☕ Kopi

Agentic skills for project managers and software architects.

Connecting evidence, decisions, meetings, delivery state, analysis, presentations, and performance reviews — while keeping each workflow independently usable.

License: MIT Codex Plugin Agent Plugins Skills Validation

The central router is kopi-mode. Start there when a request spans more than one deliverable. Invoke a specialist skill directly when the outcome is already clear.

📦 Installation

Kopi packages the same eleven skills for Claude Code, Claude chat/Cowork, ChatGPT Work/Codex, and Cursor. See the installation guide for local installation, updates, prerequisites, and host smoke checks.

From the repository root, build the portable plugin ZIP and OpenAI local marketplace:

python3 scripts/package_plugin.py

Use dist/kopi-0.1.2.zip for Claude upload or extraction into Cursor's local plugin folder. For ChatGPT Work/Codex, register the generated dist/kopi-openai-0.1.2 marketplace and install through the Plugins Directory. Claude Code can also load this checkout directly with claude --plugin-dir ..

These are local distribution packages, not public marketplace listings. Host and account capabilities determine which workflows can execute. Presentation, spreadsheet, calendar, and source-access tools are supplied by your environment; Kopi does not install libraries or connect accounts.

🧱 Layers

#	Layer	Responsibility
1	Router	Chooses the smallest complete workflow and sequences connected work.
2	Workflow skills	Own one professional outcome each.
3	Playbooks	Describe the end-to-end path from request to verified deliverable.
4	References & rubrics	Provide schemas, prompts, source rules, and quality gates only when needed.
5	Adapters	Use available presentation, calendar, tracker, document, browser, and spreadsheet capabilities without binding the reasoning layer to one vendor.
6	Evaluation	Checks structure, routing, authority boundaries, and deliverable behavior.

🧰 Skill catalog

Skill	Use it for	Primary output
`kopi-mode`	Connected or ambiguous requests	Routed workflow and coordinated deliverables
`build-presentation`	PowerPoint and executive narratives	Render-verified deck
`run-meeting`	Meeting preparation, scheduling, and closure	Agenda, pre-read, event, decisions, and actions
`control-portfolio`	Multi-initiative visibility and intervention	Reconciled register and executive intervention view
`research-technology`	Current tools, frameworks, releases, and AI capabilities	Time-stamped, cited research brief
`decide-architecture`	Architecture options and target-state decisions	Decision record with tradeoffs and impact
`analyze-data`	Delivery, operational, and business datasets	Reproducible analysis and honest visuals
`recall-work`	Catch-up, resume, and handoff	Current-state status capsule
`challenge-deliverable`	Red-team review of plans, designs, and decks	Prioritized review findings without silent edits
`capture-learning`	Recurring workflow failures and lessons	Evidence-backed structural improvement proposal
`write-performance-review`	Direct-report reviews from 360 feedback and self-reflection	Employee report and separate confidential manager report

🔗 Operating model

Kopi follows a shared chain:

research + data ─▶ architecture decisions ─▶ meetings + work items ─▶ portfolio ─▶ presentation

Each workflow can run alone. When composed, records retain source pointers and stable identities so claims can move into decisions, actions, metrics, and slides without being rewritten from memory.

Performance reviews start with the person's role, supplied 360 feedback, and self-reflection; formal goals are optional. Kopi first presents a confidential evidence synthesis and numbered ideas, then invites the manager's observations and selection. It produces an employee report with reviewer identities protected and a separate manager report containing attribution, disagreements, evidence gaps, and working notes. The employee copy still needs the manager's check for identifying context known only to the participants.

The writing idea selection reference can also be composed by other writing workflows. It takes supported ideas and returns a stable selection record for drafting; it does not require multiple agents or activate a comparison step for every request.

Important

External changes are gated. Local analysis and drafts may proceed when in scope; sending invitations, changing trackers, posting messages, or updating shared systems requires explicit authority and destination checks. Create and update operations must be safe to retry and must be read back after execution.

💬 Example requests

“Use kopi-mode to turn these delivery metrics and architecture notes into the right materials for tomorrow’s steering review.”
“Build a decision deck from these findings and create the actual PowerPoint.”
“Research recent changes in our agent frameworks and assess whether the architecture roadmap should change.”
“Reconcile active initiatives and tell me which three need intervention.”
“Prepare an architecture review; do not send invitations until I approve.”
“Use write-performance-review for my direct report, an engineering manager. Here are their 360 feedback and self-reflection. Show me the evidence and numbered ideas before drafting the employee and confidential manager reports.”

🗂️ Package map

kopi/
├── plugin.json
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── .codex-plugin/plugin.json
├── INSTALL.md
├── skills/
│   ├── kopi-mode/
│   │   ├── SKILL.md
│   │   ├── playbooks/
│   │   └── references/
│   └── <specialist-skill>/
│       ├── SKILL.md
│       └── references/
├── evals/
├── scripts/validate_pack.py
├── scripts/package_plugin.py
└── tests/

✅ Validation

Run the deterministic pack checks from the repository root:

python3 -m unittest discover -s tests -v
python3 scripts/validate_pack.py .

The validator checks cross-platform manifest consistency, Claude marketplace resolution, skill metadata, resource reachability, internal links, unfinished placeholders, and legacy implementation terms. Packaging tests also check clean, reproducible archives and OpenAI marketplace resolution. These checks do not replace a fresh install in each host. The behavioral cases and run records are in evals/scenarios.md and evals/results.md.

📄 License

Released under the MIT License. See LICENSE.