wiki-ingest

Wiki Ingest

Ingest a source document into the knowledge wiki — creating structured, interlinked pages that compound over time.

Usage Template

Prompt

Use wiki-ingest on this source. Create source notes, concept pages, entity pages, navigation updates, and a verification summary.

Use Case

• Turning an article, PDF, transcript, or rough note into durable linked knowledge.

Expected Result

• The agent creates an immutable source note, 3-7 key insights, linked wiki pages, and a log entry.

Output Example

• SOURCES_DIR/src-YYYYMMDD-title.md, CONCEPTS_DIR/concept-name.md, ENTITIES_DIR/entity-name.md, and LOG_FILE update.

Verification Case

• New wiki pages have frontmatter, source references, at least two [[wikilinks]], and timeline entries.

Verified Effect

• Knowledge stops being a loose summary: the source becomes traceable, linked, and reusable across future sessions.

Success Metrics

• Creates one immutable source note, 3-7 key insights with block refs, and at least one linked wiki page.
• Each new wiki page has frontmatter, source references, at least two wikilinks, and a timeline entry.
• Final report lists created/updated files and flags single-source claims.
• Targeted post-ingest lint confirms no missing frontmatter, broken source refs, zero-inlink pages, or pages with fewer than two outbound wikilinks.
• Clippings are archived after successful ingest and the clipping queue is updated when applicable.
• Each concept page passes the Karpathy understanding gate: one-line thesis, source boundary, mechanism, connections, and what remains uncertain.
• V6 receipt states Trigger, Context, Steering/verification, State/write-back, and whether any rule was queued rather than promoted.

V5.2 Closure Add-on

• Classify every input as external-fact, human-experience, internal-state, or environment-signal.
• For high-value ingests, decide whether to create one behavior experiment and one creativity experiment.
• Record governance risks: single-source claims, missing provenance, stale-page risk, and review queue items.

V6 Knowledge OS Add-on

• Treat input-skills-obsidian as the fast front door and this skill as the durable compiler.
• Use Trigger -> Context -> Steering -> State before writing:
• Trigger: user request, clipping queue, file drop, scheduled daily note, or research handoff.
• Context: source boundary, existing wiki pages, relevant maps, and only the rules needed now.
• Steering: source-risk label, human review boundary, independent verification, and stop condition.
• State: source note, wiki pages, maps, log, dashboard queue, or skill/SOP promotion candidate.
• Promote an insight into a skill, SOP, schema rule, or automation only after repeated support or local verification, bounded macro-action shape, no provenance relaxation, and a cheap check.
• Prefer zero-overhead context: do not preload broad maps, whole reports, or historical summaries unless the current ingest needs them.

When to Use

• User drops a file and says "ingest this"
• User shares a link/article/video/PDF to be processed
• User says "add this to the wiki" or "save this to my knowledge base"
• User mentions a source that should be captured

Workflow

Step 0: Resolve Paths

Before writing, read system/config.md when available and resolve SOURCES_DIR, CONCEPTS_DIR, ENTITIES_DIR, MAPS_DIR, LOG_FILE, BEHAVIORS_DIR, and CREATIVITY_DIR. If no config exists, use the default STOW layout.

Step 0A: Define the Ingest Macro Action

Treat each ingest as an agentic macro action:

Objective:
Trigger:
Source boundary:
Owned vault paths:
Expected source note:
Expected wiki updates:
Verification evidence:
State / write-back:
Non-goals:
Stop condition:

Do not start a broad autonomous expansion unless the source has objective metrics and cheap verification. Otherwise ingest once, record uncertainty, and queue follow-up research.

