pensieve

Original：🇺🇸 English

Translated

13 scripts

Project knowledge base and workflow routing. Knowledge contains previously explored file locations, module boundaries, and call chains — reuse directly instead of re-locating; decisions/maxims are settled architectural decisions and coding principles — follow them, don't re-debate; pipelines are reusable workflows. Use self-improve to capture new findings after tasks. Provides init, upgrade, migrate, doctor, self-improve, and loop tools.

2installs

Sourcekingkongshot/pensieve

Added on2026-03-07

NPX Install

npx skill4agent add kingkongshot/pensieve pensieve

SKILL.md Content

View Translation Comparison →

Pensieve

Route user requests to the correct tool. When uncertain, confirm first.

Intent Routing

Explicit intent takes priority: if the user explicitly names a tool or trigger word, route directly.
Infer from session stage (when no explicit command is given):
- New project or blank context →
```
init
```
  | Version uncertainty →
```
upgrade
```
  | Migration/residue uncertainty →
```
migrate
```
- Exploring codebase or locating issues → check
```
<USER_DATA_ROOT>/knowledge/
```
  first, then use
```
self-improve
```
  to write findings
- Development complete or retrospective signals →
```
self-improve
```
  | Complex task needing decomposition →
```
loop
```
When uncertain, confirm first.

<example> User: "Initialize pensieve for me" → Route: tools/init/_init.md User: "Check if there are any data issues" → Route: tools/doctor/_doctor.md User: "This task is complex, run it with loop" → Route: tools/loop/_loop.md </example>

Global Rules (Summary)

Boundary priority: version/compatibility issues go through
```
upgrade
```
; migration/residue cleanup goes through
```
migrate
```
.
Confirm before executing: when the user has not explicitly issued a command, confirm first.
Keep links connected:
```
decision/pipeline
```
must have at least one
```
[[...]]
```
link.
Read spec before writing data: read the corresponding README before creating/checking user data.

Full rules in
references/shared-rules.md

Tool Execution Protocol

Before executing any tool, read its

### Use when

section to confirm applicability. For tool boundaries and redirect rules, see

references/tool-boundaries.md

.

Routing Table

Intent	Tool spec (read first)	Trigger words
Initialize	`<SYSTEM_SKILL_ROOT>/tools/init/_init.md`	init, initialize
Version update	`<SYSTEM_SKILL_ROOT>/tools/upgrade/_upgrade.md`	upgrade, version
Structure migration	`<SYSTEM_SKILL_ROOT>/tools/migrate/_migrate.md`	migrate, migration, clean old paths
Health check	`<SYSTEM_SKILL_ROOT>/tools/doctor/_doctor.md`	doctor, check, format check
Archive experience	`<SYSTEM_SKILL_ROOT>/tools/self-improve/_self-improve.md`	self-improve, archive, reflect
Iterative execution	`<SYSTEM_SKILL_ROOT>/tools/loop/_loop.md`	loop, iterative execution, execute pipeline

Routing Failure Fallback

Ambiguous intent: return candidate routes and ask the user to confirm.
Tool entry unreadable: stop and report the missing path.
Incomplete input: collect missing information before executing.

<SYSTEM_SKILL_ROOT>

is injected by the SessionStart hook; user data path is fixed at

<project>/.claude/skills/pensieve/

.