Knowledge Acquisition

LLM Wiki

Build and maintain a persistent, compounding knowledge base as interlinked markdown files. Based on Karpathy's LLM Wiki pattern.

Unlike RAG (which rediscovers knowledge from scratch per query), the wiki compiles knowledge once and keeps it current. Cross-references are already there. Contradictions have already been flagged. Synthesis reflects everything ingested.

First gather raw sources of knowledge, then summarize, cross-reference, file, and maintain consistency.

Key Principle: Distill to State-of-the-Art

The wiki is not an archive of everything ever written on a topic. It is a distilled, current view of the best known approaches. When writing or updating wiki pages:

• Keep pages small and focused — each page should capture the current best approach to one concept, not a historical survey. Target ~50–150 lines.
• Obsolete methods get removed, not accumulated. If a newer paper supersedes an older technique, update the page to reflect the new state-of-the-art. Mention the predecessor briefly for context, not in detail.
• Prioritize recency and impact. When multiple approaches exist, lead with the one that has the best results or widest adoption today. Don't give equal weight to a 2019 method and a 2025 method that dominates it.
• Archive aggressively. If an entire page becomes obsolete (e.g., a technique fully superseded), move it to _archive/ rather than letting it clutter the wiki.
• The wiki should answer: "What is the best way to do X right now?" — not "What are all the ways people have tried X since 2017?"

When This Skill Activates

Use this skill when the user:

• Asks to research a topic, find papers, or gather sources
• Asks to create, build, or start a wiki or knowledge base
• Asks to ingest, add, or process a source into their wiki
• Asks a question and an existing wiki is present
• Asks to lint, audit, or health-check their wiki

Architecture: Three Layers

The wiki directory name is configurable — it does NOT have to be wiki/. Set WIKI_DIR to the wiki's root directory. When working with the platform (multiple wikis), each wiki has its own name (e.g., llm-wiki/, security-playbook/). The default is wiki/ for single-wiki projects.

# Set WIKI_DIR to the wiki's root directory (default: wiki/)
export WIKI_DIR=wiki  # or: llm-wiki, security-playbook, etc.

$WIKI_DIR/
├── SCHEMA.md           # Conventions, structure rules, domain config
├── index.md            # Sectioned content catalog with one-line summaries
├── log.md              # Chronological action log (append-only)
├── raw/                # Layer 1: Immutable source material
│   ├── papers.db.jsonl # Local paper index
│   ├── papers/         # Downloaded PDFs and LaTeX sources
│   ├── articles/       # Web articles, clippings
│   ├── code/           # Cloned GitHub repositories
│   └── assets/         # Images, diagrams
├── entities/           # Layer 2: Entity pages (people, orgs, models)
├── concepts/           # Layer 2: Concept/topic pages
├── comparisons/        # Layer 2: Side-by-side analyses
└── queries/            # Layer 2: Filed query results worth keeping

Layer 1 — Raw Sources (raw/): Immutable. The agent reads but never modifies these. Layer 2 — The Wiki: Agent-owned markdown files. Created, updated, and cross-referenced by the agent. Layer 3 — The Schema: SCHEMA.md defines structure, conventions, and tag taxonomy.

Core Operations

1. Ingest

When the user provides a source or asks to research a topic:

① Gather raw sources using the acquisition scripts (see below). Start with a web search to discover relevant blogs, articles, and leads — then narrow to academic sources with the scripts below. ② Discuss takeaways with the user — what's interesting, what matters. ③ Check what already exists — read index.md, search for existing pages. ④ Write or update wiki pages — create entity/concept pages, add cross-references ([[wikilinks]]), use tags from the taxonomy. ⑤ Update navigation — add to index.md, append to log.md. ⑥ Report what changed — list every file created or updated.

2. Query

When the user asks a question:

① Read index.md to find relevant pages. ② Read the relevant pages. ③ Synthesize an answer with citations to wiki pages. ④ File valuable answers back as new pages in queries/ or comparisons/.

3. Lint

Periodically health-check the wiki: orphan pages, broken wikilinks, stale content, contradictions, missing cross-references, index completeness.

Source Acquisition Scripts

Scripts live in this skill's scripts/ directory. They use PEP 723 inline dependencies resolved automatically by uv run.

