Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-02 · high
Eight builds and 38 model calls per skill later, the knowledge transfers but the persona and decision changes are much harder to prove.
Why now: The paper is new, and most existing advice on building agent skill files hasn't had to answer for its findings yet.
A new paper splits into three testable pieces the claim that an expert's writing teaches an AI agent knowledge, persona, and judgment all at once.
Built with mimeo, a tool the paper uses to test each piece on its own, the compiled file answered all 20 obscure, quotation-heavy questions. An agent working from memory alone never topped 10, and BM25 keyword search over the same source pages reached 15.
Building these files isn't free. Eight logged builds averaged 38 model calls each, and a quotation-accuracy check rejected 13.2% of the quotes the extraction step pulled from source material.
The persona and decision-change tests don't get the same clean result. That cuts against advice suggesting one writing sample buys knowledge, persona, and judgment together.
That leaves a practical test for anyone building a skill file this way. The model calls buy better recall on niche facts. They don't guarantee the agent will decide things the way the source person would.
Each link below shares sources, entities, or timing with this story.
Triple-stream retrieval (BM25 keyword, vector embeddings, knowledge-graph traversal) fused via Reciprocal Rank Fusion on the iii engine, with SQLite for state and an in-memory vector index, no external database. The economic claim: ~170K tokens/year (~$10) versus ~650K tokens...
mksglu/context-mode (19,893 stars, 1,430 forks, ELv2) is an MCP server attacking context burn four ways: sandboxing tool output so raw data never enters the window, tracking every file edit, git operation, task, and error in SQLite with FTS5/BM25 retrieval after compaction, en...
Everyone writing SKILL.md files has absorbed the same folklore. Keep the top file thin. Push detail into reference files. Let the agent walk the tree as needed. More layers, more context efficiency. A controlled study submitted July 20 tested that across InfiniteBench, three a...
MCP server that virtualizes agent context windows by sandboxing tool call outputs. Claims 98% context reduction (986KB to 62KB). SQLite+FTS5 with BM25 ranking. Every Playwright snapshot costs 56KB; twenty GitHub issues cost 59KB — this solves fundamental scaling. GitHub
1. Flip your multi-model pipeline to review-then-generate. Instead of using a reasoning model to plan before code generation, let the specialist generate freely and use reasoning tokens for review. Paper shows 90.2% pass@1 vs 87.2% for the planning pattern. Source 2. Audit you...
arXiv 2607.25936 shows models maintain an assigned role and reproduce its behaviors even when doing so produces wildly inefficient reasoning. RolePlay constructs adaptive personas that induce coherent but computationally expensive output, averaging 7.64x token amplification wi...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.