菜单
Use to store, recall, query, or audit durable project knowledge — atomic facts and compound documents alike — plus a working notepad that survives conversation compaction. One skill, one storage convention, file-backed under `.vibekit/`. Cross-runtime portable.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use after verify-gate returns `ready`, as an optional security pass over a vibe run's diff before finish-branch. Runs a tiered threat-model review — universal code-security checks on every diff, plus AI-artifact checks when the diff builds a skill, agent, prompt, or MCP server — quotes verbatim file:line evidence, scores findings, and hard-gates the handoff on CRITICAL/HIGH with a written-waiver escape. Peer to review-pack; both optional, either may run before finish-branch.
Use to diagnose vibekit installation health — skill file integrity, runtime registration parity, `.vibekit/` directory health, `docs/` subdir presence, authoring-contract alignment. Reports verdict only by default; `--fix` auto-repairs safe items. Cross-runtime portable.
Use after verify-gate returns `ready` and before finish-branch or any outward-facing action. Performs a self-review pass (Reflexion-style) — the skill critiques its own completed work against the spec and the plan, surfaces issues in order of severity, and presents the diff to the user for explicit sign-off. No outward-facing action runs until the user approves.
Use when starting any conversation — establishes vibekit's auto-trigger discipline so pipeline skills (brainstorm, brief, plan, exec, verify, review, finish, memory) actually fire at their trigger points instead of being silently skipped.
Use when the user says "vibe" or gives a short intent like "add X", "build Y", "fix Z". Runs the full disciplined pipeline end-to-end — brainstorm → plan → isolated execution → verification — with minimal user interaction outside the brainstorm phase. The user only has to answer design questions; every other step is autonomous, guardrail-protected, and token-efficient. If the user does not say "vibe" explicitly, do not assume this skill applies.
You MUST use this before any creative or implementation work — creating features, building components, adding functionality, or modifying behavior. Runs a disciplined Socratic brainstorming loop (hard gate, one-question-at-a-time, 2-3 approaches, design doc, user approval) while keeping token overhead minimal. Questions, user answers, approaches, design, and the written spec are never compressed.
| name | memory-dual |
| description | Use to store, recall, query, or audit durable project knowledge — atomic facts and compound documents alike — plus a working notepad that survives conversation compaction. One skill, one storage convention, file-backed under `.vibekit/`. Cross-runtime portable. |
The single durable-knowledge surface for vibekit. Two file-backed regions under .vibekit/:
type: frontmatter, not by which directory it lives in.A single append-only log.md chronicles all writes. An auto-maintained INDEX.md catalogs every entry.
All five CLIs (Claude Code, Codex, Gemini, opencode, Pi) read and write the same files identically. No database, no daemon, no runtime API.
brainstorm-lean step 1, brief-compiler CONTEXT enrichment).Do NOT invoke for:
docs/specs/ artifacts produced by brainstorm-lean. Specs reference memory; memory distills patterns from many specs..vibekit/
├── memory/
│ ├── INDEX.md # auto-maintained catalog
│ ├── log.md # append-only operation chronicle
│ └── <key>.md # one file per key, frontmatter + body
└── notepad.md # single file, three sections
<key> is kebab-case, ≤60 chars, descriptive: test-runner, auth-architecture, cache-stampede-postmortem-2026-q1, merge-freeze-2026-q2.
Every entry has this exact frontmatter, then a markdown body:
---
key: <kebab-case>
type: convention | decision | preference | gotcha | contact | deadline | architecture | pattern | debugging | environment | session-log | reference
tags: [<tag>, <tag>]
created: YYYY-MM-DD
updated: YYYY-MM-DD
status: active | stale | superseded
supersedes: <key, optional>
confidence: high | medium | low
---
# <Title>
<body — single paragraph for atomic facts; multi-section markdown for compound documents.
Use h2/h3 for structure when the body has more than one section.>
## See also
- [[<other-key>]] — <one-line reason for the link>
Atomic vs compound is signaled by length and structure, not by separate storage. A convention: test runner is vitest is one paragraph. An architecture: auth flow may be a multi-section document with cross-links. Both are entries.
confidence: low means the entry is a best guess — surface it but flag it. confidence: high means the user confirmed or the codebase verifies it.
Pick exactly one. If a finding does not fit any type, it probably does not belong in memory:
convention — coding/naming/process rules in force.decision — a choice and its rationale (often supersedes another).preference — user/team operational preference.gotcha — non-obvious pitfall.contact — person, team, or service reference.deadline — time-bound fact.architecture — how a subsystem fits together (compound).pattern — a repeatable approach used in multiple places (compound).debugging — playbook for diagnosing a class of issue.environment — how to run/build/test/deploy.session-log — chronicle of a notable run or incident.reference — external links, glossaries.Adding a new type is a deliberate decision that updates this list, not a per-entry expedient. Free-tagging happens via tags, not type.
[[<key>]] is the cross-link syntax for any entry. Plain markdown links ([text](url)) are also fine for external URLs but do not count as cross-references.
[[<key>]] must resolve to an existing entry. audit flags broken links. The skill never auto-creates missing target entries — that path leads to junk.
Single file, three sections, fixed headings. The skill writes/edits within sections; never reorders them.
# Vibekit Notepad
## Priority Context
<HARD LIMIT 500 chars total. Critical facts the next turns need. Bullet list.>
## Working Memory
- [YYYY-MM-DDTHH:MMZ] <one-line breadcrumb>
- ...
