Confluence To Agent Knowledge Base

Confluence to Agent Knowledge Base

Convert a company's Confluence spaces into a private, LLM-powered knowledge base hosted on GitHub. Works for ANY Confluence Cloud instance. No API tokens required — uses the Atlassian MCP connector for all Confluence access.

Prerequisites

Before starting, verify:

1. The user has the Atlassian MCP connector connected in Claude Code
2. The user has Python 3.9+ installed
3. The user has git configured with push access to their GitHub org

If the Atlassian connector is not connected, tell the user: "You need to connect your Atlassian account first. Go to Claude Code settings and add the Atlassian connector, then come back and run this skill again."

Overview — 6 Phases

Walk the user through these phases interactively:

1. Discovery — Find their Confluence, list spaces
2. Access Tiers — Group spaces by who should see them
3. Ingest — Pull all pages via MCP into markdown files
4. Compile — Set up the CLI tool and compile the wiki
5. GitHub — Create private repos and push
6. Auto-Sync — Schedule weekly updates

Always explain what you're doing at each step. Ask for confirmation before proceeding to the next phase.

Phase 1: Discovery

Step 1.1: Find Confluence Instance

Call getAccessibleAtlassianResources (no parameters needed).

This returns a list of Atlassian sites. Extract:

• id — the cloud ID (use this for all subsequent MCP calls)
• url — the site URL (e.g., "https://acme.atlassian.net")
• name — the site name

If multiple sites are returned, ask the user which one to use.

Store the cloudId and url for all subsequent operations.

Step 1.2: List Spaces

Call getConfluenceSpaces with:

• cloudId: the site URL or cloud ID from step 1.1

Present results to the user as a table:

SPACE KEY    NAME                          TYPE      STATUS    PAGES
─────────────────────────────────────────────────────────────────────
ENG          Engineering Team              global    current   ~
HR           Human Resources               global    current   ~
MGMT         Management                    global    current   ~
...

Filter out:

• Archived spaces (unless user requests them)
• Personal spaces (unless user requests them)

Store the full list with id (numeric space ID), key, name, type, status for each space. The id is required for getPagesInConfluenceSpace.

Step 1.3: Select Spaces

Ask the user: "Which spaces do you want to include in your knowledge base?"

Accept:

• "all" — include everything (excluding archived/personal)
• A list of space keys — e.g., "ENG, HR, MGMT, OPS"
• Exclusions — "all except ARCHIVE, TEST"

Store the selected spaces.

Phase 2: Access Tiers

Step 2.1: Explain Access Tiers

Tell the user: "Each access tier becomes a separate private GitHub repo. This lets you control who sees what — for example, management strategy docs stay separate from the general team knowledge base.

Most companies use 1-3 tiers:

• General — everyone on the team (operations, engineering, onboarding, etc.)
• Leadership — executives only (strategy, financials, board docs)
• Sales — sales team only (pricing, contracts, pipeline)"

Step 2.2: Group Spaces

Present the selected spaces and suggest groupings based on space names. Use these heuristics to suggest tiers:

• Spaces with "management", "executive", "board", "finance", "accounting" in the name → leadership tier
• Spaces with "sales", "pricing", "revenue", "pipeline" in the name → sales tier
• Everything else → general tier

Ask the user to confirm or modify. They might want:

• A single tier (no access restrictions)
• Custom tier names and groupings
• More than 3 tiers

Step 2.3: Name the Repos

For each tier, determine the repo name:

• General: {company-slug}-kb
• Additional tiers: {company-slug}-kb-{tier-name}

Example: acme-kb, acme-kb-leadership, acme-kb-sales

Ask the user for their GitHub org/username for the repo URLs.

Step 2.4: Suggest Categories

For each tier, suggest wiki categories based on the space names in that tier. Common categories:

• operations, technical, sales, management, onboarding
• finance, customer-success, meetings, projects, general

Ask the user to confirm or customize. Every tier should have a general catch-all.

Phase 3: Ingest

Step 3.1: Choose Working Directory

Ask the user where to create the knowledge base directories. Default suggestion: current working directory or ~/Documents/.