Step 1: Capture

Create an immutable source note in SOURCES_DIR:

---
source_id: "src-YYYYMMDD-short-slug"
source_date: "YYYY-MM-DD"
source_title: "Original Source Title"
source_author: ""
source_type: "article | book | video-transcript | pdf | epub | notebooklm-mediated | local-synthesis | primary-filing | company-interview"
source_url: ""
input_class: "external-fact | human-experience | internal-state | environment-signal"
created: "YYYY-MM-DD"
knowledge_stage: captured
evidence_level: "single-source | multi-source | curated-map"
trust_level: "1-unverified | 2-expert-source | 3-primary-source"
hash: "sha256-16char"
status: "raw | ingested"
---

Rules:

• Never modify source files after creation
• Extract 3-7 Key Insights with block refs (^ki-short-name)
• Flag single-source claims with > [!warning] Single source
• Assign an input class: external-fact | human-experience | internal-state | environment-signal

Step 1A: Classify Source Risk

Use a source-risk label before writing claims:

| Source type | Default evidence | Required caution | |---|---|---| | primary filing / official docs | high | still check date, draft/final status, and missing fields | | article / book / transcript | medium | mark author perspective and single-source claims | | founder / company / investor interview | medium-low | treat metrics, adoption, roadmap, and performance as self-reported | | NotebookLM-mediated source | low | state that the original transcript/source is not fully archived | | local synthesis / PDF stack | low | treat as secondary-source synthesis; do not promote numbers without primary refs | | fast-changing financial/product claim | low | add review queue item for current docs or filings |

If a clipping duplicates an existing source, create or update provenance only when it adds useful block refs; otherwise link to the canonical source and log the duplicate.

Step 2: Auto-create Entity Pages

For every person, company, product, or project mentioned in the source:

---
tags:
  - domain/strategy
  - type/entity
type: entity
status: seed
created: "YYYY-MM-DD"
knowledge_stage: single-source
evidence_level: single-source
---

# Entity Name

• Check if entity page already exists → update if needed
• Use consistent naming (English for global entities, Chinese for local)

Step 3: Create/Update Concept Pages

For every key concept, framework, or methodology:

• Check if concept page already exists
• If exists: add new source reference, flag contradictions
• If new: create with status: seed, ≥2 [[wikilinks]], knowledge_stage: single-source

Standardized page structure (all concept pages should follow this pattern):

# Title

> Core thesis in one line — what is the single most important thing to know?
> (Source: [[source]])

---

## Core Mechanism

ASCII diagram showing causal relationships.

## Classifications / Comparisons

Tables comparing with related frameworks.

## Key Data (if applicable)

Bulleted list of critical numbers.

## Implications / Applications

What this means for decision-making.

## Connections

- Related concepts with brief explanation of relationship
- Source references

---

## Evolution Timeline

- **YYYY-MM-DD**：Created.

Rules:

• Start every page with a quote block that captures the core thesis
• Use ASCII diagrams for mechanisms (causal chains)
• Use tables for comparisons (frameworks, classifications, data)
• End with connections (grouped wikilinks) + evolution timeline (separated by ---)

Karpathy understanding gate:

Before writing a concept page, answer:

Core thesis: What changed in my understanding?
Mechanism: What causes what?
Boundary: What is this source allowed to prove?
Counterpoint: What could make this wrong?
Reusable bit: Is this a concept, SOP, skill, decision rule, or review item?

The page should compile understanding for humans and agents. Do not store a long summary that future agents must re-understand from scratch.

Step 4: Update Navigation

1. Update the wiki overview / living synthesis
2. Update the central index
3. Update relevant MOC (map of content) pages
4. Update source batch index when the source starts or extends a topic batch
5. For Web Clipper inputs, move the processed clipping to Clippings/archive/ and update Clippings/README.md if that queue exists
6. If ingest came from the daily knowledge loop, update only the supervised KR evidence section or log entry; do not rewrite the automated snapshot block.