Important: Always run scripts from the user's project directory (where the wiki lives), not from the skill directory. Use the full path to the script:

# SKILL_DIR is the directory containing this SKILL.md file.
# The agent should resolve this from the skill's installed location, e.g.:
#   ~/.kiro/skills/knowledge-acquisition/scripts/

# WIKI_DIR is the wiki's root directory (default: wiki/)
# Set this to match the wiki name you're working with.
export WIKI_DIR=wiki  # or: llm-wiki, my-research, etc.

uv run $SKILL_DIR/scripts/<script>.py [args...]

All default output paths use $WIKI_DIR/raw/papers/, $WIKI_DIR/raw/code/, etc. Scripts accept --output-dir to override. When using the default, ensure WIKI_DIR points to the correct wiki directory.

Wiki Naming

When creating a wiki (locally or on the platform), choose a slug-like name: lowercase, hyphens, no spaces. Examples:

• wiki — default for single-wiki projects
• llm-wiki — research on LLMs
• security-playbook — security knowledge base
• project-alpha-kb — project-specific knowledge

The name becomes:

• The local directory name (e.g., ./llm-wiki/)
• The CodeCommit repo suffix (e.g., wiki-llm-wiki)
• The SSM registry key (e.g., /wikis/llm-wiki)
• The URL path in the web app (e.g., /wiki/llm-wiki)

Typical Workflow

# Set SKILL_DIR to the absolute path of this skill's installation directory.
# (The agent should derive this from where it found SKILL.md)
# Set WIKI_DIR to the wiki's root directory.
export WIKI_DIR=wiki

# 0. Discover — use web search to find relevant sources, blogs, and leads
#    (Use the agent's built-in web search capability before narrowing to academic sources)

# 1. Orient — get background
uv run $SKILL_DIR/scripts/fetch_wikipedia.py summary "Mixture of Experts"

# 2. Search — find papers across multiple sources
uv run $SKILL_DIR/scripts/search_papers.py --query "mixture of experts routing" -o results.jsonl

# 3. Download — get the PDFs (saves to $WIKI_DIR/raw/papers/)
uv run $SKILL_DIR/scripts/download_papers.py --jsonl results.jsonl --output-dir $WIKI_DIR/raw/papers/ --max-downloads 10

# 4. Extract — convert to readable text
uv run $SKILL_DIR/scripts/pdf2md.py $WIKI_DIR/raw/papers/2101.03961.pdf

# 5. Find implementations (saves to $WIKI_DIR/raw/code/)
uv run $SKILL_DIR/scripts/clone_github_repo.py --repo google/flaxformer --output-dir $WIKI_DIR/raw/code/

# 6. Track what you've collected
uv run $SKILL_DIR/scripts/paper_db.py add --jsonl results.jsonl --db $WIKI_DIR/raw/papers.db.jsonl

Rate Limits

Script Reference

Paper Search

Unified Search (search_papers.py)

Search across arXiv, Semantic Scholar, and OpenAlex in one command with deduplication.

uv run $SKILL_DIR/scripts/search_papers.py --query "mixture of experts" --sources arxiv,s2,openalex -o results.jsonl
uv run $SKILL_DIR/scripts/search_papers.py --query "LLM reasoning" --sources arxiv,s2 --max-results 30

Options: --query, --sources (default: all), --max-results (default: 30), --min-citations, --api-key, -o.

arXiv Search (search_arxiv.py)

uv run $SKILL_DIR/scripts/search_arxiv.py --query "LLM agent" --categories cs.AI cs.CL --sort-by lastUpdatedDate -o results.jsonl

Options: --query, --max-results, --categories, --sort-by, --start-date, --end-date, -o.

Semantic Scholar Search (search_semantic_scholar.py)

uv run $SKILL_DIR/scripts/search_semantic_scholar.py --query "sparse MoE" --min-citations 50 -o results.jsonl
uv run $SKILL_DIR/scripts/search_semantic_scholar.py --citations-of "ARXIV:2101.03961" -o citing.jsonl
uv run $SKILL_DIR/scripts/search_semantic_scholar.py --references-of "ARXIV:2101.03961" -o refs.jsonl
uv run $SKILL_DIR/scripts/search_semantic_scholar.py --related-to "ARXIV:2101.03961" -o related.jsonl

Options: --query, --max-results, --min-citations, --year-range, --venue, --citations-of, --references-of, --related-to, --api-key, -o.

