I've been running Anthropic's Claude Desktop (Cowork mode) across several project folders at my consulting firm. One of them… an enterprise transformation programme… has 462 files. Strategy docs, delivery plans, pricing models, brand guidelines, session notes, old drafts, the lot. Accumulated over months of real work.
It was going well until about three weeks in. Claude started stalling mid session. I'd ask it to update a delivery schedule and it would pull context from a pricing model we'd replaced three months earlier. It contradicted analysis it had done two days before. Sessions that used to complete cleanly were taking longer and producing worse output.
I couldn't work out what had changed. The files hadn't moved. The folder structure was the same. Then I watched what Claude was actually doing at the start of each session.
It was reading everything.
Every file. Every draft. Every archived version. Every superseded document. A 462 file folder and Claude was loading the lot before getting to my question. Half the context window gone on stuff that didn't matter.
The problem, once you name it
Claude has no concept of which files matter. A superseded draft from October and the canonical strategy document sitting next to it… they're the same to Claude. Both get loaded. Both compete for context window space. And when the window is finite, filling it with noise means the signal gets lost.
This is the same problem enterprise infrastructure solved decades ago with tiered storage. Frequently accessed data on fast disks. Less frequently accessed data further away. Archival data on tape until someone explicitly asks for it.
What I built
I applied near-line tiering to Claude's file access. Three components, all lightweight.
Global instructions. A block of text you paste into your Cowork settings. Tells Claude to always look for a manifest file first, read canonical docs before anything else, load other files only when the task touches their domain, and scope sub-agents to minimum context when decomposing tasks.
Folder manifests. A _MANIFEST.md template you drop into each working folder. It's a simple markdown table:
Tier 1 (Canonical) lists the source of truth docs… the files Claude must read before doing any work. Tier 2 (Domain) maps subfolders to domains so Claude only loads them when relevant. Tier 3 (Archival) flags old drafts, superseded versions, anything Claude should ignore unless explicitly asked.
Takes about five minutes to fill out per folder. Longer for big folders, but you only do it once.
Tiered Context Navigator skill. A Cowork skill that enforces the whole thing consistently. It runs a four step sequence every session: find and read the manifest, load Tier 1 docs, assess which Tier 2 files are relevant to the task, load only those. It also handles conflict resolution when two documents disagree (Tier 1 wins).
What changed
The difference was immediate. Claude stopped reading everything. It oriented from the manifest, identified the canonical docs, worked out which domain files it needed, and left the rest alone.
The 462 file folder that was causing stalls started completing cleanly. Claude stopped contradicting itself because it wasn't pulling from superseded documents any more. The quality of analysis went up because the context window was full of relevant, current, authoritative content instead of noise.
I've been running this for several weeks now. The big transformation folder is the heaviest test and it's holding up well.
Setup
- Clone or download the repo: github.com/hughtopian-gif/cowork-context-kit
- Copy the global instructions into your Cowork settings
- Drop the
_MANIFEST.mdtemplate into your working folders, fill in the tiers - Optionally install the skill for consistent enforcement
About 10 minutes for the basics.
Who it's for
Anyone using Cowork with folders bigger than about 10 files. The problem scales with folder size. If you're running Cowork on real project work and you've noticed quality dropping or sessions stalling, this is probably why.
MIT licensed. Use it, fork it, build on it.
Hugh Evans runs FromHereOn, a specialist enterprise architecture and transformation firm. He builds with AI and occasionally writes about it.
Top comments (0)