The reading room · Digna Legi
Audit your Agent files
/100
80–100: high value. 70–79: worth the time. Below 70: below the usual publication threshold.
Evidence-reviewed score based on available publisher text. Evidence is substantive, but some cited studies are summarized rather than fully shown, so research quality cannot be independently assessed from the excerpt.
Scores reflect one reader’s profile, not an objective quality rating. Best is a separate personal selection.
How scoring works →This brief · about 3 min with detail
Why read this
Agent instructions are perishable configuration: useful for hard-to-infer workflow constraints, harmful when they accumulate stale generic prose.
AI brief · Checked against source text
The main idea
The central claim is that AGENTS.md, CLAUDE.md, skills and memory should be treated as perishable configuration, not a growing knowledge base. They help most when they encode hard-to-infer workflow constraints, concrete task examples or recurring team rules; they can hurt when they accumulate stale preferences, generic advice, redundant prose and token cost. The author distinguishes useful steering from implementation judgment: context can shape how agents work without reliably making them better at subtle coding decisions.
Some background helpful. Basic familiarity with coding agents, local instruction files, skills, hooks, and context windows.
Go a little deeper
Prose is the wrong place for invariants
The author’s most useful distinction is between preferences that guide behavior and requirements that must hold. If a rule is non-negotiable, prose is fragile because the model may ignore, dilute or misprioritize it. The stronger pattern is to archive questionable instructions, then move durable constraints into tests, hooks or permissions where failure is observable and enforcement is mechanical.
Personalization is weaker than repeated specificity
The article pushes against the assumption that personal agent memory is automatically valuable. In the research described, one developer’s personalized skill did about as well as someone else’s, while a generic skill built from many developers was more useful overall. Personal rules earn permanence only when the same preference recurs across similar tasks, or when a skill contains concrete examples for a domain with real quirks.
Repository context helps execution more than taste
Context files appear most defensible when they tell the agent things source code alone may not reveal: slow commands, generated files, unusual conventions, architectural boundaries and project-specific safety rules. The author argues they are less able to teach subtle design judgment. That makes them operational guides, not substitutes for reading the code or encoding quality checks.
Audit quality, loading behavior and memory separately
The article is not just a call to delete aggressively. It separates three hygiene problems: installed skills can pile up locally, low-quality skills can have vague triggers or thin workflows, and auto-memory can retain stale preferences after project files look clean. It also notes that installed skill bodies do not all load into every prompt, so auditing should focus on discoverability, invocation and relevance, not only raw existence.
A case from the article
The slow test-suite guide
In one repository described by the author, a guide warned that the full test suite was very slow. Claude reacted by running more targeted tests, which saved time without making the underlying feature implementation more correct. The example illustrates the practical boundary of context files: they can improve workflow efficiency when they encode local operating facts, but they may not solve design or judgment failures.
How the case is made
The case is made through reported studies, tool behavior notes, developer discourse, and the author’s own configuration audit.
Where the idea has limits
The cited prompt-reduction result is explicitly limited: the author says it is not a target, uses unpublished internal evaluations, and applies to specific models in a specific harness.
A question to take away · from Digna Legi
Which rules in your agent setup would still be useful if the underlying model became substantially better tomorrow?
What the original adds
The source adds concrete audit targets: unused skills, MCP servers, plugins, slow hooks, stale memory, over-specified project files, skill descriptions, triggers, examples, security checks, and the distinction between `/doctor` and `claude doctor`.
About this brief
AI-written, then separately checked for source support, useful detail and clarity. The author’s claims and our editorial question are kept separate. The original remains the author’s work. How we select and summarise →
How was this brief?
Rate this summary, separately from the author’s article.
Optional. Saved in this browser; shared only if you allow analytics.
How was the original article?
Rate the author’s original after reading it.
Optional. Saved in this browser; shared only if you allow analytics.
Digna legi. Worth reading.