OpenAlex Search (search_openalex.py)

Free, no API key, 250M+ works.

uv run $SKILL_DIR/scripts/search_openalex.py --query "attention mechanism" --max-results 50 -o results.jsonl

Options: --query, --max-results, --min-citations, --year-range, --type, --sort, -o.

CrossRef Search (search_crossref.py)

Find DOI-registered publications, generate BibTeX.

uv run $SKILL_DIR/scripts/search_crossref.py --query "attention is all you need" --bibtex --output refs.bib

Options: --query, --rows, --output, --bibtex.

Paper Download

Generic Download (download_papers.py)

Downloads PDFs from any search JSONL output (reads pdf_url field).

uv run $SKILL_DIR/scripts/download_papers.py --jsonl results.jsonl --max-downloads 10 --sort-by-citations

Options: --jsonl, --output-dir (default: $WIKI_DIR/raw/papers/), --max-downloads, --delay, --sort-by-citations.

arXiv Download (download_arxiv.py)

uv run $SKILL_DIR/scripts/download_arxiv.py --arxiv-id 2101.03961
uv run $SKILL_DIR/scripts/download_arxiv.py --title "Attention Is All You Need"
uv run $SKILL_DIR/scripts/download_arxiv.py --arxiv-id 2101.03961 --format source

Options: --arxiv-id, --title, --format (pdf/source), --output-dir (default: $WIKI_DIR/raw/papers/).

Semantic Scholar Download (download_semantic_scholar.py)

uv run $SKILL_DIR/scripts/download_semantic_scholar.py --paper-id "ARXIV:2101.03961"
uv run $SKILL_DIR/scripts/download_semantic_scholar.py --jsonl results.jsonl --max-downloads 10

Options: --paper-id, --jsonl, --output-dir (default: $WIKI_DIR/raw/papers/), --max-downloads, --delay.

PDF Processing

PDF to Markdown (pdf2md.py)

uv run $SKILL_DIR/scripts/pdf2md.py $WIKI_DIR/raw/papers/2101.03961.pdf

Options: --margins LEFT TOP RIGHT BOTTOM.

Extract PDF Text (extract_pdf.py)

Structured extraction with section detection.

uv run $SKILL_DIR/scripts/extract_pdf.py --pdf $WIKI_DIR/raw/papers/2101.03961.pdf --sections-only
uv run $SKILL_DIR/scripts/extract_pdf.py --pdf-dir $WIKI_DIR/raw/papers/ --output-dir texts/

Options: --pdf, --pdf-dir, --output-dir, --sections-only.

Summarize Paper (summarize_paper.py)

Outputs structured JSON (title, abstract, sections array).

uv run $SKILL_DIR/scripts/summarize_paper.py $WIKI_DIR/raw/papers/2101.03961.pdf -o paper.json

Options: pdf (positional), -o, --sections-only.

Code & Implementations

Papers With Code (search_paperswithcode.py)

Find GitHub repos linked to papers (searches GitHub for repos referencing arXiv IDs in their READMEs).

uv run $SKILL_DIR/scripts/search_paperswithcode.py --arxiv-id 2101.03961 --output repos.jsonl
uv run $SKILL_DIR/scripts/search_paperswithcode.py --query "multi-agent" --output repos.jsonl

Options: --arxiv-id, --arxiv-ids, --arxiv-ids-file, --query, --output.

GitHub Search (search_github.py)

uv run $SKILL_DIR/scripts/search_github.py --query "mixture of experts" --language Python --min-stars 100 --output repos.jsonl

Options: --query, --output, --language, --min-stars, --sort, --max-results, --topic.

GitHub Code Search (search_github_code.py)

Search code snippets across GitHub.

uv run $SKILL_DIR/scripts/search_github_code.py --query "FlashAttention" --language Python --output results.jsonl

Options: --query, --output, --language, --filename, --max-results.

Clone Repository (clone_github_repo.py)

uv run $SKILL_DIR/scripts/clone_github_repo.py --repo google/flaxformer
uv run $SKILL_DIR/scripts/clone_github_repo.py --repo https://github.com/google/flaxformer.git

Options: --repo (owner/name or full GitHub URL), --output-dir (default: $WIKI_DIR/raw/code/), --depth, --branch.

