Claude Commands LibraryTransform Ideas into Precise Prompts

The Problem ​

You write a prompt. Claude guesses what you meant. You get something close but not quite right. You clarify. Claude tries again. Repeat.

This wastes time and tokens.

The Solution ​

Claude Commands Library ensures every prompt is perfect before execution:

# Before: Vague prompt leads to wrong implementation
"Add login to my app"

# After: Phase 0 asks what you actually need
/prompt-technical Add login to my app

# Claude asks:
# - Which authentication method? (JWT / OAuth / Session)
# - Which database stores users?
# - What UI framework for the login form?
# - Should I add password reset functionality?

# Result: Precise implementation matching your exact needs

Quick Start ​

1. Install ​

# Download and run installer
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/Tadzesi/claude-ideas/main/install-claude-commands.ps1" -OutFile "install.ps1"
.\install.ps1

2. Use Commands ​

# Perfect any prompt (2 seconds)
/prompt Fix the authentication bug

# Technical analysis with agent assistance (5-30 seconds)
/prompt-technical Implement caching with Redis

# Deep multi-agent research (1-3 minutes)
/prompt-research Perform security audit of payment system

# Save your work
/session-end

3. Track Progress ​

Commands show real-time progress as they run.

Command Overview ​

| /prompt-article | Article writing wizard | Interactive | Blog posts, documentation | | /session-start | Load previous context | ~2s | Beginning of session | | /session-end | Save current context | ~5s | End of session | | /reflect | Improve skills from feedback | 5-15s | After using a command |

How It Works ​

Phase 0: Prompt Perfection ​

Every command runs your prompt through validation:

1. Recall - Load known facts from project profile
2. Detect - Language, type, and intent
3. Check - 6 completeness criteria, pre-filling from memory
4. Ask - Only truly unknown information
5. Structure - Transform into executable format
6. Approve - Wait for your confirmation

Hybrid Intelligence ​

Tasks are scored for complexity (0-50+):

• Simple (0-4): Fast inline validation
• Moderate (5-9): Optional agent assistance
• Complex (10-19): Automatic agent exploration
• Research (20+): Multi-agent deep analysis

Continuous Learning ​

The system tracks:

• Successful transformations
• Your modification patterns
• Common missing information
• Agent effectiveness

After 3+ occurrences, it suggests smart defaults.

What's New ​

v5.2 - Real Anthropic Subagents ​

Multi-agent research now uses real Claude Code subagents instead of prose simulation. Each specialist lives in .claude/agents/ with valid YAML frontmatter (name, description, tools, model, color) and runs in its own isolated context window.

5 specialist subagents:

• research-explore (haiku, blue) — codebase discovery, architecture mapping
• research-pattern (haiku, green) — convention detection
• research-security (sonnet, red) — OWASP Top 10, vulnerability scanning
• research-performance (sonnet, orange) — N+1, async, caching audit
• research-citation (haiku, purple) — file:line evidence on every finding

Orchestration in main thread. The prompt-research/SKILL.md orchestrator spawns subagents via the Task tool (single message, multiple parallel invocations), runs gap analysis, and persists citation entries — because subagents cannot spawn other subagents (Anthropic limit).

Net diff: -1 172 LOC. Removed 5 prose-only library/research-agent-*.md files (~6 000 words) and config/agent-roles.json (408 lines, mostly aspirational fields). Debloated research-adapter.md from 2 411 → 627 words.

Tools narrowed: Bash removed from explore + security allowlists (read-only research is fully served by Read+Grep+Glob; attack surface narrowed).

See full changelog →

v5.1 - Architecture Cleanup ​

Architecture cleanup pass driven by a documented diagnostic. Net diff: -2 856 LOC.

• prompt-research/SKILL.md rewritten following Anthropic's progressive disclosure: 902 → 185 lines (-79%).
• scripts/sync-version.ps1 propagates package.json version to installer + CLAUDE.md (modes: -Check, -DryRun, default writes).
• Installer -RebaselineMemory flag, version-transition announcement, obsolete-subdir auto-detection.
• Standalone Interaction Protocol stub in .claude/CLAUDE.md.

v5.0 - Session Memory + Honest Portfolio ​

Repo reduced from 11 commands to the three that are actually used: /prompt, /prompt-article-readme, /prompt-research. Everything else removed.

