harvest

Original：🇺🇸 English

Translated

1 scripts

Use when users ask to capture conversation decisions, problems, and lessons into persistent notes (e.g. 'harvest', '/harvest', 'save this to second brain', 'document this work').

6installs

Sourceshihyuho/skills

Added on2026-02-11

NPX Install

npx skill4agent add shihyuho/skills harvest

SKILL.md Content

View Translation Comparison →

Harvest

Capture high-signal conversation knowledge into

docs/notes/

for reuse across sessions.

Terminology:
harvest
is the skill name (the action of capturing knowledge). The second brain it builds belongs to the project — name it
[Project Name] Second Brain
, not "Harvest".

When to Trigger

Explicit request:

"harvest", "/harvest"
"harvest this", "harvest this conversation"
"save this to second brain"
"document this work", "capture this knowledge"

Natural breakpoints — AI may suggest (never auto-execute):

Significant decision made
Complex problem solved after multiple attempts
Important lesson learned
Discussion wrapping up with valuable content

We just [made a decision/solved X/learned Y].
Would you like me to update the second brain?

Wait for user confirmation. Never auto-execute.

Non-Negotiables

Never auto-execute. Always wait for explicit user confirmation.
Use template files as the single source of structure.
Keep
```
context_id
```
in frontmatter for smart merge.
Keep MOCs as index files: link to context anchors, do not duplicate full lesson content.
Omit optional empty sections. Do not write empty headers or "None" placeholders.
Keep behavior logic in this file; keep structure/manifest data in
```
references/
```
.
Use list-first formatting for context content. Use tables only for short, fixed-width, easy-to-scan fields.

Quick Start

Ensure
```
docs/notes/
```
exists (initialize when missing).
Run lesson review for this harvest run.
Detect
```
context_id
```
and check whether matching context file exists.
Create new context or smart-merge existing context.
Update
```
docs/notes/00-INDEX.md
```
and related MOCs.
Confirm created/updated files to user.

Workflow

Phase 1: Prepare

Initialize second brain storage when
```
docs/notes/
```
is missing:
- Use references/initialization-manifest.md as the single source for initialization inventory.
- Follow idempotent behavior: create missing files/directories, preserve existing files unless explicit update is requested.
- Check
```
AGENTS.md
```
  then
```
CLAUDE.md
```
  ; if both exist or neither exists, ask user which file to append, then append section from references/agents-lessons-section.md.
- Do not own or maintain
```
.base
```
  template syntax in this skill. If
```
obsidian-bases
```
  is available, delegate default base-file creation to
```
obsidian-bases
```
  using references/bases-generation-spec.md.
Mandatory lesson review for this harvest run:
- Read
```
docs/notes/00-INDEX.md
```
  when available.
- Scan
```
docs/notes/mocs/lessons-learned.md
```
  when available.
- Apply matched lessons (tech, operation type, error pattern).

Phase 2: Detect + Route

Build
```
context_id
```
:
- Prefer helper script
```
scripts/context_id.py
```
  (Python stdlib only, no third-party packages).
- Default call:
```
python3 scripts/context_id.py --format json
```
  (or
```
python
```
  when
```
python3
```
  is unavailable).
- Stateless policy (no state file): env keys (
```
OPENCODE_*_ID
```
  /
```
*_ID
```
  ) -> optional
```
--infer-latest-session
```
  -> generated
```
ctx-YYYYMMDDHHMMSS-<hex6>
```
  .
- Do not pass
```
--infer-latest-session
```
  by default. Use it only when the user explicitly requests latest-session inference.
- If script is unavailable or fails, use inline fallback: first env var matching
```
*SESSION*ID*
```
  /
```
*CONVERSATION*ID*
```
  /
```
*THREAD*ID*
```
  ; otherwise generate
```
ctx-YYYYMMDDHHMMSS-<hex6>
```
  .
Look for
```
docs/notes/contexts/<context_id>-*.md
```
.

Condition	Action
Matching context file exists	Run Phase 4 (Smart Merge)
No matching context file	Run Phase 3 (New Context)
User rejects suggested filename	Ask for new slug and regenerate filename
User cancels confirmation	Stop without writing files

Phase 3: New Context

Extract high-signal items: work, decisions (with rationale), unsolved, lessons, optional source notes.
- Use stable item IDs for merge safety:
```
D-*
```
  ,
```
Q-*
```
  ,
```
LL-*
```
  .
- Keep technical signal high: merge routine process/SOP chatter into one short note instead of many standalone decisions.
- If planning files (
```
task_plan.md
```
  ,
```
findings.md
```
  ,
```
progress.md
```
  ) are present, capture medium-density snapshots (
```
conclusion + evidence + source note
```
  ) into context content.
- Do not copy full planning files. Do not create Obsidian wikilinks to files outside
```
docs/notes/
```
  .
- Exclude harvest-process artifacts from captured knowledge (command menus, filename suggestion prompts, merge/update status lines, and other capture bookkeeping).
Generate filename
```
<context_id>-<topic-slug>.md
```
and confirm:

