Context Sharpening: Ask What Your Agent Knows

Your agent knows things about your business that are written down nowhere.

Why a decision went the way it did. Which approach you tried in March that quietly failed. The constraint nobody documented because you found it by running into it. The correction you gave it on Tuesday that it has been honoring ever since. None of that is in a file. It lives in the conversation, and you cannot see any of it from the outside.

It is also carrying junk. Assumptions it made early and never confirmed. Instructions that stopped applying two projects ago. A rule it follows carefully that stopped mattering in June. You cannot see that either.

There is a simple way to find out both: ask it.

What context sharpening is

It's a prompt that makes a session stop working and look at itself. Three moves, in order.

It takes stock. What does it know now that it didn't at the start? What did it learn the hard way? Where is it still fuzzy — running on something it assumed and never checked? What is it carrying that stopped being relevant?

Then it asks you questions. Three to five, and it waits for your answers before going further.

Then it proposes. What's worth writing down so it outlives the conversation. What's worth fixing in the files it reads at startup. What it thinks it can stop carrying. And finally, given a clearer view of its own ground, the five things it would do next in your area if the choice were its.

Nothing gets built while it runs. Every output is a proposal you approve or bin.

The questions are the whole mechanism

That middle move looks like a courtesy. It's the part that makes this work at all, for two reasons.

The first is that a model asked to summarize its own work will confidently tell you things that never happened. Not maliciously — it's filling a gap the way a person misremembers a meeting. Ordinarily that costs you nothing, because you read it and move on. Here it's dangerous: you're about to write these "facts" into the files every future session reads. A made-up rule in a startup file is worse than no file at all, because every session after it inherits the mistake as truth and none of them re-checks it. Being made to stop and ask is what forces the uncertainty up instead of letting it get smoothed into confident prose.

The second is simpler. Your agent has no idea which of the hundred things it's holding matter to what you're doing next month. Only you know that. Without the questions it guesses, and it guesses generic — you get "document your decisions" instead of "the reason we do it this way is X, and it will bite you again the next time Y happens."

Answer the questions properly. That part is the actual work, and it takes a few minutes.

The prompt

Fill in the one bracket and paste it into a session that has real history behind it.

CONTEXT SHARPENING

Pause the work and take stock of yourself.

You're going to look at what you're actually carrying in this conversation, sort
out what's worth keeping, and say where you'd point your attention next. Nothing
gets built while we do it.

What you own: [THE DEPARTMENT, REPO AREA OR PROJECT THIS SESSION IS RESPONSIBLE
FOR — a line or two]

Three moves, in order. Stop at the end of the first one and ask me your questions
before going on.

Everything you produce here is a proposal for me — nothing changes on disk during
this. And if credentials, keys or personal data came up in our conversation, refer
to them by name; don't reproduce the values anywhere in this exercise.

MOVE 1 — TAKE INVENTORY

Look back over this conversation and tell me what you're actually holding.

- What do you know about this area now that you didn't know at the start?
- What did you work out the hard way — an approach that failed, a constraint you
  ran into, a correction I gave you and what it was really about?
- Where are you still fuzzy? What are you operating on that you assumed and never
  confirmed?
- What are you carrying that stopped being relevant a while ago?

Be generous, and be honest about the fuzzy parts — that half is usually more
useful than the confident half.

Then ask me three to five questions: the ones where my answer would actually
change how you work in this area. Stop there and wait for my answers.

MOVE 2 — PROPOSE A CLEANUP

With my answers, sort what you're holding into three piles and show me each:

WORTH WRITING DOWN — things that should outlive this conversation. Give each as a
short flat statement someone could act on without ever seeing it. Mark anything
you're not fully certain actually happened, so I can confirm it rather than
inherit it as fact.

WORTH FIXING — anything in the files you load at startup that is now wrong,
contradicts something else, or has stopped earning its keep. Every line in those
files is re-read on every turn of every future session, so a paragraph that no
longer teaches anything is a standing charge. Quote what's there, give the exact
replacement, and say what changed to make it necessary. If you can't work out why
a rule is there, keep it and flag it instead — most of them were written straight
after something went wrong, and the reason usually isn't in the file.

WORTH DROPPING — what you're carrying that looks like it no longer serves the
work, and why you think so. This one is you pointing, not deciding: keep following
all of it exactly as you have been until I tell you otherwise.

MOVE 3 — POINT IT FORWARD

Now that you can see your own ground clearly: if it were up to you, what would
you do next in this area?

Five ideas, specific to this system and no other — not general best practice. For
each one: what it is, why now, the first concrete step, and what it's worth if it
works. Rank them, and tell me which one you'd start and why. Give me the list and stop
there — I'll pick.

It can't damage the session it runs in

Worth knowing before you point this at something you care about, because a self-review is exactly the kind of thing that can quietly go wrong.

It changes no files, applies no edits, and starts no work. Everything it produces is a proposal.

The real hazard in an exercise like this isn't a bad suggestion — it's a session deciding on its own that some instruction has gone stale and quietly dropping it. You would never trace the degradation back here. So the prompt has it point at those rather than act on them, and keep following everything exactly as before until you say otherwise.

Two smaller guards do the rest. It flags anything it isn't certain really happened, so a guess doesn't get written down as fact. And it names credentials rather than reproducing them, so nothing sensitive gets copied out of a conversation into a document.

The worst case is that it hands you a list you disagree with.

When to run it, and on what

Run it on sessions carrying real history — a long build, a project thread, whatever holds an area of your business. A fresh session has nothing to take stock of and will hand you a page of generic advice.

The moments it pays best: when a session starts feeling muddled or repeating itself, before a long break, at the end of a work cycle, or before you hand an area to somebody else.

There's one thing worth knowing about timing. The sharpening only reaches what's still in the session's window. A conversation that has been running so long it has compacted repeatedly has already forgotten some of what you're trying to collect. Long but not ancient is the sweet spot, so if you have one enormous session you have been feeding for weeks, do that one first.

And if you're on a plan with a weekly limit that resets, this is the best possible use of capacity that would otherwise expire. Unused capacity doesn't roll over. Spending the last of it on the files every future session reads turns something that was about to disappear into a permanent improvement. /usage in Claude Code shows your reset, and so does Settings then Usage in the app.

What it actually does

It does not make the model smarter. A cheaper model next week is still a cheaper model.

What changes is the quality of what it reads before it starts, and how clearly it understands its own job. That's a briefing effect, not an intelligence one — and it's worth a great deal, because the difference between an agent that starts informed and one that starts blank is most of the difference in what you get out of it.

Ask your agent what it knows. You will be surprised twice: once by how much of it was never written down, and once by how much of what it's carrying stopped mattering a long time ago.