Conciseness

The body is lean, dense prose that assumes Claude's competence: 'Core memory is the agent's prompt, refined over its lifetime. Treat each change as a revision of that prompt, not an append to a log.' It never explains concepts Claude already knows, and every line ('Never store secrets. Reference them by name, such as $SECRET_NAME.') earns its place.

Actionability

Guidance is concrete and executable for an instruction-only skill: an exact compiled-prompt template, a copy-paste frontmatter example, named config fields ('maxCoreMemoryCharacters', 'readOnlyFiles'), and a labeled contrastive example. It falls short of 5 because the commit workflow that activates edits is only implied ('Edits take effect only after they are committed') with no concrete steps.

Workflow Clarity

The editing process is coherent with embedded checkpoints — 'Check that a change does not contradict an existing line; resolve the conflict', the pre-commit hook enforcing limits, 'After adding a directory or file, check that a natural future request would reach it' — and 'Shrink before you grow' gives a numbered 3-step sequence. It is a 4 rather than 5 because the main edit workflow (read config → edit → verify limits → commit) is assembled from scattered rules instead of one explicit sequence.

Progressive Disclosure

The body (~100 lines) has no bundle files and uses clear, well-ordered sections (Prompt format, Writing prompts, Teach with examples, deferred-memory design, Shrink before you grow) with no nested references. It stays at 4 because some self-contained detail — the full prompt-compilation spec and the .memfs.config.json limit semantics — could live in a one-level-deep reference file if the bundle grew.

Specificity

The description lists several concrete actions — 'modify the agent's core behavior, embed critical knowledge (needed on every turn), or instruct the agent on how to use deferred memory' — each a specific capability rather than generic filler. It falls short of the 5 anchor because coverage of the domain's surface (e.g., file layout, persona/human blocks, size limits) is incomplete.

Completeness

Both 'what' ('Maintains an agent's core memory, the editable part of its system prompt') and 'when' ('Use to modify... embed... or instruct...') are explicit. It sits at 4 rather than 5 because the when-clause is phrased as purposes rather than concrete user-side trigger phrases (e.g., 'use when the user asks the agent to remember something across sessions').

Trigger Term Quality

Relevant keywords like 'core memory', 'system prompt', 'core behavior', and 'deferred memory' are present and natural for someone managing agent memory, but common variations users would say — 'edit memory', 'remember this permanently', 'MEMORY.md', 'update the persona' — are missing. This matches the 3 anchor ('some relevant keywords but missing common variations or synonyms') more closely than the 4 anchor's 'good keyword coverage'.

Distinctiveness Conflict Risk

'Core memory' and 'the editable part of its system prompt' carve out a clear niche distinct from generic note-taking or file skills, with only minor overlap risk against a broader 'memory management' skill. Not a 5 because 'maintains memory' alone could still collide with a sibling deferred-memory skill without the qualifier.