Create a directory for each tier:

mkdir -p {base_dir}/{tier-repo-name}/{raw,wiki,output}

Step 3.2: Pull Pages

For each tier, for each space in that tier:

1. Call getPagesInConfluenceSpace with:

   • cloudId: stored from Phase 1
   • spaceId: the numeric space ID (NOT the space key)
   • contentFormat: "markdown"
   • limit: 250

2. Handle pagination: if _links.next exists in the response, there are more pages. Extract the cursor from the next URL and call again with the cursor.

3. For each page in the response:

   • Extract: id, title, status, createdAt, body, parentId, version.number, version.createdAt
   • Skip pages with empty or very short body (< 50 chars)
   • Slugify the title: lowercase, replace spaces/special chars with hyphens, max 80 chars
   • Clean the markdown body:
     • Remove blob image URLs: ![...](blob:https://media.staging.atl-paas.net/...) → [image]
     • Remove custom tags: <custom ...>...</custom> → empty
     • Remove zero-width spaces
     • Collapse 4+ blank lines to 3

4. Write to raw/{space_key_lowercase}/{slug}--{page_id}.md with this format:

---
confluence_id: "{page_id}"
title: "{title}"
space_key: "{SPACE_KEY}"
space_name: "{Space Name}"
status: "{status}"
created_at: "{createdAt}"
updated_at: "{version.createdAt}"
version: {version.number}
labels: []
ingested_at: "{current_iso_timestamp}"
---

# {title}

{cleaned_body}

5. After each space, report progress: "Pulled {N} pages from {Space Name} ({SPACE_KEY})"

6. Process spaces sequentially to avoid rate limits. Add a brief pause between spaces.

Step 3.3: Report Totals

After all spaces are ingested, show a summary table per tier:

TIER: acme-kb (General)
  Engineering (ENG): 245 pages
  Operations (OPS): 189 pages
  Onboarding (HR): 93 pages
  TOTAL: 527 pages

TIER: acme-kb-leadership
  Management (MGMT): 312 pages
  Accounting (FIN): 87 pages
  TOTAL: 399 pages

Phase 4: Compile

Step 4.1: Set Up Python Project

For each tier directory:

1. Copy the Python source from this skill's references directory. Read each file from ~/.claude/skills/confluence-to-agent-knowledge-base/references/ and write it to the tier directory:

   • references/confluence_kb/*.py → {tier_dir}/confluence_kb/
   • references/pyproject.toml → {tier_dir}/pyproject.toml
   • references/requirements.txt → {tier_dir}/requirements.txt
   • references/build_sync_state.py → {tier_dir}/build_sync_state.py

2. Copy setup script:

   • ~/.claude/skills/confluence-to-agent-knowledge-base/scripts/setup.sh → {tier_dir}/setup.sh

Step 4.2: Configure Space-Category Mapping

Edit {tier_dir}/confluence_kb/compile_kb.py and populate the empty space_category_map dict in the compile_fast method with the mappings decided in Phase 2. Example:

space_category_map = {
    "ENG": "technical",
    "OPS": "operations",
    "HR": "onboarding",
    "MGMT": "management",
}

Also add title-based overrides for meetings:

if any(w in title_lower for w in ["meeting", "standup", "stand-up", "weekly", "daily"]):
    category = "meetings"

Step 4.3: Generate Config

Write {tier_dir}/ckb-config.yaml:

company_name: "{Company Name}{tier suffix}"
company_description: >
  {Description of what this tier covers}

confluence:
  url: "{confluence_url}"
  email: "${CONFLUENCE_EMAIL}"
  api_token: "${CONFLUENCE_API_TOKEN}"
  cloud_id: "{cloud_id}"

spaces:
  include:
    - {SPACE_KEY_1}
    - {SPACE_KEY_2}
  exclude: []
  include_archived: false
  include_personal: false

compile:
  model: "claude-sonnet-4-20250514"
  max_tokens: 4096
  categories:
    - {category_1}
    - {category_2}
    - general
  min_content_length: 100
  generate_concepts: true
  generate_backlinks: true

raw_dir: "raw"
wiki_dir: "wiki"
output_dir: "output"
anthropic_api_key: "${ANTHROPIC_API_KEY}"

