agentic-toolkit

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# Agentic Toolkit

A set of prompts and tools I use. Treat this repo as a glimpse into my setup, not a sealed product. Use the parts you want and adapt them.

## Installation

```sh
npx skills add rstacruz/agentic-toolkit
```

Use it with [Pi](https://pi.dev/), [OpenCode](https://opencode.ai). Claude Code, Codex... pick your poison.

- **Pi:** Subagents are used in the *polish* skills, and Pi has no subagents by default. Consider [@tintinweb/subagents](https://github.com/tintinweb/pi-subagents).

- **OpenCode/Pi**: If you'd like to use a different model for polish reviews, create an agent with the name *"oracle"* and configure your preferred model for it. I personally use it with Opus, or GPT 5.5 high, or GLM 5.2 high. Otherwise, *polish* will default to whatever model you've got selected.

## Quick start

See [skills/](https://github.com/rstacruz/agentic-toolkit/tree/HEAD/skills/) and [`docs/skills.md`](https://github.com/rstacruz/agentic-toolkit/blob/HEAD/docs/skills.md) for the full skill reference.

### Plan it

Start with [`$brainstorm`](https://github.com/rstacruz/agentic-toolkit/blob/main/skills/brainstorm/SKILL.md). This is my replacement for plan mode.

- Asks questions relentlessly (inspired by [grill-me](https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/))
- Easily-skimmable output (Grounded facts, Steps, Risks)
- Works for all agents (including Pi which has no plan mode)
- Example: [example-plan.md](https://github.com/rstacruz/agentic-toolkit/blob/HEAD/docs/example-plan.md)

```
/brainstorm i want to implement config via c12 npm package
```

### For larger projects → megaplan <sup>experimental</sup>

Use [`$megaplan`](https://github.com/rstacruz/agentic-toolkit/blob/main/skills/megaplan/SKILL.md) when the work spans multiple slices (4+ tickets, evolving requirements, decisions you'll revisit weeks later).

- Living notebook that accumulates as work progresses
- Numbered requirements, decisions, risks, and grounded facts with cross-references
- Work-plan tracker with ticket/PR statuses

### Polish the plan

Use [`$polish-plan`](https://github.com/rstacruz/agentic-toolkit/blob/main/skills/polish-plan/SKILL.md) after creating a plan.

- Runs the plan through a subagent review loop (up to 7 passes)
- Applying fixes until there are no more changes needed

### Implement the plan

Implement the plan as you normally would (I suggest `/goal implement this plan`, as supported by many harnesses by default). No skill here.

<details>

<summary>Example</summary>

Here's one I would use:

```
/goal First, make empty commit then create draft PR (use pr-description skill).
Then implement the plan as described, correctly and completely.
Ensure that the `polish-implementation` skill (using Oracle agent type if available; else, Opus) is used afterwards and has completed successfully.
Ensure `trim-tests` skill was used, and that no other changes to tests are needed.
Ensure the changes are committed and pushed to Git.
Ensure `atk-pr-autofix` is used and runs successfully.
Ensure CI is passing in GitHub. Ultrathink. Plan: <plan file here>
```

</details>

### Polish the implementation

Use [`$polish-implementation`](https://github.com/rstacruz/agentic-toolkit/blob/main/skills/polish-implementation/SKILL.md) after implementing a plan.

- Requires a PR for the branch (creates a draft one automatically if missing)
- Findings are posted as self-review threads on the PR, then fixed, replied to, and resolved — durable across passes, unlike a purely local review loop
- Stops when a pass finds nothing new, up to 10 passes

### Watch CI

Use [`$atk-pr-autofix`](https://github.com/rstacruz/agentic-toolkit/blob/main/skills/atk-pr-autofix/SKILL.md) to monitor a PR until CI passes or needs intervention.

- Waits for CI to be green
- Addresses CI failures automatically

Agentic Toolkit

A set of prompts and tools I use. Treat this repo as a glimpse into my setup, not a sealed product. Use the parts you want and adapt them.

Installation

npx skills add rstacruz/agentic-toolkit

Use it with Pi, OpenCode. Claude Code, Codex... pick your poison.

Pi: Subagents are used in the polish skills, and Pi has no subagents by default. Consider @tintinweb/subagents.
OpenCode/Pi: If you'd like to use a different model for polish reviews, create an agent with the name "oracle" and configure your preferred model for it. I personally use it with Opus, or GPT 5.5 high, or GLM 5.2 high. Otherwise, polish will default to whatever model you've got selected.

Quick start

See skills/ and docs/skills.md for the full skill reference.

Plan it

Start with $brainstorm. This is my replacement for plan mode.

Asks questions relentlessly (inspired by grill-me)
Easily-skimmable output (Grounded facts, Steps, Risks)
Works for all agents (including Pi which has no plan mode)
Example: example-plan.md

/brainstorm i want to implement config via c12 npm package

For larger projects → megaplan ^experimental

Use $megaplan when the work spans multiple slices (4+ tickets, evolving requirements, decisions you'll revisit weeks later).

Living notebook that accumulates as work progresses
Numbered requirements, decisions, risks, and grounded facts with cross-references
Work-plan tracker with ticket/PR statuses

Polish the plan

Use $polish-plan after creating a plan.

Runs the plan through a subagent review loop (up to 7 passes)
Applying fixes until there are no more changes needed

Implement the plan

Implement the plan as you normally would (I suggest /goal implement this plan, as supported by many harnesses by default). No skill here.

Example

Here's one I would use:

/goal First, make empty commit then create draft PR (use pr-description skill).
Then implement the plan as described, correctly and completely.
Ensure that the `polish-implementation` skill (using Oracle agent type if available; else, Opus) is used afterwards and has completed successfully.
Ensure `trim-tests` skill was used, and that no other changes to tests are needed.
Ensure the changes are committed and pushed to Git.
Ensure `atk-pr-autofix` is used and runs successfully.
Ensure CI is passing in GitHub. Ultrathink. Plan: <plan file here>

Polish the implementation

Use $polish-implementation after implementing a plan.

Requires a PR for the branch (creates a draft one automatically if missing)
Findings are posted as self-review threads on the PR, then fixed, replied to, and resolved — durable across passes, unlike a purely local review loop
Stops when a pass finds nothing new, up to 10 passes

Watch CI

Use $atk-pr-autofix to monitor a PR until CI passes or needs intervention.

Waits for CI to be green
Addresses CI failures automatically