Step 5: Convert to Behavior / Creativity When Warranted

After a high-value ingest, run these checks:

| Check | Create / Update | |-------|-----------------| | The source implies a habit, action, identity shift, pain point, or repeated decision | BEHAVIORS_DIR/action-sops/ or BEHAVIORS_DIR/reviews/ | | The source suggests a product, content asset, analogy, cross-domain pattern, or minimum test | CREATIVITY_DIR/experiments/ | | The source contains single-source claims, missing provenance, or stale-page risk | SYSTEM_DIR/governance-dashboard.md or review queue |

Behavior experiment minimum fields:

• source concept
• behavior hypothesis
• trigger
• 15-minute action
• success metric
• review date

Creativity experiment minimum fields:

• source ingredients
• combination formula
• target user
• hypothesis
• minimum test
• success signal
• next output

Skill / SOP conversion check:

• If the source teaches a repeatable agent workflow, mark it as a candidate for skills/, commands/, or wiki/sops/.
• Extract only the few "human bits" that matter: objective, constraints, failure modes, verification, and write-back target.
• Do not create a new skill when an existing skill can absorb the rule.
• V6 promotion requires evidence support plus objective check; otherwise write the candidate to governance/review queue instead of changing a skill.

Step 6: Append Timeline

After updating or creating any wiki page, append a timeline entry at the bottom (below a --- separator):

---

## 演化时间线

- **YYYY-MM-DD**：创建初始版本。
- **YYYY-MM-DD**：更新了关于 X 的理解（来源：[[source]]）。旧观点 Y，新证据表明 Z。

Rules:

• Content above --- is compiled truth — current best understanding, rewritable
• Content below --- is the timeline — append-only, never edited or deleted
• Add an entry on creation, and every time a core claim changes

Step 7: Log

Append to LOG_FILE with:

• Source ingested
• Entity pages created/updated
• Concept pages created/updated
• Maps updated
• Behavior / creativity experiments created or explicitly skipped
• Governance risks recorded

Step 8: Targeted Post-Ingest Lint

Before final reporting, run a targeted health check over the files touched:

• source file exists and block refs resolve
• every new/updated wiki page has V6-compatible frontmatter
• every new/updated wiki page has V6-compatible provenance fields when the local schema requires them
• every new/updated wiki page has at least two outbound wikilinks
• every source reference points to an existing source
• every new concept/entity is linked from at least one relevant MOC or page
• no empty files were created
• concept pages pass the understanding gate rather than being raw summaries
• weak links are reported, not auto-created, unless the source clearly defines the concept/entity

For large batches, write or update system/lint-report.md with P0/P1 results and queue P2/P3 debt instead of trying to fix all historical structure issues.

Quality Gates

• [ ] Source note created with frontmatter
• [ ] ≥2 entity pages created or updated
• [ ] ≥1 concept page created or updated
• [ ] Every wiki page has ≥2 [[wikilinks]]
• [ ] Navigation (overview + index + maps) updated
• [ ] Behavior conversion assessed
• [ ] Creativity conversion assessed
• [ ] Governance risks recorded
• [ ] Log appended
• [ ] Targeted post-ingest lint completed
• [ ] Clipping archived or left in queue with reason
• [ ] Understanding gate passed for new/updated concepts
• [ ] V6 receipt includes Trigger, Context, Steering/verification, State/write-back, and rule-promotion decision

Page Format

Every wiki page uses this frontmatter:

---
tags: [domain/xxx, type/concept|entity|sop]
type: concept | entity | sop
status: seed | growing | evergreen | stale
created: YYYY-MM-DD
knowledge_stage: captured | stored | cross-checked | applied
evidence_level: single-source | multi-source | curated-map
---

Key Principles

• Immutable sources: raw/inbox files are never modified after ingest
• Compound: every new source makes the wiki richer
• Linked: every page ≥2 outbound [[wikilinks]]
• Sourced: every claim cites (Source: [[source-file#^ref]])
• Contradictions flagged: > [!warning] Contradiction when sources disagree
• No fake provenance: do not invent hashes, source IDs, or primary-source confidence for old or mediated materials
• Debt is explicit: weak links, stale metadata, duplicate sources, and fast-changing claims go to lint report or governance queue
• Markdown first: durable Markdown pages are the knowledge product; vector search is optional retrieval support, not a substitute for understanding