Quote It Back

---

I’d spent an evening on my CLAUDE.md. That’s the file Claude reads at the start of every session: who I am, how I like to work, what I’m trying to get better at. I’d added my role, a pointer to our design system, working modes and coaching rules. I was rather pleased with it.

So I opened a new session and, before anything else, asked it to quote the file back to me word for word and tell me where it came from.

It quoted the old version. Four short paragraphs about honesty and curiosity, sitting in a cached copy I’d forgotten existed. Every improvement I’d made was in a file the agent wasn’t reading.

Here’s the thing: if I hadn’t asked, I’d never have known. The answers would have been fine, just a bit generic, and I’d have blamed the model.

---

The prompt

This is the whole method. Paste it at the start of a session:

Before anything else: quote back the CLAUDE.md instructions you received at the start of this session, word for word, and tell me which file(s) they came from. If you received none, say so and stop.
Then critique that file as a colleague, not a cheerleader:
1. Which lines will actually change how you respond, and which are too vague to act on? Rewrite the weak ones.
2. Any contradictions or tensions between instructions?
3. What's missing for someone in my role?
4. Is it the right length, or is anything diluting the rest?
Finish with the one change that would make the biggest difference.

Each part has a job.

Quote it back checks that the file was actually loaded. You can’t critique what you can’t see, and you can’t trust behaviour you haven’t confirmed.

Which lines will actually change how you respond is the question that matters most. A line the model would follow anyway is just decoration.

Tensions catches rules that pull in opposite directions. The model resolves them silently, and not always the way you would.

What’s missing for my role turns the critique from “is this well written” into “is this useful for the work I do”.

One change stops the critique becoming a to-do list I’ll never get through.

---

Round one: all personality, no context

Even the stale file was worth critiquing. The verdict was blunt: almost everything in it was something the model does by default. “Be honest and respectful” changed nothing. “Find the right time to coach me” was impossible to act on, because the right time according to whom? “Become a professional augmented by AI” was an aspiration, not an instruction.

The useful line was the one that set a direction against the model’s natural pull: I’d rather you tell me I’m wrong than try to prove I’m right. It flagged the line that told it to change its behaviour and called most of the rest filler.

The biggest gap was that the file described how I wanted to be treated and said nothing about what I do. I’d written a personality brief for an agent working without context. Tone the model can work out for itself. Context it can’t.

---

Rounds two and three: the small things that break

Then I pasted the version I’d meant it to read: my role, where the design system lives, house spelling, “answer first, reasoning after”, a couple of working modes, a coaching section. It’s much better.

The critique still found things I’d missed, and they were embarrassing in a useful way.

A broken cross-reference. A coaching rule pointed to the wrong goal number. A rule pointing at the wrong goal works against the goal you actually care about.

The starred goal was the unfinished one. My current focus still had template brackets in it: [audience] through [formats]. Every other goal was concrete. The one I’d marked as most important was the one I hadn’t decided.

Paths it couldn’t reach. The file pointed to a repository and a local folder that weren’t available in that session, and it didn’t say what to do when they weren’t. Without that, an agent guesses. The fix was one line: if you can’t access it, say so before you start.

A rule in the wrong place. My “Don’ts” section had a single item: a visual rule that belonged in the design system, not in my personal instructions. If I have to repeat it there, the system’s guidance isn’t reaching agents, and that’s a distribution problem, not a CLAUDE.md problem.

Too many add-ons. Coach’s note, a question, a “worth sharing” line, a suggested addition to the file. Each was reasonable, but together they could stack four extras onto a reply I’d also asked to keep short.

And honestly? I’d written that file to sound like someone who welcomes challenge. The critique showed the difference between sounding like it and setting the file up to deliver it.

The cross-reference bug survived my next edit too. I was applying the critique by hand, one line at a time, and fixes kept getting lost along the way.

---

Round four: stop editing, start reviewing

By the fourth version I’d added a long section on connecting to our ticketing and code tools, and the file had passed the 60-line cap I’d written into it myself. The critique pointed out that this setup only works in one environment but was loaded into every session, which is what skills are for. A skill loads only when the task needs it.

So I stopped hand-editing. I asked for the full corrected file plus the tooling as a separate skill, and reviewed the diff instead. It went from 80 lines to 52, and every line left is about how I work.

---

What I’m taking from it

Check what’s loaded, not what you wrote. Your instructions are only as good as the copy the agent actually reads. Quote-back is a ten-second test. Run it whenever you change the file.

Separate behaviour from context. Behaviour instructions mostly adjust things the model already does. Context (your role, the systems you work in, where the source of truth lives) changes what goes into every answer. If you only have time for one, write the context.

Every rule needs a trigger and an action. “Coach me” isn’t a rule. “When I make a choice you’d make differently, say so in one line, at the end of the task” is.

Goals need a signal you can see along the way. An outcome that arrives once, late, can’t steer a single reply. Pair it with something you can notice week to week.

Keep it about you. Setup for a particular environment belongs in skills. Domain rules belong in the system that owns them. A misplaced rule is a bug report for wherever it should have lived.

Let the agent write, you review. Ask for the full corrected file and read the diff. You stay the editor, and nothing gets lost in transcription.

Let the file maintain itself. The first line of my CLAUDE.md now asks the agent to propose an addition whenever I correct it, and to propose cuts when the file gets too long.

---

Where it ended

The most useful moment wasn’t a rewrite. The critique didn’t fill in my starred goal. It asked what was stopping me.

That’s the part I didn’t expect. I’d asked the file to make the agent challenge me, and the most useful challenge was about the file itself.

Make the instructions earn their place. Then ask the agent to read them back.