agentic-coding（Agent 编码）

Setup

If ~/agentic-coding/ does not exist or is empty, read setup.md, ask a concise kickoff question, and keep any persistence explicitly opt-in.

Positioning

This skill is intentionally different from agentic-engineering and vibe-coding:

• agentic-engineering focuses on multi-agent operating patterns and team throughput.
• vibe-coding focuses on prompt-led exploration and fast idea shipping.
• agentic-coding focuses on contract-first implementation, proof of fix, and reviewer-ready handoff.

When to Use

User needs merge-ready code from an AI agent with explicit quality gates. Use for production features, risky refactors, bug fixes with reproducible failures, and Xcode-centered work such as Swift feature delivery, iOS/macOS regressions, and release-branch hotfixes.

Architecture

Memory lives in ~/agentic-coding/. See memory-template.md for setup.

~/agentic-coding/
|- memory.md       # Persistent preferences and operating mode
|- contracts.md    # Accepted task contracts and non-goals
|- evidence.md     # Test evidence and verification snapshots
`- handoffs.md     # Delivery notes and rollback hints

Quick Reference

Load these files on demand to keep context focused and execution fast.

Core Rules

1. Lock a Contract Before Writing Code

Start every task with a compact contract:

• Objective: exact outcome in one sentence
• Acceptance: checks that prove success
• Non-goals: what must stay untouched
• Constraints: stack, style, limits, deadlines

No contract, no code.

2. Run the PACT Loop

Use the same execution loop every time:

1. Problem framing: restate objective and assumptions
2. Acceptance design: define checks before edits
3. Change set: produce the smallest useful diff
4. Trace and test: show evidence and residual risk

This skill is execution discipline, not brainstorming. For Xcode workflows, tie acceptance to a concrete target, simulator/device, and test command before editing.

3. Keep Diffs Surgical

One user objective maps to one focused change set:

• Prefer file-local edits over broad rewrites
• Separate behavior change from style cleanup
• Avoid hidden side effects outside declared scope

If scope grows, split into a second contract.

4. Prove Failure Then Prove Fix

For bugs and regressions:

• Capture the failing condition first (test, log, or reproduction)
• Apply minimal fix
• Re-run the same check to prove resolution

Never claim fixed without before and after evidence.

5. Deliver Handoff-Grade Output

End each cycle with a delivery packet:

• What changed and why
• Files touched and blast radius
• Validation run and results
• Known risks and rollback path

If handoff is unclear, the task is not finished.

6. Escalate With a Structured Fallback

When blocked after two failed attempts:

• Stop editing
• State what was tried
• Propose two grounded alternatives
• Request a decision with tradeoffs

Do not keep guessing in loops.

Common Traps

• Starting implementation without acceptance checks -> endless iteration and unclear done state.
• Asking the agent for full rewrites -> noisy diffs and avoidable regressions.
• Mixing feature work with architecture overhaul -> weak reviewability and hard rollback.
• Reporting success without reproducible evidence -> false confidence in production.
• Treating AI output as final draft -> quality debt moved to code review.

Security & Privacy

Data that leaves your machine:

• None from this skill itself

Data that stays local:

• Contracts, evidence notes, and handoff summaries in ~/agentic-coding/

This skill does NOT:

• Trigger undeclared network requests
• Access files outside its own memory path
• Write to global or platform memory stores
• Auto-approve risky code without explicit evidence

Scope

This skill ONLY:

• Improves execution quality of AI-assisted coding
• Enforces contract driven implementation and verification
• Produces clear handoff packets for reviewers

This skill NEVER:

• Replaces security review for high risk domains
• Encourages blind trust in generated code
• Overrides project specific contribution rules

Related Skills

Install with clawhub install <slug> if user confirms:

• agentic-engineering - Multi-agent collaboration and operating patterns.
• coding - General coding support across stacks and tasks.
• code - Broad code authoring and editing assistance.
• copilot - Companion style IDE assistance patterns.
• delegate - Structured task delegation to autonomous agents.

Feedback

• If useful: clawhub star agentic-coding
• Stay updated: clawhub sync