Found: [N] decisions, [N] unsolved, [N] lessons
Suggested: contexts/<context_id>-<topic-slug>.md
1. Use this  2. Change slug  3. Cancel

Create
```
docs/notes/contexts/<filename>.md
```
from references/context-template.md.
Use
```
obsidian-markdown
```
when available for wikilinks/frontmatter/anchors.
Update
```
docs/notes/00-INDEX.md
```
according to references/index-template.md:
- Recent Updates (top 5), Topics, Key Decisions, Open Questions, Recent Lessons, Stats.
Manage MOCs:
- Lessons MOC: for error-related lessons (>15 min impact), ensure
```
docs/notes/mocs/lessons-learned.md
```
  exists via references/lessons-learned-moc-template.md, then append links only.
- Topic MOC: when a topic appears in 3+ contexts without MOC, ask user first; create from references/moc-template.md after confirmation.
Confirm:

✓ Created: contexts/<filename>.md
✓ Updated: 00-INDEX.md
✓ Updated: mocs/lessons-learned.md (when relevant)
✓ Created: mocs/<topic>.md (when relevant)

Phase 4: Smart Merge

Read the existing context file.
Merge new items from current conversation.

Section	Existing Topic	New Item
Decisions Made	Match by `D-*` first, update entry and preserve anchor	Append
Still Unsolved	Match by `Q-*` ; move resolved items to Decisions with trace	Append
Lessons Learned	Match by `LL-*` first; merge same lesson and preserve anchor	Append
What We Worked On	Keep existing	Append
Source Notes	Merge by source note signature if duplicated	Append

Update frontmatter (
```
updated
```
, tags), keep
```
created
```
unchanged.
Update
```
00-INDEX.md
```
stats and recent updates.
Update
```
mocs/lessons-learned.md
```
with links only when new error-related lessons were added.
Confirm:

✓ Updated: contexts/<filename>.md
Changes: [added/updated/moved items]

Content Quality Rules

Quality over quantity: concise, reusable, high-signal notes.
One idea per bullet; include rationale for decisions.
Keep content relevant for future reuse (3+ months horizon).
Skip raw transcripts, dead ends without insight, and obvious process noise.
Exclude harvest operational chatter: confirmation menus,
```
created/updated/skipped
```
summaries, and "how harvest processed this" narration unless it records a reusable lesson.
Use stable IDs (
```
D-*
```
,
```
Q-*
```
,
```
LL-*
```
) with anchored headings for merge-safe entries.
Prefer list-first sections for long text. Use tables only when each cell stays short (for example
```
ID | Question | Next
```
).
Avoid duplicate restatement: decision headings should carry title intent; item body should add why/impact.
For lessons, default to compact triad (
```
Issue
```
,
```
Root Cause
```
,
```
Fix
```
); add
```
Guardrail
```
/
```
Apply When
```
only when they provide non-obvious value.
Avoid repeating
```
Related
```
links on every item; prefer one section-level related MOC link.
Omit low-signal metadata (
```
Deadline
```
,
```
Carry-Over
```
) unless it materially changes follow-up decisions.
For carry-over items from prior sessions, label clearly and avoid presenting them as new outcomes.
For planning-derived content, store medium-density snapshots (
```
conclusion + evidence + source note
```
) in context files; avoid external-file wikilinks outside
```
docs/notes/
```
.

Recommended section limits:

Section	Requirement	Typical Size
Summary	Required	1-2 sentences
What We Worked On	Required	5-7 bullets
Decisions	Optional	up to 5 items
Still Unsolved	Optional	up to 3-5 items
Lessons Learned	Optional	up to 3 items
Source Notes	Optional	up to 3-5 items
Notes	Optional	short snippets only

Integration

Recommended:

obsidian-markdown

skill for enhanced Obsidian compatibility. Without it, AI may write files directly and suggest installation:

For optimal Obsidian compatibility, consider installing obsidian-markdown skill:
npx skills add <obsidian-markdown-repo>

Optional companion:

obsidian-bases

for

.base

file generation.

This skill does not own
```
.base
```
syntax templates.
When
```
obsidian-bases
```
is available, delegate generation of default base files to that skill.
Keep desired base content in references/bases-generation-spec.md, not
```
.base
```
syntax templates.
If
```
obsidian-bases
```
is unavailable, skip
```
.base
```
creation and continue markdown-only harvest flow.