Topic
Techniques for handling finite LLM context limits, including full history replay, truncation, and compaction or summarization of past turns and tool outputs.
Claude Code screens MCP output for character length before it counts tokens, so a dev.to test measured a result at roughly twice the documented 25,000-token cap sitting in the conversation untouched. The docs describe the same guard in two units.
Reality
An engineer writing on dev.to spent two weeks on prompts and a larger model before concluding that his ten-step agent workflow broke at the fourth handoff because one supervisor was holding every worker's output.
Reality
A post arguing the Model Context Protocol was built for 2024-era models pushed a 165-point Hacker News thread into a fight about token cost. The post reports no benchmark, so the measurement is left to your own agent.
Reality
An adversarial four-model review panel had been catching real bugs for months. On 2026-08-10 its operator swapped file contents for repository paths to save context, and the findings kept arriving in the same confident format.
Reality
A dev.to postmortem of the central orchestrator agent argues for LangGraph's explicit edges, and the 60 percent latency win it cites only adds up once the manager's own turns come off the critical path.
Reality
A dev.to walkthrough of Claude Code compaction says the system prompt and the last 10 to 15 turns survive while the messages between them are summarized and deleted without an error. Constraints you need later belong in a file.
Reality
Claude Code's configuration-debugging page says nested memory files load on demand through the Read tool, so a report that the agent ignored the rules starts with /context and the question of whether the file was in the window at all.
Reality
Anthropic's memory documentation sets a 200-line target per instruction file and reserves real blocking for a hook. Everything written in markdown arrives as context the model weighs against your last message.
Reality
A dev.to walkthrough of MCP internals locates the integration saving in deployment coupling and leaves the residual risk with the host process that validates each tool call and raises the consent prompt.
Reality
An XDA Developers writer moved his desktop sticky-note prompt stash into Claude Code skill folders. Claude keeps only each skill's name and description in context and pulls in the body it judges to match.
Reality
An XDA Developers writer says his Claude allowance stretched four times further once he stopped letting the agent orient itself inside a 500-file repository. The post does not include token counts, so read it as one practitioner's log.
Reality
A quantised 27B on one 16GB GPU audited a home Splunk and Sysmon install with nothing leaving the host. Its first repair attempt ran the 128K context window dry. A halt rule in the system prompt got it under control.
Reality
A dev.to write-up cut a 194,492-character CLAUDE.md to 17,283 by changing what triggers each rule to load, sorting them by what breaks if one stays unloaded at the moment it matters. The harness that checks the result runs one headless Claude Code session per prompt.
Reality
A dev.to walkthrough prices one research agent off Anthropic's published rates for Claude Sonnet 5 and finds prompt caching takes the same twenty turns from $1.37 to $0.33. The workload behind those numbers is the author's own invention.
Reality
A dev.to test put the same 2,004-word rule set in four locations on Claude Code v2.1.263 and measured the first request of a fresh session. Only the path-scoped file kept its words out of the launch prompt.
Reality
A Hook fires on a lifecycle event whether the model agrees or not, while a Skill is pulled in by a per-turn judgement on its description. The exit code that enforces a Hook can also deadlock it.
Reality
A dev.to account of a four-hour agent run shows compaction preserving an abandoned fix in full and losing the operator's correction. The long-context benchmarks in the same post say a bigger window would not have saved it.
Reality
Across 60 coffee-shop listings, an agent with browser tools and a plain Playwright script pulled the same fields. The agent's per-place token count fell by about a third once it started reading aria-label attributes.
Reality
The same breakdown also points at subagent-heavy sessions and at the Playwright MCP server, and the shares overlap enough that one line on its own cannot size a saving. The overage that started his audit came from a different tool.
Reality
Claude Code loads skills in three stages, so an idle skill charges only its description to the context window. Thomas Tartrau says the same folder runs unchanged in Cursor, and his article is the only evidence for that.
Reality
Leadership · September 10, 2026 · 1 publisher
Leadership · September 6, 2026 · 1 publisher
Build · August 24, 2026 · 1 publisher
Build · August 22, 2026 · 1 publisher
Build · August 15, 2026 · 1 publisher
Build · August 15, 2026 · 1 publisher
Build · August 15, 2026 · 1 publisher