AI CLI Bootstrap Skill

Build tiny task-scoped CLIs on demand when an agent cannot directly access the data, page, API, or platform action a user needs.

This is a starter repo for agent unblock workflows: reuse an existing tool if possible, otherwise scaffold the smallest useful helper, run it immediately, and answer from the result.

When To Use This

Use it for prompts like:

• "Read this X post for me"
• "Analyze this YouTube script"
• "Summarize this page the agent cannot read directly"
• "Compare these API responses"

Do not use it when the helper is turning into a real product with config, subcommands, browser login flows, or framework structure.

30-Second Quick Start

python scripts/bootstrap_ai_cli.py youtube-script \
  --purpose "Fetch transcript text and metadata for a YouTube video" \
  --template transcript_stub \
  --workspace "$PWD"

This creates:

.agent-tools/
  youtube-script/
    cli.py

Patch the generated file with the smallest missing fetch or transform logic, run it on a real sample, and answer from the output.

If your repository already uses another scratch-tool directory, override it:

python scripts/bootstrap_ai_cli.py youtube-script \
  --purpose "Fetch transcript text and metadata for a YouTube video" \
  --template transcript_stub \
  --workspace "$PWD" \
  --tools-dir ".codex-tools"

End-To-End Example

Representative example: examples/youtube_transcript

1. User request

Analyze this YouTube script for pacing and retention.

2. Bootstrap command

python scripts/bootstrap_ai_cli.py youtube-script \
  --purpose "Fetch transcript text and metadata for a YouTube video" \
  --template transcript_stub \
  --workspace "$PWD"

3. Patched cli.py excerpt

def extract_video_id(target: str) -> str:
    if "youtube.com" in target or "youtu.be" in target:
        ...
    return target.strip()


def fetch_segments(video_id: str, lang: str) -> list:
    return YouTubeTranscriptApi.get_transcript(video_id, languages=[lang])

4. Example dependency

python -m pip install youtube-transcript-api

This dependency is only required for the representative YouTube example, not for the scaffold script itself.

5. Example execution

python examples/youtube_transcript/cli.py "https://youtu.be/dQw4w9WgXcQ" --json

6. Example JSON output

{
  "status": "ok",
  "tool": "youtube-transcript",
  "video_id": "dQw4w9WgXcQ",
  "lang": "en",
  "segment_count": 3
}

7. Final answer built from the output

The script gets to the point quickly, but the payoff arrives almost immediately after the hook. To improve retention, expand the middle section with one concrete example before the closing action.

Template Types

Template	Best for	What it gives you
generic	Anything custom	Minimal one-file scaffold with TODO logic
fetch_json	APIs and JSON endpoints	urllib fetch + JSON normalization
fetch_page_text	Articles and webpages	HTML fetch + title/text extraction
parse_local_file	Local JSON or text artifacts	File read + simple normalization
transcript_stub	Video or audio transcript flows	Transcript-shaped output and segment scaffold

Agent Install

Agent	Entry point	How to use it
Codex	SKILL.md	Place this repo in your Codex skills directory
Claude Code	prompts/claude-code.md	Copy the prompt into project instructions or task preamble
Gemini CLI	prompts/gemini-cli.md	Copy the prompt into project instructions or task preamble
Other coding agents	prompts/core.md	Paste the core instructions into the agent's working context

Codex

git clone https://github.com/Siwoo4985/ai-cli-bootstrap-skill.git \
  "${CODEX_HOME:-$HOME/.codex}/skills/ai-cli-bootstrap-skill"

Claude Code, Gemini CLI, And Others

1. Clone this repository anywhere in your workspace.
2. Start with prompts/core.md.
3. Add prompts/claude-code.md or prompts/gemini-cli.md if you want agent-specific defaults.

Examples

Example	Purpose	Notes
youtube_transcript	Representative end-to-end transcript flow	Best default onboarding path
webpage_extractor	Stable page-to-text extraction	Standard-library example
x_post_reader	Illustrative post reader	Shows normalization shape, not a guaranteed stable fetch path

Design Principles

• references/ai-only-cli-rules.md
• references/decision-guide.md
• references/anti-patterns.md

Validation And Tests

Run the scaffold tests:

python -m unittest discover -s tests

Validate the Codex skill metadata:

python "${CODEX_HOME:-$HOME/.codex}/skills/.system/skill-creator/scripts/quick_validate.py" \
  "${CODEX_HOME:-$HOME/.codex}/skills/ai-cli-bootstrap-skill"

Contributing

Small, focused improvements are the best fit:

• tighten prompts
• improve scaffold templates
• add clearer examples
• keep docs, templates, and behavior consistent

If the generated helper starts turning into a framework, move that work into a separate project instead.

License

This project is released under the MIT License.