memory-search-best-practices

记忆系统多刀检索 + recall coverage 策略（8 类题型 recipe）。 Use when: 任务是 "哪些地方提过 X" / "X 的来源 / source map" / "有没有提过 Y / absence check" / "上次到现在变了什么 / delta" / 冷启动 onboard 复杂主题 / 任何召回任务搜了一刀觉得不够。 Not for: 只是选哪个入口走第一刀（用 memory-navigation）/ 已知精确 anchor 单 Read（直接 Read）/ 代码符号查（Grep/LSP）/ 新功能开发（不是 recall 任务）。 Output: 多 query 多 scope 召回 union 结果 + coverage matrix（item/source/谁提到/直接 vs 间接）+ "何时停下来"判据。 GOTCHA: 和 memory-navigation 互补不重叠 — memory-navigation 决定**第一刀走哪个工具**（search vs graph vs list_recent），本 skill 决定**要不要补刀 + 题型对应几刀几路 + 何时停**。Ragdoll家族必加载：opus 系治"我能猜出来 / 碎片够了"停太早病，fable 系治"再确认一轮"停太晚病（双向校准见正文）。

Quality

72%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./cat-cafe-skills/memory-search-best-practices/SKILL.md

Quality

Content

62%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A thorough, operator-validated recall skill with concrete per-query-type recipes and stop criteria; its weaknesses are repetition of the same stop/calibration guidance across four sections and a monolithic structure with no supporting reference files despite referencing several specs. Time-sensitive dated anecdotes also slightly undercut conciseness.

Suggestions

Consolidate the stop criteria: keep one authoritative '何时停下来判据' table and have the Ragdoll and Common Mistakes sections link to it instead of restating the ≥3-route rule three more times.

Move the dated dogfood anecdotes (2026-05-17, 2026-07-08, 2026-07-15) into a short 'changelog / provenance' section or a reference file so the main body stays lean and date-stable.

Pull the F256 'Related directions' provenance-trust material into a one-level-deep reference (e.g. references/f256-related-directions.md) and summarize the three trust buckets inline, improving progressive disclosure.

Dimension	Reasoning	Score
Conciseness	Dense and mostly efficient, but the ~120-line body re-iterates the same ≥3-route / stop-early-vs-late material across the recipe table, the Ragdoll calibration section, the stop-criteria table, and Common Mistakes, with dated dogfood anecdotes adding padding.	3 / 5
Actionability	The 8-row recipe table and the 5-step AUDHD recipe give concrete, near-executable calls (search_evidence(..., hybrid, scope=docs, limit=10), graph_resolve with relations filter), with only minor gaps around exact arguments for every step.	4 / 5
Workflow Clarity	Workflows are explicitly sequenced (recipe table per question type + a 5-step worked example) with a per-type stop-criteria table as validation checkpoints, though the 'when to stop' rules are spread across three sections rather than one authoritative checklist.	4 / 5
Progressive Disclosure	No bundle files exist (references/scripts/assets absent) and the body is a single inlined monolith; the '相关' section links specs and sibling skills but detailed material (Ragdoll calibration, provenance trust buckets) is inlined rather than split into one-level-deep references.	3 / 5
	Total	14 / 20 Passed

Description

82%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A strong, dense description that explicitly states capabilities, a 'Use when' trigger set, exclusions, output format, and a conflict-avoidance GOTCHA relative to its sibling skill. Its main weakness is reliance on project-internal jargon that may be less natural as a trigger for general users.

Suggestions

Add a few plain-language synonyms alongside the in-domain terms (e.g. 'where was X mentioned / who said it / has anyone raised Y') so the triggers read as phrases a user might naturally say.

Consider trimming the GOTCHA line slightly; the core distinction ('第一刀 vs 补刀') is the load-bearing part and the Ragdoll family note can live in the body.

Dimension	Reasoning	Score
Specificity	Names concrete actions like '多 query 多 scope 召回 union 结果 + coverage matrix' and '题型对应几刀几路' rather than vague verbs, with only minor coverage gaps around the exact search primitives.	4 / 5
Completeness	It explicitly answers both 'what' (多刀检索 + recall coverage 策略, 8 类题型 recipe, coverage matrix output) and 'when' (a concrete 'Use when' clause plus a 'Not for' exclusion list), with concrete trigger phrases.	5 / 5
Trigger Term Quality	The 'Use when' clause lists natural operator phrasings ('哪些地方提过 X', 'source map', 'absence check', 'delta') and a triggers list adds more, though it leans on in-domain jargon over universally natural user terms.	4 / 5
Distinctiveness Conflict Risk	A dedicated GOTCHA plus 'Not for' clause cleanly separates it from the adjacent memory-navigation skill, giving it a clear niche with minimal conflict risk.	5 / 5
	Total	18 / 20 Passed

Validation

93%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: zts212653/clowder-ai
Path: cat-cafe-skills/memory-search-best-practices/SKILL.md
Commit: 090626a

Reviewed: ~23 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.