claude-core
Foundation plugin - planning, brainstorming, verification, tracing, memory governance, debugging, learning pipelines, and workplace integrations.
claude-core
The foundation of the hjemmesidekongen/ai methodology - the infrastructure that makes structured AI-assisted development possible. Without it, the other plugins have nothing to stand on: no state management, no verification protocol, no learning pipeline.
Version: 0.3.0 · 40 skills · 13 commands · 12 agents · 23 hooks
What It Does
claude-core turns Claude Code from a stateless Q&A tool into a stateful development environment. It remembers what you decided, tracks what you're working on, verifies that work is actually done, and learns from how you use it.
Commands
| Command | What It Does |
|---|---|
/brainstorm:start | Open-ended exploration. Claude acts as a sparring partner, challenges assumptions, pushes back on weak reasoning. |
/brainstorm:decide | Extract structured decisions from a brainstorm session. Each decision gets user confirmation before it's saved. |
/plan:create | Convert a task into a wave-based plan. Resolves dependencies via topological sort, assigns model tiers, sets up verification gates. |
/plan:execute | Run the plan wave-by-wave with parallel agent dispatch and verification between waves. |
/plan:status | See where you are - wave progress, task status, errors, recovery options. |
/plan:resume | Pick up an interrupted plan from exactly where it stopped. |
/full-review | 5-phase code review with 2 parallel reviewers per phase. Severity gates between phases. |
/autopilot-run | Autonomous iteration loop. Give it a task, it works through iterations until completion, respecting stop hooks and verification gates. |
/autopilot-cancel | Stop the loop and get an iteration report. |
Key Concepts
Verification & Safety
No task gets marked "done" without proof.
verification-gate enforces a 5-step protocol: identify the proof command, run it, read the full output, verify it matches expectations, and only then claim completion. Inference-based completion ("it should work") is explicitly banned.
plan-verifier runs two stages after each wave: Stage 1 checks mechanical spec compliance (files exist, schemas valid, ownership correct). Stage 2 runs quality review with judgment calls. Both must pass.
Hooks enforce this at the system level:
- verification-gate-stop prevents premature completion claims
- check-wave-complete blocks plan advancement without verification
- scope-guard prevents modifications to out-of-scope files
Cross-Session Continuity
Claude Code forgets everything when you start a new session. claude-core doesn't.
session-handoff creates structured handoff documents capturing: current state, blockers, open questions, and actionable next steps. When you resume, it loads the handoff, classifies staleness (FRESH, SLIGHTLY_STALE, STALE, VERY_STALE), and verifies assumptions still hold.
PreCompact snapshot - before Claude Code compresses context, a hook saves the full session state to .ai/context/snapshot.yml. On session start, another hook restores it. Your working context survives compaction automatically.
Learning Pipeline
claude-core observes how you work and extracts patterns:
- observation-recorder (hook) - captures tool call patterns automatically during every session
- instinct-extractor - reads observations, identifies recurring patterns (3+ occurrences), extracts them as "instincts" with confidence weights
- instinct-evolve - reviews mature instincts (confidence ≥0.75, evidence ≥3) and proposes promotion to CLAUDE.md rules, MEMORY.md entries, or new skills
The pipeline is self-correcting: contradictions decay confidence, and instincts that don't recur expire naturally.
Debugging & Investigation
Three skills that enforce evidence-first debugging:
root-cause-debugging - 4 mandatory phases: Investigate, Pattern Analysis, Hypothesis, Implement. You can't jump to a fix without understanding the problem first. If 3+ fixes fail, it escalates.
hypothesis-generator - generates 3 competing hypotheses, investigates each in parallel, and synthesizes evidence. Eliminates confirmation bias by forcing you to consider alternatives before committing to one explanation.
reasoning-trace-optimizer - analyzes trace-light.log for reasoning degradation: context loss, tool confusion, instruction drift, goal abandonment, circular reasoning. Produces specific diagnosis with actionable fixes.
Component Reference
Planning & Execution Skills
The planning engine handles dependency ordering, file-ownership isolation (so parallel agents don't conflict), and automatic verification gates between waves.
| Skill | Purpose |
|---|---|
| brainstorm-session | Open-ended brainstorm |
| brainstorm-decision-writer | Extract decisions from brainstorms |
| brainstorm-decision-reader | Load past decisions for context |
| plan-engine | Task to wave plan conversion |
| plan-verifier | Two-stage wave verification |
Code Review Skills
/full-review runs a 5-phase review with 2 parallel reviewers per phase:
- Code quality & best practices
- Architecture & design patterns
- Security vulnerabilities
- Performance optimization
- Testing coverage & documentation
Severity gates between phases - critical findings in Phase 1 block progression to Phase 2. Final consolidated report deduplicates across all reviewers.
parallel-reviewer dispatches 4 specialized reviewers (security, performance, architecture, testing) simultaneously for faster targeted review.