Session Memory added as a workflow layer:

• PreCompact hook — writes a diary entry to .claude/memory/diary/ before every compaction. Captures task summary, design decisions (with WHY), challenges, solutions, and observed preferences. Per-project; never committed.
• /reflect-diary — reads accumulated diary entries, identifies recurring patterns, and proposes updates to project-profile.md. Nothing applied without explicit approval.
• Installer — install-claude-commands.ps1 now deploys the hook to ~/.claude/hooks/ and registers it in ~/.claude/settings.json on any machine.

See full changelog →

v4.9 - Opus 4.7 Optimisation ​

Token + context savings targeted at claude-opus-4-7. Seven additive changes, all backward compatible.

Highlights:

• Prompt caching strategy — cache_control: ephemeral breakpoints on stable library files (core-library, model-router, exec-plan-template, model-tiers, adapters). Expected ~90% input cost reduction on warm cache hits. 5m default TTL, 1h beta TTL via extended-cache-ttl-2025-04-11.
• Fast Path in Phase 0 — trivial prompts (complexity score < 5, single file, no security flags) short-circuit to Steps 0.1 → 0.5 → 0.6. Saves ~40% Phase 0 tokens on simple tasks.
• Model tier split — opus is now opus-fast (Opus 4.6, 200K ctx, 4K thinking, interactive) vs opus-smart (Opus 4.7, 1M ctx beta, 8K thinking, interleaved thinking beta). Router picks based on depth-vs-latency tradeoff.
• Per-tier thinking budget — thinking_budget_tokens in model-tiers.json: haiku 0 / sonnet 2K / opus-fast 4K / opus-smart 8K.
• Context-editing adapter — applies context-management-2025-06-27 with clear_tool_uses_20250919 to prevent 200K context exhaustion during /prompt-research iterations.
• Memory-tool adapter — bridge to native memory_20250818 with a Phase A/B/C/D progressive migration plan.
• Consolidated CHANGELOG-skills.md — Version History deduped across all seven skills (~5K tokens saved per skill load).
• Batch API hint in /reflect for 50% cost reduction on non-urgent SDK-driven runs.

See full changelog entry →

v4.8 - Interaction Protocol ​

Plan-first discipline promoted from individual /prompt* skills to a global session protocol loaded from CLAUDE.md. Applies to every interaction, not only slash commands.

Four rules:

• Language — SK in / EN internal / SK out; technical terms verbatim
• Plan-First — summarise understanding, present 2-3 options for non-trivial work, emit an execution plan, wait for explicit approval
• Proactive Option-Finding — Claude proposes better paths BEFORE executing; names tradeoffs, recommends, user decides
• Never Auto-Execute — no git commit, git push, npm install, installer runs without explicit consent

Why it matters: previously, normal conversation bypassed Phase 0 discipline and auto-executed. v4.8 fixes the root cause — CLAUDE.md is loaded into every session automatically.

Read the full protocol →

v4.7 - Dynamic Model Routing + Smarter Phase 0 ​

Three major additions to Phase 0 that make every prompt cheaper to run and clearer to approve:

Curiosity Gate (Step 0.25) — confidence score 0-100 with a mandatory assumption ledger whenever confidence drops below 100%. No more silent guessing:

ASSUMPTIONS I AM MAKING (correct me if wrong)
- Block goes BEFORE Step 0.6, not after.
- Minor version bump (v4.1) is appropriate.

Confidence: 85%

Options-First (Step 0.35) — 2-3 alternatives shown BY DEFAULT for Task/Feature/Bug Fix/Refactor/Config prompts. Each option carries a model tier (haiku/sonnet/opus) and token-cost estimate so you choose knowing the tradeoff.

Execution Plan + MODEL HINT (Step 0.55) — mandatory block before the approval gate. Lists verified files, numbered steps, tools, risks, verification, effort, and assumptions. Plus one line:

MODEL HINT: sonnet (claude-sonnet-4-6) — multi-step single-file edit.
Savings vs current: ~80% (switch from opus).

