Turn Claude Code into a failsafe, token-efficient, one-shot coding machine. Production-tested patterns from Anthropic, Manus, Devin, and 100+ real-world projects. Updated March 2026 with latest breakthroughs from 10-agent parallel research.
Most developers use Claude Code at 20% of its potential. The top 1% know:
- Scaffold > Model — SWE-bench proves: changing the scaffold changes scores 22%. Changing the model changes them 1.3%. (Source)
- Context > Prompts — For every 1 output token, 166 input tokens are read. 10% context reduction saves more than eliminating all output. (Source)
- Hooks > Instructions — CLAUDE.md rules can be "forgotten." Shell hooks execute deterministically, always.
- TDD + Agents = Gold — Anthropic officially endorsed Red/Green TDD as the agentic coding pattern. (Source)
- 65% is the real limit — Context quality degrades suddenly (not gradually) past 65% capacity. Plan for 650K in a 1M window. (Source)
This repo gives you everything in a single bash setup.sh.
https://github.com/Supersynergy/awesome-agentic-coding — clone this, run setup.sh, confirm what's installed.
That's it. Claude clones the repo, runs the installer, and shows you what was set up.
git clone https://github.com/Supersynergy/awesome-agentic-coding.git ~/awesome-agentic-coding
cd ~/awesome-agentic-coding && bash setup.shClone https://github.com/Supersynergy/awesome-agentic-coding and run bash setup.sh — then detect this project's stack, create an optimized CLAUDE.md, add path-specific rules, run tests, and show me what's available.
What it installs: 5 hooks, 10 skills, 3 agents, 3 rules, optimized settings.json, CLAUDE.md template. What it preserves: Your existing config (backup + merge, never overwrite).
See
ONESHOT_SETUP_PROMPT.mdfor all 4 setup variants including a zero-clone option.
YOU
|
"Plan this. Don't code."
|
v
┌────────────────┐
│ ORCHESTRATOR │ Opus/Sonnet — plans only, never codes
└───────┬────────┘
│ spawns parallel
┌───────────┼───────────────┐
v v v
[Researcher] [Coder] [Reviewer]
Haiku Sonnet Haiku
Read-only Edit+Bash Read-only
$0.02/task $0.15/task $0.03/task
| | |
└───────────┼───────────────┘
v
┌──────────────────┐
│ QUALITY GATES │ <-- Hooks (deterministic, can't bypass)
│ TDD | Lint | Secrets │
└──────────────────┘
v
Production Code
Evidence: Orchestrator + Specialists beats single-agent by 90.2% (Anthropic). Scaffold matters 22x more than model choice (SWE-bench).
| # | Breakthrough | Impact | Source |
|---|---|---|---|
| 1 | Scaffold > Model | Models within 1.3pts on SWE-bench. Scaffold = 22pt difference | Morph LLM |
| 2 | 65% context limit | Quality collapses suddenly past 65%, not gradually at 95% | Chroma Research |
| 3 | Prompt caching GA | 90% savings on cached reads, automatic since Feb 2026 | Anthropic |
| 4 | Tool Search | 46.9% MCP token reduction via deferred tool loading | Medium |
| 5 | TDD + Agents | Red/Green TDD officially endorsed by Anthropic | Simon Willison |
| 6 | 1:166 ratio | Every 1 output token costs 166 input tokens of context reads | Dev.to |
| 7 | Memory tiering | Letta/Mem0 = 26% accuracy gain, 90% token reduction | Mem0 |
| 8 | Batch API | 50% discount on all tokens for async processing | Anthropic |
| 9 | Self-healing CI | Standardized pattern: failures trigger repair agents | Medium |
| 10 | Inference cost 1000x drop | $20/MTok (2022) to $0.40/MTok (2026) | Introl |
| 11 | /btw command |
Side questions at ~0 marginal cost (reuses parent cache) | Claude Code Docs |
| 12 | opusplan alias |
Opus planning + Sonnet execution = 40% cheaper | Claude Code Docs |
| 13 | PreToolUse input modification | Hooks can now modify tool inputs, not just block | Claude Code Hooks Guide |
| 14 | Memory timestamps | Stale memories (>7 days) now flagged by /context |
Claude Code Changelog |
| Hook | Event | What It Does |
|---|---|---|
quality-gate.sh |
Stop | Auto-detects project type, runs tests before Claude "finishes" |
secret-guard.sh |
PreToolUse | Blocks writes to .env, .pem, credentials, warns on API keys in content |
context-inject.sh |
SessionStart | Injects git branch, recent commits, project type detection |
