Topic
Technique for allocating an LLM's limited context window among system prompt, conversation, tool outputs, and memory, with truncation when budgets run out.
Crystals, a memory design written up on dev.to, deliver notes to an agent just before a matching tool call runs, from a hook firing about 300 times a day. Its most useful finding is a matched note that the token budget cuts before the model sees it while the logs still count a hit.
Reality
A dev.to harness guide puts numbers on two ways MCP breaks at scale, 8,000 to 15,000 tokens of tool schema on every prompt and a SQLite session lock duplicate stdio processes cannot share. Its fix moves stateful tools onto a supervised loopback daemon.
Reality
Browser-use's perception pass strips scripts and hidden nodes, then paints numbered badges on a screenshot so the model clicks by index instead of by XPath. The teardown puts action precision above 95% per action.
Reality
The Django tool caps each input document at 8,000 characters and the whole grounding context at 40,000. Its system prompt tells the model to name the missing input in one sentence when the data is not there.
Reality
Three postures exist for an agent instruction file. Only the third holds at commit speed. The team running it watched its own CLAUDE.md reach 548KB before a context-measurement pass cut it to 34KB.
Reality
The file published in the CL4R1T4S repository reads as tool schemas, search rules and safety policy flattened into one context. That makes it a measurement of one deployment, not a secret worth copying.
Reality
Nobody tells you to uninstall an agent skill, so the pile only grows. One author finally priced his: the cost tracks the size of the install, not anything he asked the model to do.
Reality
One developer found roughly 70 of his 112 installed skills reduced to bare names in the system prompt, and nothing in the tool had flagged it, because the listing budget drops descriptions without erroring.
Reality
A worked example from a dev.to post shows a fixed-size split severing "unless defective" from a refund rule. Retrieval still ranks the mutilated chunk first, and no component reports a fault.
Reality
Codex now has a documented precedence chain: one global file, then one file per directory from repo root down to your working directory, capped at 32 KiB. Determinism is the useful part.
Reality
One developer found seven dead MCP connections and could not say when they broke. His answer was measurement: a weekly snapshot that turns "feels slow lately" into a dated diff.
Reality
Every auto-generated Claude Code skill is injected into context on every conversation. One developer's weekly curator treats that like log rotation: stale at 30 days, archived at 90, nothing deleted.
Reality
A dev.to series on harness engineering splits persistence into in-task scratch and cross-session recall, and argues that building them as one system produces neither.
Reality