Dev-Skill-Unslop

KO | EN

gitlite — search

#swift #nlp #open-source #rust #macos #awesome #redis #cli #bash #security

Dev-Skill-Unslop

(★ 9)

The Dev Skill Reborn

파일 탐색기

ZIP 다운로드

- dev-skill.md
- openai.yaml
- - .gitignore
  - package-lock.json
  - package.json
  - README.md
- - .gitignore
  - devskill-guard.js
  - install.ps1
  - opencode-adapter.bat
  - opencode.mcp.jsonc
  - package.json
  - README.md
  - start-guard-core.bat
  - start-opencode-without-guard.bat
  - stop-guard-core.bat
- - README.md
  - rebon-adapter.bat
  - rebon-guard-hook.mjs
  - rebon-plugin.json
  - start-rebon-with-guard.bat
  - start-rebon-without-guard.bat
  - stop-guard-core.bat
- README.md
- - codebase-design.md
  - common-sense.md
  - context-optimization.md
  - decision.md
  - diagnosing-bugs.md
  - domain-modeling.md
  - grilling.md
  - handoff-and-boundaries.md
  - host-enforcement.md
  - implementation-and-tdd.md
  - improve-codebase-architecture.md
  - manual-procedure.md
  - markdown-tables-and-diagrams.md
  - opencode-host-adapter.md
  - project-discovery.md
  - prototype.md
  - rebon-host-adapter.md
  - research.md
  - review.md
  - talk-like-king.md
  - teaching.md
  - to-spec.md
  - to-tickets.md
  - two-layer-development-planning.md
  - write.md
  - writing-style.md
- common-sense-table.md
- design.md
- plan.md
- release.md
- stage-0-route.md
- work.md
.gitignore
LICENSE
manifest.json
mode-gate.md
README.md
SKILL.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

> AI can write code faster than you can regret the architecture.
>
> — ChatGPT 5.6

DevSkill Unslop turns an AI agent from a prompt-driven helper into a developer with a working method: real-world professional workflows, a route that triggers the right capability, and a rule against work that looks safe but does not move the human goal.

## What it changes

- **Professional workflows** — design through release
- **Automatic routing** — no remembered prompts
- **Less defensive slop** — work must move value

## Lineage

DevSkill Unslop builds on Matt Pocock’s [Skills for Real Engineers](https://github.com/mattpocock/skills).

Matt’s project brought real engineering methods into agent work. This release rebuilds their runtime topology around stage families with more developed modules, explicit module contracts, automatic routing, durable re-entry, and a stricter answer to AI slop.

## A complete development route

```mermaid
flowchart TB
    A["Human task"] --> B["Route"]
    B --> C{"What is needed now?"}
    C -->|"Meaning unclear"| D["Design"]
    C -->|"Design settled"| E["Plan"]
    C -->|"Approved slice"| F["Work"]
    C -->|"Ready result"| G["Release"]
    D --> E
    E --> F
    F --> G
    G --> H["Complete or reopen owner"]
```

| Family or overlay | Module or sub-operation | Job |
|---|---|---|
| Admission | Mode Gate | Confirm host, mode, adapter choice, and capability fit |
| Runtime | Host Enforcement | Enter and advance declared operation checkpoints when the host can intercept them |
| Route | Route | Select one current consumer |
| Route | Project Discovery | Establish greenfield, brownfield, or uncertain facts |
| Route | Common Sense — Vibe and Attention | Retrieve a relevant pattern reference |
| Design | Grilling | Present reviewed design choices |
| Design | Review | Return bounded findings only |
| Design | Decision — Think and Humour | Represent, test, and select semantic options |
| Design | Research — Pathfinding | Find deep factual routes under fog |
| Design | Domain Modeling | Resolve terms, rules, and exact semantic writes |
| Design | Prototype | Answer one bounded runnable question |
| Design | Manual Procedure | Guide a temporary human operation |
| Design | Teaching | Build demonstrated human understanding |
| Design | Codebase Design | Choose seams, modules, and dependencies |
| Design | Improve Codebase Architecture | Surface evidence-backed architecture candidates |
| Plan | Two-Layer Development Planning | Show workflow and structure together |
| Plan | To-Spec | Shape accepted meaning into a specification |
| Plan | To-Tickets | Form outcome-bearing vertical slices |
| Work | Implementation and TDD | Change one accepted behavior at a time |
| Work | Diagnosing Bugs | Narrow symptoms to a supported cause |
| Work | Handoff and Boundaries | Transfer work across a real boundary |
| Work | Context Optimization | Bound reading, reasoning, writing, and continuation |
| Runtime | Rebon Host Adapter | Use Rebon-native tools and classify Guard capability |
| Runtime | OpenCode Host Adapter | Classify OpenCode Guard capability |
| Presentation | Write | Render reports, records, conversations, and instructions |
| Presentation | Markdown Tables and Diagrams | Render a selected structure clearly |
| Presentation | Writing Style | Select an explicit writing treatment |
| Presentation | Talk Like King | Style non-instructional conversation and creative work without changing meaning |
| Release | Release | Reconcile work and return an effect or reopening path |

Each module has a task trigger, inputs, operation, authority boundary, return, and recovery path. The user may name a module, but naming it is never the only way it starts.