protect-prod.sh |
PreToolUse | Blocks edits to production configs |
syntax-check.sh |
PostToolUse | Validates Python/JSON/YAML syntax after every edit |
| Skill | Description |
|---|---|
/oneshot <task> |
Context-primed 1-shot coding (4-phase protocol) |
/review [target] |
Multi-agent code review (3 parallel Haiku reviewers) |
/debug <bug> |
Scientific debugging with hypothesis testing |
/refactor <target> |
Safe refactoring with test verification at each step |
/plan <feature> |
Create implementation plan before coding (research → plan → review) |
/test <target> |
Auto-detect framework, write tests following TDD principles |
/commit |
Smart commit with auto-generated message from staged changes |
/github <cmd> |
GitHub via gh CLI — replaces GitHub MCP (saves 55K tokens) |
/docs <library> |
Fetch live docs via WebFetch — replaces Context7 MCP (saves 5-8K tokens) |
/db <query> |
Database via CLI — replaces Postgres/SurrealDB MCP (saves 3-8K tokens) |
/search-code <what> |
Deep codebase search with parallel agents |
| Agent | Model | Tools | Cost/Task |
|---|---|---|---|
researcher |
Haiku | Read, Grep, Glob | $0.02 |
reviewer |
Haiku | Read, Grep, Glob, Bash | $0.03 |
architect |
Sonnet | Read, Grep, Glob | $0.10 |
| Rule | Paths | Focus |
|---|---|---|
security.md |
**/* |
OWASP Top 10, input validation, no hardcoded secrets |
api.md |
src/api/**, routes/** |
Validation, error format, pagination, rate limiting |
tests.md |
**/*test*, **/*spec* |
Behavior testing, independence, happy + error paths |
| Guide | What You'll Learn |
|---|---|
CONTEXT_ENGINEERING.md |
The 6 principles + annotation cycle (90% quality improvement) |
CONTEXT_TOOLS.md |
Context7 MCP, Greptile, Repomix, 1M window strategies |
TOKEN_ECONOMICS.md |
Cost per task, 7 optimization levers, batch API, prompt caching |
PATTERNS.md |
Architecture patterns ranked (Orchestrator: 59/70, Blackboard: 41/70) |
TOP_50_TOOLS.md |
Top 50 dev tools for 2026, ranked by impact |
NATS_GUIDE.md |
NATS.io for AI agents: pub/sub, JetStream, KV, code examples |
DATABASES.md |
SurrealDB, Turso, EdgeDB, Dragonfly comparisons |
UI_AND_CRUD.md |
shadcn alternatives, Server Actions, tRPC, optimistic updates |
ZERO_MCP.md |
Zero-MCP workflow: 97% token savings — replace every MCP with CLI + skills |
WHY.md |
Evidence and sources behind every decision |
| Guide | What's Inside |
|---|---|
CRUD_CHEATSHEET.md |
Every CRUD pattern: Server Actions, tRPC, SurrealDB, Hono REST, NATS events |
AGENT_PATTERNS.md |
5 agent architectures, model routing, Claude Code agent best practices |
MODERN_STACK.md |
6 stack recipes: SaaS, AI agents, real-time, e-commerce, API-first, admin |
Copy-paste into Claude Code. 3 variants: full (clone + install), quick (no clone), project-specific.
Run bash validate.sh to check all hooks, skills, agents, settings are properly installed.
| File | Purpose |
|---|---|
CLAUDE.md.template |
Optimized starter (copy, customize, ship) |
settings.json.optimal |
Full settings with 30+ pre-approved commands |
oneshot-prompts.md |
10 proven one-shot prompt templates + meta-prompt |
mcp-config.json |
MCP server configs: Context7, GitHub, Postgres, SurrealDB |
agent-team.md |
Agent Teams examples: full-stack, debugging, multi-service |
nats-surrealdb-starter.md |
Event-sourced CRUD with NATS + SurrealDB + Hono in 50 lines |
"Getting to 80% is fast. Getting to 95%+ takes discipline." — Addy Osmani
The secret isn't the prompt. It's the context scaffold:
1. CLAUDE.md loaded → Claude knows your patterns
2. Rules auto-applied → Security, API, test standards enforced
3. Context injection hook → Git state, project type detected
4. /oneshot "add X" → 4-phase protocol: gather → plan → implement → verify
5. Quality gate hook → Tests run before "done"
You: "Plan this feature. Don't code."
Claude: Returns plan.md
You: Add inline notes → "@claude: use Zod here", "@claude: skip auth for MVP"
You: "Address all notes. Still don't code."
Claude: Refined plan with 0 ambiguity
You: "Implement."
Claude: Done in 1-2 turns. Every decision pre-made.