## Manual
<user-controlled, free-form, never auto-pruned>
Working Memory entries older than 7 days are pruned on audit. Priority and Manual are never auto-pruned.
read --key <K>Return the entry contents verbatim. If the key does not exist, return not found with three closest keys from INDEX. Never guess.
query --text <T> [--tags <list>] [--type <T>] [--include-superseded]Match against status: active entries by default. Matching is keyword + tag + type — no vector embeddings, no fuzzy. Returns:
hits:
- key: <key>
type: <type>
snippet: <first matching line, verbatim>
score: <integer; tag/type matches weighted higher than body>
The caller synthesizes an answer with [[key]] citations. The skill never paraphrases bodies into the answer — it returns hits and the caller reads entries it cares about.
list [--type <T>] [--tags <list>] [--include-superseded]Read INDEX.md and filter. Returns key + title + type + updated date. Bodies are not loaded.
read --notepad [--section priority|working|manual]Return the full notepad, or the named section verbatim.
write --key <K> --type <T> --confidence <C> [--tags <list>] [--body @file | --body-stdin]Create or update .vibekit/memory/<K>.md. If the key exists, update updated: and replace the body wholesale; never append silently. Preserve created: on update. Update INDEX.md. Append a one-line entry to log.md.
If a different key exists with the same body content (heuristic: identical first line + ≥80% body overlap), REJECT and propose update or supersede.
write --notepad --priority <text>Append <text> to Priority Context. After append, count chars. If total >500, REJECT with the diff between current and limit; do not truncate silently.
write --notepad --working <text>Prepend a timestamped bullet to Working Memory: - [YYYY-MM-DDTHH:MMZ] <text>.
write --notepad --manual <text>Append to Manual section verbatim.
supersede --old <K> --new <K>Mark <old> with status: superseded. Set supersedes: <new> on <new>. Both stay on disk; nothing is deleted. query hides superseded entries by default unless --include-superseded.
delete --key <K> --reason <R>Hard delete. REJECT unless --reason is given. Logged in log.md with the reason verbatim. Cross-references to the deleted key are NOT auto-updated; they show as broken on the next audit and the user fixes them. Default to supersede over delete — project history matters.
classify <finding>Given a free-text finding from a session, decide the right destination. Output:
finding: "<verbatim quote>"
proposed:
surface: entry | notepad-priority | notepad-working | notepad-manual | none
key: <key if entry>
type: <type if entry>
tags: [<tag>, ...]
confidence: <low|medium|high>
reason: <one line>
surface: none is correct when the finding is a code fact (lives in code), a spec fact (lives in docs/specs/), or already-known (already in INDEX). Default to none unless the finding is durable, reusable, and not derivable from the code. Do not write everywhere — that is the dump anti-pattern.
auditconfidence: low and updated: >30 days old as candidates for review.status: active and updated: >180 days old as stale.[[key]] cross-link resolves; report broken links per source key.oversized (consider splitting).superseded entries still linked from active entries.Output:
pruned_working: <count>
stale_low_confidence: [<key>, ...]
stale: [<key>, ...]
index_rebuilt: <bool>
duplicates: [{a: <key>, b: <key>, reason: <string>}]
broken_links: [{from: <key>, to: <key>, line: <n>}]
oversized: [<key>, ...]
superseded_with_active_inbound: [<key>, ...]
Audit is read-mostly. The only writes are the prune + INDEX rebuild. Never delete an entry without explicit user instruction.
These join the never-compress list. Compress narration around them; never the rules themselves.
classify verdict is surface: none. Memory rots when overfilled.confidence: low. Lying about confidence destroys the surface.supersede over delete. Project history matters; rot is preferable to silent deletion of decisions.review-pack may PROPOSE a write after sign-off; the user approves.@./.vibekit/notepad.md or @./.vibekit/memory/INDEX.md to its own entry doc — that is the runtime owner's choice, not this skill's responsibility.write writes one entry + INDEX + log. write --notepad --working prepends one bullet. No drive-by reorganization.git diff-able.classify defaults to surface: none. delete REJECTS without --reason. The skill challenges every write that risks rotting the surface.| Region | Policy |
|---|---|
| Operation announcements ("Stored.", "Pruned 2 entries.") | Compress |
| Frontmatter values | Verbatim |
| Entry bodies | Verbatim |
| Query hits / snippets | Verbatim |
| Notepad Priority Context | Verbatim, hard 500-char limit |
classify reasons | One terse line |
| Audit YAML report | Verbatim |
delete --reason text | Verbatim in log.md |
Cross-link syntax [[key]] | Verbatim, never paraphrased |
| Confidence flags / uncertainty markers | Verbatim |
Other skills MAY invoke memory-dual but are not required to:
brainstorm-lean step 1 — query for prior decisions/architecture in the relevant domain. Surface 1-3 hits before asking clarifying questions.brief-compiler — pull entries cited by the user into the CONTEXT block (verbatim; no compression).review-pack after sign-off — propose a classify pass on findings; with user approval, propose write for entries that generalize.verify-gate — does NOT write. Verification is not knowledge capture.These are suggestions, not hard wires. Each calling skill decides whether the round-trip is worth the tokens.
All operations are file I/O under .vibekit/. Verified portable to:
.claude-plugin/plugin.json..codex-plugin/plugin.json.GEMINI.md via @./skills/memory-dual/SKILL.md..opencode/plugins/vibekit.js (skills directory scan).No runtime-specific shim. No capability gate. No degradation.
confidence: high on a guess. The whole surface depends on this honesty.supersede.[[key]] is a claim that the linked entry is relevant; verify before writing.INDEX.md. Run audit instead.For read-class operations: the requested content, verbatim, with no narration.
For write-class operations: a single short line — Stored: <key>. or Notepad priority updated (412/500 chars). or Updated <key> (+a/-r lines).
For classify: the YAML block exactly as specified above.
For query, list, audit: the YAML structure exactly as specified above.
For delete: Deleted <key>. Reason: <verbatim>.
No structured return value beyond what is documented per operation. The skill is file-producing and user-facing.