaps in your own writing.
Explaining something I need to understand quickly. Enough grasp of an unfamiliar area to ask sensible questions of the person who actually knows.
Every one of those is a task where I already have the substance and need help with the shape. That is the pattern.
How I actually prompt
There is no technique here worth mystifying. Three parts, and the first is the one people skip.
1. The context I would give a new colleague
Who this is for, what they already know, what happened before, and what I am worried about. This is the part that determines whether the output is useful or generic, and it takes longer to write than the request itself.
2. The raw material, however untidy
My actual notes, the actual thread, the actual numbers. Untidy input with real substance beats a tidy prompt with none. I paste far more context than feels necessary.
3. The specific ask and the constraint
What I want back, roughly how long, and what it must not do. The constraint matters as much as the ask: no bullet points, do not invent figures, keep my phrasing where it works.
What I do not do is write elaborate persona instructions or hunt for magic phrasings. In my experience the returns on that are small compared with simply supplying better context.
Where I stop, and why
| What it is good for | Where it stops | What it costs me |
|---|---|---|
| Drafting from my own material | Nothing goes out unedited. The voice drifts within a few messages if I am not paying attention. | Minutes of editing, always. |
| Summarising long threads | It will summarise confidently past the point where it should say the thread is ambiguous. | A skim of the original for anything that matters. |
| Research and facts | I do not use it as a source. Anything factual I verify at the primary source before it appears in my work. | Real time, and it is time well spent. |
| Anything client-facing | It never sends. It drafts, and a person decides. | One approval step, deliberately kept. |
The strongest example of that last row is on this site. Every statistic in every article here was traced to a primary source, and several widely repeated figures were rejected precisely because they could not be. That checking was mine. The tool made the writing faster; it did not make the sourcing trustworthy.
What it has not changed
It has not made me a better judge of what matters, and it does not know which of two conflicting priorities wins in a business it cannot see. It does not know that a particular client is sensitive about tone, or that an internal deadline is aspirational.
Those are the things my actual job consists of, and they remain exactly as hard as they were.
How it breaks
Voice drift. Accept enough drafts unedited and everything starts sounding the same. The fix is editing every time, and the failure is invisible from the inside.
Confident wrongness on specifics. Dates, figures, names of things. It is fluent about all of them and should be trusted on none.
Using it to avoid thinking. The genuine risk. If I ask for a decision rather than for help articulating one, I get something plausible and I have skipped the part of the job that is actually mine.
How to tell whether it is working for you
One test: would you be comfortable showing someone the input you gave it? If the input was thin, the output is generic and you will spend longer fixing it than writing it yourself. If the input was substantial, the output is usually a real head start.