CLAUDE.md

Super Skill - Claude Code Operating Instructions

Session Start Protocol

Every session starts with two files only:

1. SUMMARY.md -- domain snapshot, 80 lines max, start here
2. PENDING.md -- open items awaiting owner approval

Read both before anything else. Do not load additional files unless the specific task requires it. See the Load Protocol table in SUMMARY.md.

How to Find Available Prompts

All activation and operational prompts live in ONBOARDING.md. Read ONBOARDING.md to find the correct prompt for any task. Do not ask the user to copy prompts manually. Run the appropriate prompt yourself based on the task requested.

Available prompts: Prompt 0: Set up a new Super Skill Prompt 1: Activate domain layers and Expert Council Prompt 2: Create or update the global router Prompt 3: Evaluate something new Prompt 4: Generate a learning module Prompt 5: reserved for future use Prompt 6: Approve a PENDING item Prompt 7: Run an Expert Debate session

Slash commands for ongoing use: /ss-eval, /ss-graph, /ss-learn, /ss-pending, /ss-query, /ss-sync, /ss-summary, /ss-drift, /ss-council, /ss-expert, /ss-synthesize, /ss-dashboard

When to Load Full Files

Never load a file the current task does not require. Individual expert profiles load on demand only, not by default.

Graph Navigation Protocol

Before loading any layer file, check wiki/graph.json if it exists. It contains:

• layer_index: which file answers which type of question
• hotspots: which files have the most active changes right now
• navigation_hints: pre-computed answers for common question types
• open_pending count: how many items need the owner's attention

Use graph.json as a routing layer, not as a replacement for the actual files. graph.json tells you WHERE to look. The layer files tell you WHAT is there.

If graph.json does not exist yet: suggest the user run /ss-graph to generate it.

Operating Rules

• Never modify a layer file without explicit approval from the owner
• All proposals go to PENDING.md -- never directly to layer files
• If a task affects multiple sub-domains, note all impacts in PENDING.md
• If asked about something not in the current state, say so honestly
• Generate learning content on request -- write output to notebooks/ folder

What Requires Owner Approval

• Any change to CURRENT_STATE.md
• Any new entry in DECISIONS.md
• Any adoption or rejection in EVALUATION.md
• Any update to DOMAIN_MAP.md

What Does Not Require Approval

• Reading and explaining existing content
• Generating learning modules to notebooks/
• Writing proposals to PENDING.md
• Running scripts (feed, generate, check, update_summary)

Keeping SUMMARY.md Current

After any approved change to any layer file, run:

python scripts/update_summary.py

Or let super-skill-sync handle it automatically on the next pull.

Expert Perspectives Protocol

When returning answers on architecture, evaluation, or decisions:

1. Check if experts in experts/COUNCIL.md hold conflicting positions on this topic.
2. If yes: append an Expert Perspectives block after your answer.
3. If consensus or the topic is routine: no block needed.
4. Active debates in COUNCIL.md always surface when their topic is touched.

Expert Perspectives block format:

Expert Perspectives: [Expert Name, Layer N]: [one sentence -- agrees / disagrees / challenges] [Expert Name, Layer N]: [one sentence]

Maximum 3 experts per block. Maximum 1 line per expert. Do not mention experts on routine tasks or settled decisions with no friction.

Channels Audit Protocol

When an approval, rejection, or deferral action is performed via Claude Code Channels (Telegram or Discord), write the following entry to LOG.md immediately after the action completes:

LOG-NNN: [item title]

Date: [YYYY-MM-DD] Action: [approved | rejected | deferred] Item: [PENDING item ID or description] Source: channels Session: [tmux session name if available] Files changed: [list of affected files]

The source: channels field distinguishes mobile actions from desktop actions in the audit trail. The Iron Principle applies identically to Channels actions: the model proposes, the user decides via Telegram, the Super Skill records with full provenance.

Iron Principle

The model proposes. The user decides. The Super Skill records. The system executes.