Quickstart
This page takes about 20 minutes. You end with a diagnostic summary and a draft decision for one piece of work.
Pick something specific: an approval, a recurring meeting, a report, a policy, a workaround, or a software feature. A whole department is too large; pick one process inside it.
1. Download the kit
Download discovery-kit.zip. It contains the process-discovery skill, the codebase-discovery skill, two commands, and the agent definitions.
2. Load the kit
Claude Code
Run this in a terminal, from any folder, with the path to the downloaded file:
claude --plugin-dir ./discovery-kit.zip
The kit is loaded for that session. Start Claude Code the same way each time you want it.
Codex
- Unzip the file.
- Copy
skills/process-discoveryandskills/codebase-discoveryinto~/.agents/skills/. To limit them to one repository, copy them into.agents/skills/in that repository. - Copy the files in
.codex/agents/into.codex/agents/in your project. - Start Codex.
The desktop app
If you use the desktop app, it installs the kit for Claude Code and Codex from Connect your AI tool. You do not need the ZIP.
Other AI tools
Paste the contents of skills/process-discovery/SKILL.md as instructions. The skill links to files in references/; load each one when the skill asks for it.
3. Start the conversation
In Claude Code, run:
/process-discovery:discover our purchase approvals
If the desktop app installed the kit, the command has no prefix: /discover our purchase approvals.
In Codex or another tool, write a request in plain words:
Our purchase approvals take three weeks. Help me work out which steps we still need.
Replace the example with your own piece of work.
4. Answer the questions
The skill asks one question at a time and moves through these stages:
- Frame. What you want to simplify, why now, your role, and what you can change.
- Size. Whether it is an area, a process, or one thing, and how often it runs, how many people it touches, how long it takes elapsed, and how much working time it takes.
- Break down. For a process, the steps, approvals, meetings, reports, forms, systems, and workarounds in the last real case. You choose up to five to assess.
- Ask. The questions for each part’s type.
- Follow up. Extra questions when an answer reveals something: a rule that requires it, a check that never rejects anything, a report nobody can name a reader for, work done by hand because a system fails.
- Decide. A draft of keep, change, stop as a trial, or investigate.
Answer from the last real case, not from the written procedure. If you do not know, say so. The skill records an open question and who could answer it.
Paste or attach documents, logs, or calendars when the skill asks for evidence. They are recorded with a location and a quote.
5. Read the summary
Every conversation ends with a diagnostic summary. It states:
- what the work is and who it is for
- the evidence for and against keeping it
- what is still unknown, next to the claim it affects
- the draft decision
The draft is not a decision. You and the people involved decide.
6. Ask for what you need next
Ask for any of these in the same conversation:
- an item brief or decision brief
- a stakeholder register
- a stop trial plan
- a message to the owner of the work
- an interview guide built from the open questions
- a workshop plan
Next
- To keep the result and run a workshop, create a workspace and connect your AI tool.
- To assess the software that carries the work, see Map a codebase.