Track the source-of-truth code, skills, bootstrap rules, docs, and packaging files before adding new features. Exclude local vault config, generated skill mirrors, temporary exports, virtual environments, and business data files from version control. Validation: inspected staged file list and git ignored status before commit.
4.8 KiB
name, description, compatibility, metadata, allowed-tools
| name | description | compatibility | metadata | allowed-tools | ||
|---|---|---|---|---|---|---|
| wiki-quick-chat-capture | Fast, zero-friction capture of technical findings from the current conversation to the wiki's _raw/ staging area. Use this skill when the user says "/wiki-quick-chat-capture", "quick capture", "capture this finding", "save this bug fix", "capture this gotcha", "drop this to raw", "quick save to wiki", or wants to capture a non-obvious discovery mid-session without a full wiki-ingest run. Writes one _raw/ file per topic cluster in under 60 seconds — no subagents, no QMD updates, no manifest writes. Run /wiki-ingest or /data-ingest later to promote raw files to proper wiki pages. | Requires ~/.obsidian-wiki/config or OBSIDIAN_VAULT_PATH env var. qmd CLI optional. |
|
Bash Read Write |
Wiki Quick Chat Capture
Extract reusable technical findings from the current conversation and stage them in _raw/ for
later promotion. The goal is zero-friction capture of discoveries that would otherwise be lost
when the session ends.
Speed contract: Inline only. No subagents. No QMD. No manifest writes. Target: <60 seconds.
When to Use This Skill
Right tool when:
- The user just hit a non-obvious bug or confirmed a framework gotcha mid-session
- There's a concrete finding worth keeping but a full
/wiki-ingestrun would be too disruptive - The user wants to offload the knowledge now and defer promotion to end-of-day
NOT a replacement for wiki-capture (promotes directly to a final wiki page) or
wiki-ingest (full document ingestion with cross-links and manifest tracking).
Before You Start
- Resolve config — follow the Config Resolution Protocol in
llm-wiki/SKILL.md(walk up CWD for.env→~/.obsidian-wiki/config→ prompt setup). Extract:OBSIDIAN_VAULT_PATHOBSIDIAN_RAW_DIR(default:$OBSIDIAN_VAULT_PATH/_raw)
- Ensure
$OBSIDIAN_RAW_DIRexists. If not, create it.
Step 1: Scan the Conversation for Findings
Extract reusable technical knowledge — things worth having in 3 months with no memory of this session.
Capture:
- Non-obvious bugs and their root causes
- Framework or library gotchas (undocumented behavior, edge cases)
- API behavior that surprised the user
- Workarounds or fixes that required investigation
- Environment/toolchain quirks
- Patterns that emerged from debugging or testing
Skip:
- Project management updates, roadmap changes, config already in CLAUDE.md
- Exploratory back-and-forth where no conclusion was reached
- Things obvious from the docs or boilerplate any developer finds on first read
- Pleasantries, meta-conversation, status updates
If nothing material emerged, tell the user and stop.
Step 2: Cluster by Topic
Group related findings — one raw file per topic cluster, not per individual finding.
Examples: "Swift 6 concurrency gotchas", "Next.js hydration edge cases", "Postgres advisory locks".
Name each cluster as a kebab-case slug: swift-actor-reentrancy, nextjs-hydration-mismatch.
Step 3: Infer Project Context
Check the conversation for clues — repo names, file paths, framework mentions, error messages.
Use the most specific project name you can reliably infer. If unclear, use null.
Step 4: Write Raw Files
For each cluster, write $OBSIDIAN_RAW_DIR/<ISO-date>-<slug>.md.
Read references/RAW-FORMAT.md for the full frontmatter spec, body structure (finding block
format), and the provenance/confidence calibration table.
Quick reference for frontmatter fields that vary per cluster:
title— descriptive cluster titletags— 2–4 domain tags matching the vault taxonomysummary— 1–2 sentences, ≤200 charsproject— inferred project name ornullbase_confidence— 0.6 (discussed, unconfirmed) → 0.75 (fix applied) → 0.9 (test confirmed)provenance.extracted/provenance.inferred— must sum to 1.0lifecycle_changed— today's ISO datesources—"<project> session (<YYYY-MM-DD>)"
Step 5: Confirm to User
Staged to _raw/:
_raw/2026-05-27-swift-actor-reentrancy.md — "Actor reentrancy causes deadlock in async forEach"
_raw/2026-05-27-xcode-derived-data-cache.md — "Stale derived data silently breaks incremental builds"
Run /wiki-ingest (or /data-ingest) to promote these to full wiki pages.
If nothing was captured: "Nothing worth capturing found in this session."
What This Skill Does NOT Do
- No manifest writes —
_raw/files are not tracked in.manifest.json - No
index.md,log.md, orhot.mdupdates — those happen during promotion - No QMD refresh — raw files are drafts, not indexed content
- No subagents — everything runs inline in this context window
These constraints are intentional. Speed is the point. Promotion via /wiki-ingest handles
all of the above when the user is ready.