Fetch README (fetch_github_readme.py)

Get README without cloning.

uv run $SKILL_DIR/scripts/fetch_github_readme.py --repos pytorch/pytorch --output readmes.jsonl

Options: --repos, --input, --output, --max-chars.

Fetch Repo Metadata (fetch_repo_metadata.py)

uv run $SKILL_DIR/scripts/fetch_repo_metadata.py --repos pytorch/pytorch --output metadata.jsonl

Options: --repos, --input, --output, --delay.

Reference & Context

Wikipedia (fetch_wikipedia.py)

uv run $SKILL_DIR/scripts/fetch_wikipedia.py summary "Transformer (machine learning model)"
uv run $SKILL_DIR/scripts/fetch_wikipedia.py sections "Gradient descent"
uv run $SKILL_DIR/scripts/fetch_wikipedia.py section "Gradient descent" --index 3

Subcommands: summary, sections, section --index N, wikitext.

DeepXiv Semantic Search (fetch_deepxiv.py)

Semantic (not keyword) paper search with progressive reading. Requires deepxiv-sdk.

uv run $SKILL_DIR/scripts/fetch_deepxiv.py search "low rank adaptation" --limit 10
uv run $SKILL_DIR/scripts/fetch_deepxiv.py brief 2106.09685
uv run $SKILL_DIR/scripts/fetch_deepxiv.py section 2106.09685 Introduction

Subcommands: search, brief, head, section, raw, trending, social.

Utilities

JSONL to BibTeX (jsonl_to_bibtex.py)

Convert any search output to .bib.

uv run $SKILL_DIR/scripts/jsonl_to_bibtex.py --jsonl results.jsonl --output refs.bib

Paper Database (paper_db.py)

Local index to track collected papers and avoid re-downloads.

uv run $SKILL_DIR/scripts/paper_db.py add --jsonl results.jsonl
uv run $SKILL_DIR/scripts/paper_db.py list
uv run $SKILL_DIR/scripts/paper_db.py search --query "attention"
uv run $SKILL_DIR/scripts/paper_db.py has --arxiv-id 2301.07041

Options: --db (default: $WIKI_DIR/raw/papers.db.jsonl).

Harvest Citations (harvest_citations.py)

Scan LaTeX for uncited claims, find candidates via Semantic Scholar.

uv run $SKILL_DIR/scripts/harvest_citations.py --tex main.tex --bib references.bib --output candidates.bib
uv run $SKILL_DIR/scripts/harvest_citations.py --tex main.tex --bib references.bib --dry-run

Options: --tex, --bib, --output, --max-rounds, --api-key, --dry-run, --verbose.

Infrastructure: Wiki Platform (AWS)

The skill includes a serverless AWS infrastructure for persisting and sharing wikis across multiple agents. The infrastructure is in $SKILL_DIR/infrastructure/.

Deploy the Platform (One-Time)

cd $SKILL_DIR/infrastructure
pip install -r requirements.txt
cdk bootstrap   # first time only
cdk deploy WikiPlatformStack

Stack outputs (saved to CloudFormation):

• WebAppURL — CloudFront domain for the React web app
• ApiEndpoint — base URL for API calls
• WikiAgentRoleArn — IAM role for agents to assume
• WikiChangesTopicArn — SNS topic for push notifications
• ReactUIBucketName — S3 bucket for deploying the React build

Create a Wiki (Runtime — No Redeploy)

# Human (Cognito-authenticated):
curl -X POST https://<ApiEndpoint>/api/wikis \
  -H "Authorization: Bearer <token>" \
  -d '{"name": "my-wiki", "description": "Research knowledge base", "owner": "research-agent"}'

# Agent (IAM SigV4):
# Sign request with WikiAgentRole credentials and POST to /api/wikis

Agent Access (Git Protocol)

Agents read and write wiki content via git. No write API — agents use git-remote-codecommit.