Component Tooling Skills
Create, review, and maintain plugin components:
| Create | Review |
|---|---|
| hook-creator | hook-reviewer |
| skill-creator | skill-reviewer |
| command-creator | agent-reviewer |
| agent-creator | plugin-reviewer |
| plugin-creator | component-reviewer |
| mcp-creator | - |
Creators enforce correct frontmatter, naming conventions, and registration. Reviewers are read-only - they produce verdicts, never modifications.
Workplace Integrations
Five MCP-based skills for connecting to your existing tools:
| Skill | Service | What You Can Do |
|---|---|---|
| jira-basics | Jira | Read tickets, search with JQL, transition status, add comments |
| confluence-basics | Confluence | Search pages, read content, navigate space hierarchies |
| slack-basics | Slack | Read channels, search messages, post updates, reply to threads |
| bitbucket-basics | Bitbucket | Create PRs, check pipeline status, manage branches |
| outlook-basics | Microsoft 365 | Search emails, read threads, check calendar |
Each integration requires its MCP server configured. All degrade gracefully when the server isn't available - clear error messages, never fabricated data.
Architecture & Documentation Skills
| Skill | Purpose |
|---|---|
| c4-architecture | Generate C4 model diagrams (context, container, component, deployment) as Mermaid |
| mermaid-diagrams | 7 diagram types: class, sequence, flowchart, ERD, state, git graph, gantt |
| doc-checkpoint | After any task, compares changes against documentation files and produces a checklist of what needs updating |
| auto-doc | Automatically syncs CLAUDE.md structure trees, component counts, and ecosystem.json entries |
| reducing-entropy | Deletion-first mindset for codebase health: measure success by how much code remains |
Writing & Communication Skills
| Skill | Purpose |
|---|---|
| writing-clearly-and-concisely | Strunk's 18 rules applied to technical writing, with AI anti-pattern detection |
| prompt-optimizer | Sharpen vague prompts into precise instructions. Two modes: auto-sharpen (hook-triggered) and builder (/prompt:create) |
Hooks
claude-core runs 23 hooks across 6 event types:
| Event | Count | Purpose |
|---|---|---|
| PreToolUse | 5 | Prevent direct pushes, validate scope, gate TDD/planning, detect compaction |
| PostToolUse | 6 | Debug output, protect CLAUDE.md, trace operations, record observations |
| PreCompact | 1 | Snapshot session state before context compression |
| SessionStart | 2 | Restore session state, check memory health |
| Stop | 8 | Verify completion, check documentation, clear caches |
| UserPromptSubmit | 1 | Grade prompt quality |
Hooks are the safety net. They prevent you from pushing to main, skipping verification, modifying out-of-scope files, or claiming work is done when it isn't.
Agents
12 specialized agents, each with a specific role and model tier:
| Agent | Tier | Role |
|---|---|---|
| plugin-validator | sonnet | Full plugin structure validation |
| skill-auditor | sonnet | 8-dimension quality scoring (120-point scale) |
| security-auditor | sonnet | Infrastructure security: injection, secrets, unsafe eval |
| component-reviewer | sonnet | Pre-commit validation for new components |
| architect-reviewer | opus | Plugin boundary and coupling analysis |
| refactoring-specialist | sonnet | Behavior-preserving code transformations |
| knowledge-synthesizer | sonnet | Cross-session pattern mining |
| tdd-orchestrator | sonnet | Red-green-refactor cycle enforcement |
| context-manager | sonnet | Token-budgeted context assembly for multi-agent dispatch |
| error-detective | sonnet | Read-only error investigation and root cause analysis |
| incident-responder | sonnet | Post-investigation fix coordination and post-mortems |
| agent-reviewer | sonnet | 5-dimension agent quality review |
All review/audit agents are read-only. They produce verdicts. They never modify code.
See also
- dev-engine - the multi-agent development pipeline that builds on claude-core
- taskflow - task management that uses claude-core's MCP and session infrastructure
- Architecture - how the three plugins relate and share data