Coding Harness

Reusable agent skills and workflows for everyday development.

Prerequisites

Before installing an agent, review the prerequisites for required CLI tools, terminal setup, and provider access.

Installation

Recommended terminal coding agents

• Claude Code - Anthropic’s terminal-native coding agent.
  • revdiff plugin - Interactive diff review and annotation workflow.
  • Codex plugin for Claude Code - Adds OpenAI Codex integration to Claude Code.
• Pi Coding Agent - Extensible terminal coding agent with packages, tools, skills, and TUI support.
  • revdiff package - Interactive code review and diff navigation.
  • pi-subagents package - Delegates work to focused subagents and chains.
  • pi-web-access package - Adds web search, content fetching, and library research tools.
  • pi-intercom package - Lets multiple Pi sessions communicate locally.
  • pi-voice-stt package - Adds multi-provider voice dictation and audio file transcription.

Also worth trying

• Codex CLI with revdiff plugin or Codex app
• OpenCode
• GitHub Copilot CLI

Skills

To use coding-harness/skills in your projects, copy or symlink the skills folder to:

• For Claude Code:
  • project-folder/.claude/skills or ~/.claude/skills
• For Pi Coding Agent, OpenCode, Codex, Copilot:
  • project-folder/.agents/skills or ~/.agents/skills

Usage and workflow

Ask your coding agent in natural language or use the skill command syntax supported by your host. The examples below use portable prompts; direct skill invocations use /<skill-name> explicitly.

Check understanding and sharpen a task

These skills are manual-only. Claude Code uses /<name>; Pi uses /skill:<name>. In OpenCode, ask to use the skill by name; slash commands require the skill's opencode/slash metadata to be enabled.

• bro: restate the assistant's last answer in plain language.
• in-your-own-words: restate your goals and the problem to check understanding, without solving it.
• sharpen: deepen a task's methods and evidence without broadening its scope, then rewrite the prompt.
• ketchup: catch up on chat progress and every decision still waiting on you, without starting more work.

/sharpen Review this migration plan for correctness, keeping its scope unchanged.

Implement a user story

Activate the brainstorm skill:

let's brainstorm this user story: """User story description"""

If brainstorming produced an implementation plan, use the project-appropriate coding skill, such as java-coder or typescript-coder:

implement the plan YYYY-MM-DD-user-story-implementation-plan.md
  using the applicable project coding skill and provide human-verifiable evidence

Architecture design

The architect skill is manual-only. Use it after behavior and important constraints are understood to produce a caller-first, implementation-ready structural design.

/architect Define the types, signatures, and module boundaries for this approved design.

It returns a design by default. Implementation requires an explicit request or later approval.

Candidate arena

The arena skill is manual-only for direct use. Invoke it when several independent attempts at the same bounded artifact would improve the decision.

/arena Produce three independent designs for this parser contract and synthesize the strongest result.

It returns one coherent synthesis and the applicable verification result rather than a collection of candidate fragments.

Code review

Activate the code-review skill:

code review uncommitted changes

Blast-radius analysis

The blast-radius skill is manual-only. Invoke it with a bounded local comparison:

/blast-radius Analyze HEAD~1..HEAD.

It traces credible effects beyond the obvious diff and seeks executable proof for decision-critical safety assumptions.

Session reflection

The reflect skill is manual-only. Invoke it after a substantial session:

/reflect Review this session and propose evidence-backed improvements.

It proposes evidence-backed improvements to existing skills and waits for individual approval before editing.

Learn a codebase

The teach skill is manual-only. Invoke it with the learning request:

/teach Explain how request retries work in this service.

It explains repository-backed code progressively without modifying the project.

Write tests

Activate the write-tests and test-coverage skills:

write tests for new and modified code

add test and branch coverage for UserController

Fix a bug

Activate the debug-error and/or clarify skills:

debug this error: """stack trace or errors in logs"""

clarify why the response status is 500 instead of 201

For an explicit test-first bug fix, invoke tdd:

/tdd Fix the parser bug and demonstrate a regression test failing before the fix and passing afterward.

It uses a focused existing test boundary, avoids broad harness setup, and reports missing execution evidence rather than claiming an unobserved result.

Refactoring

Activate the refactor-code and java-coder skills:

refactor code in the user login flow, following guidelines from the java-coder skill

Technical writing

Use the technical-writing skill for substantial technical documents and design artifacts:

review this deployment runbook for incorrect assumptions, missing prerequisites, and unsafe recovery steps

Use unslop alone for small style-only edits.

Documenting

Choose the diagram skill based on the artifact you need:

• Use pretty-mermaid for Mermaid-as-code, themed SVG rendering, ASCII output, or batch rendering.
• Use diagram-design for editorial standalone HTML/SVG diagrams, branded styling, or redraws of Mermaid and draw.io sources.

Use the pretty-mermaid skill to create an ASCII terminal-friendly sequence diagram of the endpoint /api/v1/users.

Use the diagram-design skill to create an editorial architecture diagram of this service as standalone HTML.