Instalar en Claude Code
Copiargit clone --depth 1 https://github.com/NateBJones-Projects/OB1 /tmp/updating-thoughts && cp -r /tmp/updating-thoughts/skills/updating-thoughts ~/.claude/skills/updating-thoughtsDespués abre una sesión nueva de Claude Code; el skill carga automáticamente.
Definición
SKILL.md
# Updating Thoughts
## Overview
The core Open Brain MCP server captures and reads thoughts but cannot change
them. `update_thought` (the `open-brain-update-thought` connector) fills that
gap. Choosing *how* to update is the non-obvious part: tagging and rewriting
behave and cost differently, and the tool acts on a thought's **UUID**, which
it will not look up for you.
## When to Use
- Correcting or rewriting a thought's text
- Re-classifying or tagging it (status, review flags, cross-references)
- Annotating it without touching the wording
Not for: creating a thought (use capture), or removing one — if the thought is
outdated or wrong, prefer updating or tagging it `superseded` over deleting
(see the **deleting-thoughts** skill).
## Process
1. **Resolve the id first.** `update_thought` takes a UUID, not a description.
Use `search_thoughts` / `list_thoughts` to find the target and confirm it is
the right one. Never guess a UUID.
- **Inspect before editing.** Before you overwrite anything, read the target's
full content with `get_thought` (or the connector's `fetch` tool), and use
`related_thoughts` to see what it connects to — so you don't clobber context
other thoughts depend on. Tool names may carry a connector prefix; use
whatever the environment exposes.
2. **Pick the mode:**
- Tag / re-classify only → pass `metadata_patch` (shallow-merges keys; leaves
content and unmentioned keys alone; **no re-embedding**).
- Change the wording → pass `content` (replaces the text **and re-embeds** so
semantic search stays accurate).
- Both → pass both; they compose in one call.
3. **Guard read-modify-write (optional).** If you read the thought, reasoned,
then write back — and other writers may exist — pass `if_unchanged_since` set
to the `updated_at` you read. The write is rejected with `STALE_READ` if the
row changed underneath you; re-fetch and retry.
## Quick Reference
| Goal | Field | Re-embeds? |
|------|-------|-----------|
| Add/change tags, status, links | `metadata_patch` | No |
| Rewrite the note text | `content` | Yes |
| Prevent a lost update | `if_unchanged_since` | — |
`metadata_patch` is a **merge, not a replace**: `{"status":"done"}` adds or
overwrites only `status` and leaves everything else in the metadata intact.
## Output
A confirmation naming the thought id, what changed (content replaced and
re-embedded, and/or metadata merged), and the new `updated_at`.
## Common Mistakes
- Using `content` just to add a tag — wastes an embedding call; use
`metadata_patch`.
- Expecting `metadata_patch` to replace the whole metadata object — it merges.
- Skipping the search step and passing a half-remembered UUID.
## Notes
- A restricted/sensitive thought may refuse a content update by policy.
- Connector: `open-brain-update-thought`, auth via `?key=` or the `x-brain-key`
header — the same `MCP_ACCESS_KEY` as your core Open Brain connector.Del mismo repositorio
aiceptionSkill
Continuous learning system that extracts reusable knowledge from work sessions. Triggers: (1) /aiception command, (2) 'save this as a skill' or 'extract a skill from this', (3) 'what did we learn?', (4) after non-obvious debugging or trial-and-error discovery. Creates new skills when valuable reusable knowledge is identified. Integrates with Open Brain to prevent duplicates.
auto-captureSkill
|
auto-capture-claude-codeSkill
|
autodream-brain-syncSkill
|
competitive-analysisSkill
|
daily-digestSkill
Morning digest of yesterday's Open Brain thoughts, drafted to Gmail
deal-memo-draftingSkill
|
deleting-thoughtsSkill
|