What Is the coding Skill?

There's a skill called coding in the SFD workspace that many people install but aren't sure how to use. Simply put: it's a software development workflow guide — it tells an AI Agent how to plan, implement, and verify code to production-grade standards, not just bang out a quick script.

At SFD Lab, this skill is the core equipment for 小章鱼🐙. Every time it picks up a PRD from 小浣熊🦝, it follows the coding skill flow: understand requirements → plan modules → implement step by step → self-test → deliver a change manifest.

Installation and Setup

# Install from ClawHub
clawhub install coding

Or directly in OpenClaw
/skill install coding

After installation, the skill file loads into your workspace at skills/code/SKILL.md. OpenClaw auto-injects it into context at session start.

Real-World Workflow: From Requirement to Running Code

Here's a real task we completed using the coding skill: adding a reading progress bar to smallfiredragon.com article pages.

Step 1: Requirements Understanding
Have the Agent restate the requirements first to confirm understanding. Many people skip this, but it prevents writing in the wrong direction entirely.

Requirement: Add a reading progress bar at the top of article detail pages,
updating as the user scrolls, in brand orange (#FF6B35).
Tech stack: Nuxt 3 SSR, components in components/article/ directory.

Step 2: Planning
The coding skill requires the Agent to output an execution plan before writing any code — which files to modify, what functionality might be affected, potential risks.

Step 3: Implementation
Modify files according to plan, with explanations for each file. Critical rule: don't cross plan boundaries. If an out-of-scope issue is discovered, stop and report — no unauthorized scope expansion.

Step 4: Verification
The coding skill requires the Agent to self-verify, including at minimum:

• Syntax check (tsc --noEmit)
• Unit tests (if available)
• Local dev server functional verification
• Edge cases (very short / very long articles)

Step 5: Delivery
Output change manifest + deployment instructions (for 小蜜蜂🐝 to execute). No direct SSH deployment — this is SFD's iron rule.

Lessons Learned

Pitfall 1: Agent skips planning and starts writing immediately
Fix: Explicitly write "must output an execution plan first, wait for user confirmation before writing code" in the system prompt. The coding skill already includes this requirement, but agents sometimes skip it.

Pitfall 2: Modified files outside the plan
Symptom: fixing a button style, then "optimizing" a bunch of unrelated CSS files. CSS conflicts later, no idea where they came from.
Fix: add a line to the PRD: "this task only allows modifying files X / Y / Z."

Pitfall 3: Insufficient verification before reporting completion
Symptom: Agent says "tests passed" but only curl'd the dev environment once. SSR hydration broke in production.
Fix: acceptance criteria must be explicit, not vague like "tests passed."

Pairing with Other Skills

• security-audit (小猎鹰🦅): auto-triggered security review after code completion
• browser-automation (小刺猬🦔): E2E verification after deployment
• self-improving-agent: logs each pitfall so similar tasks auto-avoid them next time

SFD Lab Notes

We completed 32 feature development tasks using the coding skill in March. Rework rate dropped from 40% (writing code ad-hoc) to 12%. The biggest change: forcing the planning step makes the Agent expose "I'm not sure about this" before writing code — asking upfront saves far more time than rewriting later. If your AI coding keeps producing "it runs but is broken" or "fixing one bug introduces another," try installing the coding skill first.