# Setup (once per agent environment)
pip install git-remote-codecommit

# Clone a wiki
git clone codecommit::<region>://wiki-<name>
cd wiki-<name>

# Set agent identity
git config user.name "my-agent"
git config user.email "my-agent@wiki.internal"

# Workflow: pull → edit → commit → push (retry on conflict)
git pull origin main
# ... make changes ...
git add -A
git commit -m "[my-agent] Add concept page: topic-name"
git push origin main || { git pull --rebase origin main && git push origin main; }

Discover Wikis

import boto3, json

ssm = boto3.client("ssm")
wikis = []
paginator = ssm.get_paginator("get_parameters_by_path")
for page in paginator.paginate(Path="/wikis/", Recursive=False):
    wikis.extend(json.loads(p["Value"]) for p in page["Parameters"])
# Each wiki has: name, description, repositoryName, cloneUrlHttp, owner, createdAt

Subscribe to Wiki Changes

aws sns subscribe \
  --topic-arn <WikiChangesTopicArn> \
  --protocol sqs \
  --notification-endpoint <agent-queue-arn>

API Endpoints

Pre-Computed Data (Static, No Lambda Cost)

On every push to a wiki, the wiki-indexer Lambda builds:

• /data/{wiki}/graph.json — knowledge graph (nodes + edges) for visualization
• /data/{wiki}/index.json — slug → file path map for wikilink resolution
• /data/{wiki}/search.json — full-text search corpus for client-side search

These are served as static files from CloudFront.

Local Development (No AWS Required)

The web app works locally without deploying to AWS. A local dev server reads from filesystem directories instead of CodeCommit.

# From your project directory — point to one or more wiki directories:
WIKI_DIRS=./llm-wiki:./security-playbook node $SKILL_DIR/webapp/dev-server.mjs

# Or put all wikis under a single parent directory:
node $SKILL_DIR/webapp/dev-server.mjs --wikis-dir ./wikis

# In another terminal, start the React dev server:
cd $SKILL_DIR/webapp && npm run dev
# Opens http://localhost:5173 — proxies /api and /data to the local backend

Each directory passed to the dev server becomes a browsable wiki. The directory name is the wiki name.

Running the Webapp (Quick Reference)

When the user asks to "run the webapp", "visualize the wiki", or "start the local server":

1. Ensure dependencies are installed (first time only):

   cd $SKILL_DIR/webapp && npm install
   

2. Start the dev server (backend — reads wiki from filesystem):

   # Point WIKI_DIRS at the wiki directory (colon-separated for multiple)
   WIKI_DIRS=$WIKI_DIR node $SKILL_DIR/webapp/dev-server.mjs
   

3. Start the React frontend (in a second terminal):

   cd $SKILL_DIR/webapp && npm run dev
   

4. Open http://localhost:5173 in a browser.

$SKILL_DIR is the absolute path to this skill's installation directory (where this SKILL.md lives). $WIKI_DIR is the absolute path to the user's wiki directory (set via environment or inferred from context).

Operational Scripts

The skill includes shell scripts in $SKILL_DIR/scripts/ for platform operations. All auto-detect configuration from CloudFormation stack outputs.

Deploy Webapp

Build the React frontend and deploy to S3/CloudFront:

$SKILL_DIR/scripts/deploy-webapp.sh
$SKILL_DIR/scripts/deploy-webapp.sh --skip-build        # redeploy existing build
$SKILL_DIR/scripts/deploy-webapp.sh --skip-invalidation # skip CF cache bust

Sync Wiki to AWS

Push local wiki changes to CodeCommit (triggers indexer Lambda):

$SKILL_DIR/scripts/sync-wiki.sh $WIKI_DIR
$SKILL_DIR/scripts/sync-wiki.sh $WIKI_DIR --message "Added new concept pages"

Create a New Wiki on the Platform

Register a local wiki directory on AWS (CodeCommit + SSM + initial push):

$SKILL_DIR/scripts/create-wiki.sh ./my-new-wiki --description "Research KB" --owner monti

Create a User

Create a Cognito user for webapp access:

$SKILL_DIR/scripts/create-user.sh alice --email alice@example.com
$SKILL_DIR/scripts/create-user.sh bob --password "MyPassword123!"

Environment Variables (All Scripts)

Pitfalls

• Never modify files in raw/ — sources are immutable. Corrections go in wiki pages.
• Always orient first — read SCHEMA + index + recent log before any operation.
• Always update index.md and log.md — they are the navigational backbone.
• Don't create pages for passing mentions — only when an entity appears in 2+ sources or is central to one.
• Every page needs cross-references — minimum 2 outbound [[wikilinks]].
• Tags must come from the taxonomy in SCHEMA.md.
• JSONL is the interchange format — search scripts output JSONL that download scripts consume.