## The decision tree runs itself

```mermaid
flowchart TB
    A["Task arrives"] --> B["Mode and host admission"]
    B --> C["Route"]
    C --> D{"Next operation determined?"}
    D -->|"Yes"| G["Next stage or module"]
    D -->|"No"| E["Decision frames intent"]
    E --> F{"Candidate needs findings?"}
    F -->|"Yes"| I["Review"]
    I --> E
    F -->|"No"| G
    G --> H["Re-entry after each result"]
    H --> C
```

This makes the skill usable without ritual prompts such as “now review,” “now research,” or “use TDD.” A direct request stays direct. An indirect engineering concern reaches Decision first, so the next step reflects the human goal instead of a guessed workflow.

Every implementation slice receives Standards and Specification review. When accepted Design or final Grill decisions and the plan define one complete implementation, phase, roadmap, or merge outcome, that outcome receives the three-axis closure review. A checklist records the passed result; it does not choose the review.

## It refuses defensive busywork

AI often treats “more checking” as “more responsible.”

That can turn a five-minute task into a five-day ceremony: hashes, records, fallback layers, verification loops, and reports nobody will use.

DevSkill Unslop keeps a rule at the center:

> Retain a rule only when it changes a current read, question, choice, action, return, or human Value Gap.

A module may not add machinery merely because it looks safe. Review finds problems. Decision selects meaningful options. Work closes the actual gap.

## Built for short contexts and long runs

| Environment | Runtime approach |
|---|---|
| Local single agent | Sequential workflow |
| Small context window | Bounded reading and synthesis |
| Parallel agent host | Specialized bounded workers |
| Rebon optimized mode | Context optimization and durable continuation |

On compatible Rebon agents, a goal with settled design and permissions can continue for **10+ hours without constant human supervision**. Context-heavy reading, implementation, review, and verification stay bounded, so the main agent does not carry every file and failed attempt forever.

Humans still own meaning, authority, and permissions.

## Installation

### Full package

```bash
git clone https://github.com/shaggyfeng/Dev-Skill-Unslop.git
```

The repository is the full installation: the DevSkill runtime plus Guard Core, Rebon, and OpenCode adapter source. Install the repository as `dev-skill` inside your agent’s skill directory, then follow an adapter guide only if you want host-side enforcement.

For Codex:

```text
~/.codex/skills/dev-skill
```

Then describe the task normally. The skill routes the work.

### Full instruction-guided skill

Download `DevSkill-Unslop-2.3.0-soft-guide.zip` when you want the complete DevSkill instruction set without host-enforcement code. It trusts the agent model to follow DevSkill directly; it contains no host-adapter source or optional npm host code and operates as `instruction-guided`.

### Add host enforcement

Start with the instruction-guided skill, then extract the shared Guard Core patch and one matching host patch into that same DevSkill folder when code-level enforcement is needed.

| Release asset | Install when | Requires |
|---|---|---|
| `DevSkill-Unslop-2.3.0-soft-guide.zip` | You trust the model to follow DevSkill | Nothing else |
| `DevSkill-Unslop-2.3.0-guard-core-patch.zip` | You need code-level host enforcement | Instruction-guided skill |
| `DevSkill-Unslop-2.3.0-rebon-adapter-patch.zip` | You use Rebon with enforcement | Instruction-guided skill + Guard Core |
| `DevSkill-Unslop-2.3.0-opencode-adapter-patch.zip` | You use OpenCode with enforcement | Instruction-guided skill + Guard Core |

The instruction-guided skill is complete. Rebon and OpenCode patches add code-level behavior enforcement for weaker models; a smart model can follow the runtime instructions without them. Each host patch includes its setup guide.

At Mode Gate, a supported host asks whether to use code-level enforcement or the instruction-guided skill. Start Rebon or OpenCode in that matching mode first: a live host hook cannot be switched on or off truthfully mid-session.

## Author

**Shaggy Feng (嘴上云 | Mouth on Cloud)**

## License

[MIT](./LICENSE)

Dev-Skill-Unslop

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Dev-Skill-Unslop

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README