Step 4.4: Install and Compile

For each tier:

cd {tier_dir}
python3 -m venv venv
source venv/bin/activate
pip install -q -r requirements.txt
pip install -q -e .
python build_sync_state.py
ckb compile --fast

Report the compile results: "Compiled {N} pages into {M} wiki articles across {K} categories"

Step 4.5: Generate CLAUDE.md

Write a CLAUDE.md for each tier that explains:

• What this repo is (company Knowledge Base compiled from Confluence)
• The directory structure (raw/, wiki/, output/)
• How to search: ckb search "query", read wiki/_summaries.md, read wiki/_index.md
• How agents contribute: write new .md files to wiki/{category}/ with proper frontmatter
• The categories available
• That source_space: AGENT marks agent-contributed content

Step 4.6: Generate README.md

Write a README.md for each tier that includes:

• Repo name and access level
• Setup instructions (clone, bash setup.sh, source venv/bin/activate)
• Authentication options (Claude Pro/Max or Anthropic API key)
• Commands reference table
• Links to other tier repos (if multiple tiers)
• Auto-sync schedule info

Step 4.7: Generate .gitignore

__pycache__/
*.py[cod]
*.egg-info/
venv/
dist/
build/
.env
.env.local
.ckb-sync-state.json
.ckb-search-index.json
.DS_Store

Phase 5: GitHub

Step 5.1: Check for gh CLI

which gh && gh auth status

Step 5.2a: With gh CLI

For each tier:

cd {tier_dir}
git init
git add -A
git commit -m "Initial commit: {Company} knowledge base - {N} pages from {M} spaces"
gh repo create {org}/{repo-name} --private --source=. --push \
  --description "{description}"

Step 5.2b: Without gh CLI

Tell the user to create each repo manually on GitHub:

1. Go to https://github.com/organizations/{org}/repositories/new
2. Name: {repo-name}, Visibility: Private, Description: {desc}
3. Create repository

Then run:

cd {tier_dir}
git init
git add -A
git commit -m "Initial commit: {Company} knowledge base - {N} pages from {M} spaces"
git remote add origin https://github.com/{org}/{repo-name}.git
git branch -M main
git push -u origin main

Or offer to create the repos via browser automation if Chrome MCP tools are available.

Step 5.3: Access Reminder

For each restricted tier, remind the user: "Remember to add collaborators to {repo-name}: Go to https://github.com/{org}/{repo-name}/settings/access and invite the people who should have access to this tier."

Phase 6: Auto-Sync

Step 6.1: Create Scheduled Task

Use create_scheduled_task to set up weekly sync:

• taskId: {company-slug}-kb-weekly-sync
• cronExpression: "17 7 * * 1" (Monday 7:17 AM local)
• description: "Weekly Confluence sync for {Company} Knowledge Base repos"
• notifyOnCompletion: true

The task prompt should include:

1. The list of all tiers with: local path, GitHub remote URL, space keys and IDs
2. Instructions to pull pages using getPagesInConfluenceSpace MCP tool
3. Instructions to write new/updated pages to raw/
4. Instructions to run ckb compile --fast in each tier
5. Instructions to git add -A && git commit && git push for each tier with changes

Step 6.2: Suggest First Run

Tell the user: "I recommend running the sync task once now to pre-approve the tool permissions. This way future Monday runs won't pause for approval."

Offer to trigger it immediately.

MCP Tool Reference

Atlassian Tools (available via MCP connector)

Scheduled Tasks

Troubleshooting

"Atlassian MCP connector not connected" Direct user to Claude Code settings to connect their Atlassian account.

"No spaces found" Check if the user has Confluence access in their Atlassian account. Some accounts only have Jira access.

"Python not available" Instruct user to install Python 3.9+ from python.org or via their package manager.

"gh CLI not available" Provide manual GitHub repo creation instructions (Step 5.2b).

"pip install fails" Check Python version (3.9+ required). Try: python3 -m pip install --upgrade pip then retry.

"ckb compile fails" Ensure venv is activated: source venv/bin/activate. Check that pip install -e . succeeded.

"Rate limited during ingest" Add longer pauses between spaces. Process one space at a time.

"Pages have no content" Some Confluence pages are container/category pages with no body text. These are correctly skipped during ingest.