Why: 90% of iterations come from ambiguity, not capability.
| Metric | Value | Implication |
|---|---|---|
| Context read ratio | 1:166 | 10% context reduction > eliminating all output |
| Degradation threshold | 65% | Quality collapses suddenly past this point |
| CLAUDE.md adherence | 95% under 200 lines, 40% over 500 | Keep it lean |
| Cache savings | 90% | Cached reads cost $0.30/MTok vs $3.00/MTok |
| Lost-in-middle | Still unsolved | Critical info at start or end, never middle |
- Scaffold > Model — Invest in your agent architecture, not model shopping
- Cache-friendly prompts — Stable prefixes, no timestamps in system prompts
- Append-only context — Never mutate; let model learn from failures in-context
- Preserve errors — Wrong turns teach self-correction better than instructions
- External memory — Large docs on filesystem, referenced by path (not embedded)
- Progressive disclosure — Teach Claude how to find info, not dump everything
- Pointers over copies —
file:linereferences, not pasted code that goes stale - Compact at 65% — Don't wait for 95% auto-compact. Quality is already degraded.
| Task | Model | Cost |
|---|---|---|
| Simple edit | Sonnet | $0.05 |
| Standard feature | Sonnet | $0.15 |
| Complex architecture | Opus | $0.50 |
| Codebase exploration | Haiku subagent | $0.02 |
| Code review | Haiku subagent | $0.03 |
| Batch refactoring (50% off) | Batch API | $0.08/file |
| Lever | Savings | How |
|---|---|---|
| Model routing | 5x on exploration | Haiku for read-only, Sonnet for code |
| Permission allowlists | 500-1000 tok/session | Pre-approve safe commands |
| CLAUDE.md < 200 lines | ~1000 tok/session | Details → skills + topic files |
| Skills (on-demand) | 0 tokens when unused | Load only on /invoke |
| Prompt caching | 90% on repeats | Automatic since Feb 2026 |
| Tool Search | 46.9% MCP reduction | Deferred tool loading |
/compact at 65% |
Prevents quality collapse | Not 80%, not 95% — 65% |
| Batch API | 50% off | Async processing for bulk ops |
Optimized: ~$0.40/session. Unoptimized: ~$2.50/session. Savings: 84%.
Layer 1: CLAUDE.md + Rules → Persistent rules (always loaded, <200 lines)
Layer 2: Skills → On-demand workflows (0 tokens when unused)
Layer 3: Hooks → Deterministic guards (CANNOT be bypassed)
Layer 4: Subagents → Isolated specialists (failures don't crash main)
Layer 5: Memory + TDD → Cross-session learning + test-driven verification
Layer 6: Checkpoints → /resume, /rewind, session persistence
Only Layer 3 is deterministic. Everything else relies on LLM compliance. For anything that MUST happen (tests, security), use hooks.
# Context management (the #1 lever)
/clear # Reset between unrelated tasks
/compact focus on X # Manual compact at 65% (not 95%)
/context # Token breakdown — find waste
/btw question # Side question (0 history pollution)
# Effort routing
/effort low # Simple edits ($0.05)
/effort medium # Standard tasks ($0.15) [default]
/effort high # Complex reasoning ($0.50)
# Skills (this repo)
/oneshot task # 1-shot with 4-phase protocol
/review # Multi-agent code review
/debug issue # Scientific debugging
/refactor target # Safe refactoring with tests
# Session management
/resume name # Continue named session
/rewind # Roll back to earlier state
/loop 5m prompt # Repeat prompt on intervalAnthropic's official recommendation (Agentic Coding Trends Report 2026):
1. Write failing test first (Red)
2. Ask Claude to make it pass (Green)
3. Review + refactor
4. Repeat
Why: Without TDD, agents can "cheat" — writing tests that confirm broken behavior. TDD defines correctness before implementation. The test is the spec.
| Source | Key Insight |
|---|---|
| Anthropic: Multi-Agent Research | Orchestrator + Specialists = 90.2% improvement |
| Morph LLM: SWE-bench Analysis | Scaffold matters 22x more than model choice |
| Chroma: Context Rot Research | Quality degrades suddenly at 65%, not gradually |
| Manus: Context Engineering | KV-cache hit rate is the #1 cost metric |
| Anthropic: Context Engineering | Context > Prompt crafting |
| Anthropic: Agentic Coding Trends | TDD officially endorsed for agents |
| Simon Willison: Red/Green TDD | TDD + agents = the golden pattern |
| Addy Osmani: The 80% Problem | 80% is fast; 95%+ needs discipline |
| Dev.to: 70% Token Cost Reduction | 1:166 input/output ratio |
| MCP Tool Search | 46.9% token reduction via deferred loading |
| Claude Code Docs: Best Practices | CLAUDE.md < 200 lines = 95% adherence |
| Claude Code Docs: Hooks | Deterministic enforcement > LLM instructions |
| Anthropic: Prompt Caching | 90% savings on cached reads |
| Anthropic: Batch API | 50% off for async processing |
| Azure: AI Agent Patterns | Orchestrator pattern dominates production |
- awesome-claude-code — Skills, hooks, agents, applications
- awesome-claude-code-subagents — 100+ subagents
- awesome-claude-md — Exemplary CLAUDE.md files
- awesome-cursorrules — 500+ Cursor rules
- claude-code-ultimate-guide — Comprehensive guide
- agentic-coding-handbook — Tweag's handbook
- agent-flywheel — The agentic coding flywheel concept
PRs welcome! Include evidence (benchmark, source, or production experience) with every claim.
MIT — Use it, fork it, make it yours.
Built by @Supersynergy | Powered by Claude Code | Research by 10 parallel Haiku agents