The Workflow
I read Ayush Poddar’s piece on Claude’s memory one morning last week and got stuck on a single line: everything in your persistent context gets re-read at the start of every session. Not summarized. Re-read, in full, every time.
So I went to check what mine actually cost.
Three files load before I type a word: the CLAUDE.md for my Second Brain vault, the CLAUDE.md for the SoloBuilt site, and MEMORY.md, my standing-rules file. I’d been adding to all three for weeks. I had never once measured what they weighed.
Step 1. Run the audit: I asked Claude to read all three files and estimate the combined token load. The number came back around 12,000 tokens. That is roughly three to four times what these files are supposed to be. Before any real work, Claude was reading the equivalent of a short paper just to remember who I am.
Step 2. Find the dead weight: Most of the bloat was reference material I’d pasted in “so Claude would have it.” Full command lists. Long explanations. Examples of examples. None of it needed to live in always-on context. It needed to live in a doc that gets read only when it’s relevant.
Step 3. Cut to pointers: I trimmed the vault CLAUDE.md from 174 lines to 102, and the SoloBuilt one from 248 to 196. The rule I used: if it’s reference material, replace the whole block with one line saying where to find it. Keep the instruction, drop the encyclopedia.
Step 4. Rewrite the rules file: MEMORY.md got a full pass. That’s where the real surprises showed up, and they had nothing to do with token count.
When I was done, the combined load was about 3,000 tokens. A 75% cut. Same context, a quarter of the weight.
What Broke
Nothing broke because of the audit. The audit showed me what was already broken.
When I opened MEMORY.md, my standing-rules file, I found three memory files sitting in the folder that were never indexed anywhere. Claude couldn’t see them. They’d been written, saved, and quietly ignored for weeks. There was also a pointer telling Claude to read a file called session-habits.md every session. That file never existed. It had been pointing at nothing the whole time.
Then the worse one. Seven rules I thought Claude was following were not written down anywhere. The n8n sandbox limits. The rule about where my LinkedIn links go. Things I’d “told Claude” in some session months ago and assumed had stuck. They hadn’t. Every new session started fresh without them.
The fix was the same for all three: put it in the file that actually loads. I added the seven rules to MEMORY.md, deleted the dead pointer, and indexed the orphaned files so they’d load with everything else.
Here’s the part I didn’t expect. I started this to make Claude cheaper to run. What I got was a list of business processes I believed were on autopilot and weren’t. The audit didn’t just trim the context. It switched things back on that I thought had been running all along.
Time Ledger
The Prompt File
I didn’t write the original audit prompt. That one came from Ayush Poddar’s article, and credit where it’s due. What I did was turn his one-time audit into a check I can run on any context stack, on any machine, the next time a CLAUDE.md starts feeling heavy. Here’s my version.
You are a context auditor. You read every file that loads at the start of a Claude session and report what is bloated, broken, or missing before changing anything.
Inputs:
- Always-on files: [LIST_OF_SESSION_START_FILES]
- Target combined token budget: [TOKEN_CEILING]
- Rules I believe are in force: [ASSUMED_STANDING_RULES]
**Step 1. Measure the load:** Read each file. Estimate the token count per file and the combined total. Flag the total against [TOKEN_CEILING] and name the single heaviest file.
**Step 2. Find dead weight:** Identify reference material that does not need to load every session, like command lists, long examples, anything needed only occasionally. Mark each block for replacement with a one-line pointer to where it lives.
**Step 3. Test every pointer:** For each file referenced inside these files, confirm the referenced file actually exists. List any pointer that leads to a file that is not there.
**Step 4. Find orphans:** List any memory or rules file in the folder that is NOT indexed in the always-on files. These load for no one and are seen by nothing.
**Step 5. Catch unwritten rules:** For each rule in [ASSUMED_STANDING_RULES], confirm it is actually written in a file that loads. Flag any rule I assume is active but that exists nowhere in the context.
Constraints:
- Do not edit or delete anything yet. Report first, as a numbered list grouped by: bloat, broken pointers, orphaned files, unwritten rules.
- When trimming, keep the instruction and cut the reference. Replace each removed block with one line naming where the detail now lives.
- Confirm each change with me before you write it.Manager’s View
Ask whoever owns your team’s AI setup: of the rules you believe every AI-assisted task is following, how many are actually written into the instructions the tool reads, and how many just live in someone’s head?
Most leaders assume the guardrails they set in a meeting are in force. The brand rules, the compliance language, the “always do it this way” from six months ago. My seven rules felt exactly that locked in, and they were running on nothing. The trap with AI specifically is that the output looks compliant whether or not the rule is there. A confident answer and a constrained answer read the same on the page.
So don’t ask your team “are we using AI responsibly?” You’ll get a yes. Ask to see the actual file the tool loads, the system prompt or the custom instructions, and check each rule you care about against it, one line at a time. If the rule isn’t written in the file, it isn’t a rule. It’s a hope.
Field Notes
- The pointer that broke led to a file called session-habits.md. Claude had been told to read it at the start of every session. It was never created. The instruction had been firing into nothing for weeks, and no error ever surfaced, because a missing file isn’t a crash. It’s just silence.
- One of the fixes was a single line in a config file on the Mac Studio. My MacBook Air needs the exact same line typed in by hand. The change isn’t really done until I’m sitting at the other machine, which is the tax nobody mentions when you run the same setup across three computers.
- Poddar sells a $29 bundle built around a feature called a PreCompact hook. I had it in the cart. The hook doesn’t fit a workflow that runs across three machines, and the one change that actually moved the needle was a free setting I already had switched on. The audit’s first save was talking me out of the upsell.