Approval gate adds a new response: switch [haiku|sonnet|opus] — change model before approving to cut cost.

Estimated token savings: 30-45% through Haiku default for simple tasks (memory recall, single-file edits, deploy scaffolds, README generation).

v4.6 - Superprompting: Anti-Hallucination Contract ​

Every skill now has a built-in contract that prevents Claude from inventing facts:

HARD-GATE blocks — pre-flight checklist in every skill before any output:

- [ ] project-profile.md read this session
- [ ] No version numbers copied from templates — all from read files
- [ ] No file paths invented — all verified with Read or Glob

NEVER rules — explicit domain-specific prohibitions (e.g. /prompt-dotnet will never state a NuGet version without reading .csproj).

Chain-of-Thought REASONING block — Step 0.1 now requires:

REASONING
Prompt type: Task (Feature) — because input contains "add" + specific component name
Facts from project-profile.md: Stack is Node.js/VitePress, kebab-case conventions
Cannot determine: which specific file to edit

Core library v2.0 with Mermaid flowchart and Grounding Protocol.

v4.5 - Universal Skills ​

• /session-start and /session-end removed — replaced by Claude Code auto-memory
• /deploy and /new-stack are now universal — all server config read from personal-profile.md
• /prompt-dotnet and /prompt-react scan fallbacks added (monorepo support)
• Very High complexity threshold corrected: 20+ (was 15+)

v4.4 - .NET + React Project-Aware Skills ​

Two new project-aware skills that scan your codebase before asking any questions:

/prompt-dotnet - Reads .csproj, Program.cs, appsettings.json, Docker files:

PROJECT SCAN COMPLETE
Framework: net10.0
Architecture: Minimal API
Auth: JWT
ORM: EF Core | DB: PostgreSQL — ConnectionStrings:Default
Docker: yes — stack: myapp
Packages: FluentValidation, Serilog, MediatR

/prompt-react - Reads package.json, vite.config.ts, tsconfig.json:

PROJECT SCAN COMPLETE
React: 19.0.0 | TypeScript: yes, strict: yes
Router: React Router v6
State: Zustand
Data fetching: TanStack Query
Base path: /appname/

Both skills apply best practices for the detected stack automatically — no repeated questions.

v4.3 - Skills Format & Project-Aware Commands ​

Skills Format - Commands now use the native Claude Code Skills format:

---
name: prompt
description: Transform any prompt into an unambiguous, executable format...
argument-hint: "[your prompt or task description]"
---

Claude automatically suggests the right skill based on description. No more manual /prompt when Claude already knows what you need.

Project-Aware Startup - Every command loads project context first:

CONTEXT LOADED FROM PROJECT PROFILE
Project: My App v2.1 | Stack: Node.js + PostgreSQL
Platform: macOS | Branch: feature/auth
Recent work: Added JWT middleware (yesterday)

What else do you need for this task?

Zero Repeated Questions - The populated project-profile.md eliminates asking the same questions across sessions. Tech stack, preferences, structure - all pre-filled automatically.

v4.2 - Memory Recall & AI Fluency ​

Memory Recall - Commands remember your project between sessions:

Session 1: "What framework?" → "Express + TypeScript"
Session 2: Context: Express + TypeScript (from project profile) ✓

AI Fluency - Full 4Ds Framework integration:

• Delegation - Quick Delegation Check in all commands
• Description - 9 criteria completeness check
• Discernment - Post-Execution Evaluation with feedback loop
• Diligence - Responsibility reminders in Approval Gate

v4.1 - Skill Reflection ​

/reflect

• Signal detection (corrections, successes, edge cases)
• Priority-coded change proposals (HIGH/MED/LOW)
• Direct skill file modifications with approval

v4.0 - Predictive Intelligence ​

• Journey stage detection (exploring → implementing → debugging)
• Proactive warnings before you make mistakes
• Domain risk analysis (security, payment, performance)

v3.0 - Multi-Agent Research ​

• 2-5 specialized agents in parallel
• Iterative refinement with gap detection
• Comprehensive reports with file:line citations

Ready to Start? ​

Install Now → | Browse Commands → | Understand Architecture →