One of the challenges in getting good results from an LLM is maintaining enough space in the context window to get good output, while keep enough previous context that you’re not rebuilding the wheel every time you start a new work sessions. Good project hygiene is a big part of that. There are other tools to help make your context handoffs denser, and to offload some of the important stuff into “long term memory”, and we’ll cover more of those in a future post.
After six published posts, the Claudefather project had accumulated sixteen files totaling around 170K of content. Session handoffs, workflow guides, draft documents, editorial templates—each one added during the initial build sprint, each one making sense at the time.
But context windows aren’t infinite. Every file loaded into a Claude project is context that could be used for actual work. This session was about taking inventory, deciding what still earned its place, and consolidating the rest.
The Audit
I opened the session with clear expectations:
This is going to be a janitorial session. We'll evaluate our learning so far, our project structure and processes, and where we sit from a content generation perspective. Please rebuild your context from the attached project files.Claude provided a status overview—six posts published, empty queues, working infrastructure—and I followed up with three evaluation areas: what files could be deleted, whether our templates needed improvement, and what distribution prep (analytics, SEO) remained.
The file review was where things got interesting. Claude’s initial pass identified obvious candidates for deletion: draft files that had been published, obsolete workflow documents that had been superseded. But I pushed back:
I want to make sure we don't delete anything relevant that might make a good post later. Some of our sessions aren't included in the current publishing queue, but may present another opportunity for a meta post which takes learnings from multiple sessions.This changed the analysis. Claude went back through each handoff file, looking specifically at the “Raw Notes” sections for observations that weren’t captured elsewhere.
Before deleting completed work, extract the unique observations. Handoff files serve their purpose during draft generation, but the durable value lives in the insights—not the artifacts.
The Consolidation
The solution was migration, not just deletion. Claude created an updated SESSION_INDEX with a new “Session Observations” section, pulling the unique learnings from each handoff into a single reference. Issue analysis, behavioral patterns, teaching moments—all preserved in one place instead of scattered across files that would otherwise be deleted.
The same consolidation logic applied to workflow documentation. Two separate files—a workflow guide and an editorial workflow document—covered overlapping territory. Rather than maintain both, Claude merged them into a single source of truth.
Extract the unique observations and add them to the session index. Then I'll delete the other files that are no longer needed.The result: sixteen files became four. Project size dropped from 170K to around 35K—an 80% reduction.
The goal isn’t minimal files—it’s maximum signal per token of context. Everything in your project should earn its place by enabling better work.
Template Evolution
Part of the hygiene work involved updating templates based on actual usage. After six sessions, patterns had emerged that the original templates didn’t capture:
The issue taxonomy gained “Limitation” as a category—capability boundaries aren’t errors, but they’re worth documenting. The handoff template added explicit callout opportunity fields, prompting identification of TRAP/TECHNIQUE/BOUNDARY/INSIGHT moments during documentation rather than during draft generation. Image composition notes were added to prevent thumbnail monotony—tracking which layout types had been used recently.
Templates should evolve as you learn what information actually gets used. Review your templates after every few sessions and add fields for patterns you keep noting manually.
The Limitations Discovered
The session also revealed some infrastructure constraints. I’d enabled an llms.txt option in Yoast SEO—supposedly making content easier for LLMs to consume—and wanted Claude to verify the site’s metadata optimization.
Your URL caching makes it a little difficult to do onsite analysis, but here's a link I don't think you've accessed yet.Claude fetched the page content but couldn’t see the HTML head—web_fetch strips metadata. An attempt to use bash for raw HTML retrieval hit another wall: network egress is restricted to an allowlist of domains (npm, pypi, github, and similar developer resources).
`web_fetch` strips HTML metadata including Open Graph and Twitter Card tags. For SEO verification, you’ll need to use browser-based tools or provide Claude with the specific structured data files (like llms.txt) directly.
The workaround was simple—I provided the llms.txt URL directly, and Claude analyzed it alongside the sitemap. This surfaced a few gaps: a missing meta description on one post, “An Introduction” not appearing in the llms.txt output, a category potentially not indexed. Minor issues, but ones that wouldn’t have been caught without the review.
What Stayed
The final project structure:
CALLOUT_SYSTEM.md — Defines the four callout types and usage guidelines
SESSION_HANDOFF_TEMPLATE.md — Updated template for documenting sessions
SESSION_INDEX.md — Master registry with extracted observations
WORKFLOW.md — Consolidated workflow guide
Four files. Everything needed to continue the work, nothing that had outlived its usefulness.
The broader lesson here isn’t about file counts or context window optimization—it’s about treating your Claude project as a living workspace that needs occasional maintenance. The files that made sense during a build sprint may not be the files you need for ongoing work. Published content doesn’t need to stay in project files. Observations outlast artifacts.
Take the time to audit periodically. Your future sessions will thank you.
