Atlassian CLI Skill for Agents

Teach your AI coding agent to use the official Atlassian Cloud CLI (acli) correctly, so it runs real commands instead of guessing wrong ones.

The Problem

Ask a coding agent to "create a Jira bug in the TEAM project" without this skill and you'll get one of these:

• A Python script using the Jira REST API that you have to install dependencies for, configure, and run yourself
• Commands for Bob Swift's ACLI: a different, paid, third-party tool with completely different syntax
• Hallucinated flags like --issue-type (doesn't exist) or --priority Critical (wrong format)
• No safety check: it might run delete commands without asking you first

The Fix

Install this skill and the same question produces:

acli auth status                    # checks you're logged in first
acli jira workitem create \
  --summary "Fix login bug" \
  --project "TEAM" \
  --type "Bug" \
  --assignee "@me"                  # correct flags, correct syntax

The agent confirms the command with you before running it (create = moderate safety tier), then executes it live and shows you the result.

With Skill vs Without: Hard Numbers

We ran 13 eval scenarios covering Jira, Confluence, Admin, and cross-product workflows, totaling 93 individual assertions (correct command, correct flags, auth check, safety tier, no hallucinated flags, etc.).

	Assertions Passed	Score
Without skill	18 / 93	19%
With skill	93 / 93	100%

+81 percentage points. The gap comes from three failure modes that disappear with the skill:

1. Wrong tool. Without the skill, agents default to REST API scripts or confuse the official acli with Bob Swift's ACLI. The skill teaches the correct acli <product> <entity> <action> [flags] pattern.
2. Wrong flags. Without the skill, agents guess flag names and get them wrong. The skill includes 1,250+ lines of command reference with every flag, short form, and default value.
3. No safety guardrails. Without the skill, agents run destructive commands without confirmation. The skill enforces three safety tiers (safe / moderate / dangerous) with escalating confirmation requirements.

What It Covers

Product	Operations
Jira	Work items (create, search, edit, transition, delete, clone, bulk ops), projects, boards, sprints, filters, dashboards, fields, CSV/JSON export
Confluence	Spaces (list, view, create, archive), pages (view), blog posts (create, list)
Admin	User management (activate, deactivate, delete, cancel-delete)
Rovo Dev	Session management (run, auth)
Auth	OAuth browser flow, API tokens, admin API keys, Rovo Dev scoped tokens, CI/CD pipeline setup

Safety Tiers

The skill classifies every command into a safety tier before execution:

Tier	Actions	Agent behavior
Safe	search, list, view, get, auth status	Executes immediately -- read-only, no risk
Moderate	create, edit, transition, assign, comment	Shows command, asks for confirmation, then runs
Dangerous	delete, archive, bulk ops via --jql	Previews impact first (runs a count/search), shows what will be affected, requires explicit confirmation. Never auto-passes --yes

Install

npx skills add asolfre/atlassian-cli-skill

The skill auto-triggers when you ask the agent to do something with Jira, Confluence, or Atlassian from the terminal. No configuration needed.

Other install methods

# Per-project
git clone https://github.com/asolfre/atlassian-cli-skill.git .opencode/skills/atlassian-cli

# Globally for all projects
git clone https://github.com/asolfre/atlassian-cli-skill.git ~/.config/opencode/skills/atlassian-cli

Prerequisites

1. Install acli (the official Atlassian Cloud CLI):

   brew tap atlassian/homebrew-acli && brew install acli

2. Authenticate with your Atlassian instance:

   acli auth login

How It Works

The skill uses progressive disclosure to stay efficient:

1. SKILL.md (400 lines) loads on trigger -- contains the command pattern, safety tiers, execution workflow, and the most common commands for each product.
2. Reference files (1,250 lines across 4 files) load on demand -- the agent reads references/jira.md only when it needs exact flags for a Jira command. Same for confluence.md, admin.md, and auth.md.

This means the agent doesn't burn context on Confluence reference docs when you're doing Jira work.

Trigger Accuracy

The skill triggers when it should and stays silent when it shouldn't. We tested 20 queries (10 that should trigger, 10 near-miss negatives that shouldn't) across 60 individual runs:

Metric	Score
Precision	100%
Recall	100%
Accuracy	100%

Examples of queries that correctly do not trigger the skill:

• "How do I configure the columns on my Jira board?" (conceptual, not CLI)
• "Write a Python script using the Jira REST API" (REST API, not acli)
• "Help me design a Jira workflow with custom statuses" (workflow design, not execution)
• "Can you read the Confluence page at https://..." (web URL, not CLI)

File Structure

SKILL.md                    # Main skill (400 lines)
references/
  jira.md                   # Jira command reference (660 lines)
  confluence.md             # Confluence command reference (240 lines)
  admin.md                  # Admin command reference (110 lines)
  auth.md                   # Auth methods reference (245 lines)
evals/
  evals.json                # 13 eval scenarios, 93 assertions
  trigger-eval.json         # 20 trigger accuracy queries

License

MIT