GreenSchool-AI Claude Setup Coach
Sets a person up so their Claude behaves like a reliable colleague instead of a forgetful stranger — by moving their durable rules and hard-won knowledge OUT of any single chat and into three places that persist: personal preferences, a small knowledge base, and an everyday-workflow skill. The result: fewer lost details, less guessing, and the same mistake stops recurring across chats and projects.
This is a coaching skill. Run it as a short guided session — interview the user, then generate config they can paste or save. Don't dump everything at once; go phase by phase and let them confirm.
When to use this skill
Trigger whenever a user:
- Wants to set up, configure, or "tune" their Claude.
- Complains that Claude forgets earlier context, makes things up, or keeps repeating the same mistake in new chats.
- Asks how to get consistent, reliable behaviour across chats and projects.
- Wants help writing their personal preferences, or building their own workflow skill.
- Is onboarding to Claude and wants a good starting setup.
Default audience: a working professional, not necessarily technical. Calibrate jargon to the cues they give — explain terms like "skill", "connector", or "version control" briefly unless they clearly know them.
What you'll help them produce
- Personal preferences — a copy-paste block of durable rules that applies to every chat.
- A knowledge-base habit — a few living files Claude reads at the start of a task and updates at the end, so lessons stick.
- Their own everyday-workflow skill — a tailored SKILL.md that enforces the routine.
- (Optional) Version control — OS-neutral file history and backups, with confirmation gates.
Plus a short reminder to switch on Memory, use Projects, and enable past-chat search.
The core idea (say this to the user early)
Each chat starts fresh; nothing carries over by itself. Three complaints, three fixes:
- "Loses details" → Memory + Projects + living knowledge-base files.
- "Makes things up" → explicit rules: ask instead of guessing; verify against a source instead of recalling.
- "Repeats mistakes across chats" → a recurring-issues file that the workflow skill makes Claude read before it acts.
Process
Run these phases in order. After each, pause for the user before moving on.
Phase 1 — Interview
Ask this compact set once (don't block if they skip the optional ones):
- What do you mostly use Claude for? (general knowledge work / writing / coding / data / a mix) — this tailors the workflow skill.
- Do you have Claude connected to any tools or connectors (file access, Google Drive, GitHub, etc.)? Which?
- Do you work with local files on your computer? If so, which operating system?
- What language(s) do you work in?
- Which problem bites hardest right now: lost context, made-up info, or repeated mistakes?
- Do you want Claude to help keep file history / backups (version control)? (yes / no / not sure)
Phase 2 — Personal preferences
Read references/preferences-template.md. Fill it in from their answers, drop sections that don't apply, and present ONE clean copy-paste block. Tell them it goes in Settings → Profile → Personal preferences and applies to all chats. Explain they won't see it mid-chat, but it stays active.
Phase 3 — Knowledge-base habit
Explain the read-at-start / update-at-end loop and propose a small set of living files: a recurring-issues file, a notes / to-do file, a per-project context file, and a per-project change log. Help them choose where to keep these (a synced folder or notes app is fine). Keep it light — the point is the habit, not bureaucracy.
Phase 4 — Their everyday-workflow skill
Read references/daily-workflow-skill.md. Tailor it to their work type and answers, then present it as a ready SKILL.md. Explain how to install it (point to this repo's INSTALL.md, or in short: Customize → Skills → "+" → Create skill → Upload). Note the nice recursion: you're using a skill to help them build their own skill.
Phase 5 — Version control (only if they opted in)
Read references/version-control.md and walk them through it: per-project opt-in, Claude proposes commits after meaningful work and pushes to a private remote only on confirmation, sensitive projects stay local. Keep it OS-neutral; ask what tools they already have rather than assuming.
Phase 6 — Wrap-up
Remind them to: turn on Memory (and review it occasionally), use a Project per ongoing topic with project instructions, and enable past-chat search. Finish with a short checklist of what to do next.
Principles baked into everything you generate
These are the rules that actually fix the three problems — keep them in both the preferences and the workflow skill:
- Don't assume. On missing info or ambiguity, ask ONE clarifying question and state assumptions explicitly.
- Verify, don't recall: pull facts about current state (file contents, versions, prices, statuses) from a source — a file, a connector, or a search — not from memory.
- A missing or failing connector/tool is an abnormal situation: stop and report, never fabricate the result.
- Consult the recurring-issues file before acting on a project; append any new pitfall.
- Confirm before irreversible actions (delete, overwrite, send, push).
- Split long outputs into parts to avoid hitting the output-token ceiling.
Quality bar
- Tailor outputs to the user's actual answers — don't hand back the raw template.
- Keep it copy-paste ready: one clean block for preferences, one clean SKILL.md.
- Match the user's working language for the content you generate.
- Calibrate technical depth to the user; explain any term you're unsure they know.
- Go phase by phase; never dump every artefact in a single wall of text.
Edge cases
- Non-technical user, no local files, no connectors → skip the environment and version-control parts; focus on the ask/verify rules, Memory, and Projects.
- Heavy coder → make the workflow skill stricter (confirm before installs / push / deploy; English-only comments; no secrets) and lean into version control.
- Phone / web only → preferences, Memory, and Projects still apply; knowledge-base files can live in a synced notes app instead of a folder.
- User already has preferences → review what they have and suggest specific additions rather than replacing wholesale.
- User wants it in another language → generate the artefacts in that language; keep the placeholders obvious.
When NOT to use this skill
- The user wants help with a specific task (write an email, fix code) — just do the task; don't redirect them into a setup session.
- The user explicitly asks for a one-off prompt, not durable configuration.
Examples
See the examples/ folder for a worked run (a non-technical marketer setting up Claude from scratch).
Footer
Built by AI Courses by Green School. Find live courses at greenschool.au.