April 2026
10 Claude Code skills I install on every project
The exact setup I use across a prediction market platform, an AI vision prototype, and a 48-agent venture validation pipeline. Every skill explained with the problem it solves and the incident that made me add it.
1. Verify before claiming
What it does: A CLAUDE.md rule that forces Claude to actually check before saying something doesn't exist.
Why I added it: Claude confidently told me a Supabase table didn't exist. It did. I'd already created it in a previous session. One wrong assumption cascaded into 30 minutes of debugging a non-problem.
Add this to your CLAUDE.md
Verify before asserting absence. Before saying "X doesn't exist" — actually check. Read the file, query the database, list the directory.
2. Auto-save session state
What it does: At the end of every session, Claude commits all changes and writes a summary to LAST_SESSION.md.
Why I added it: I lost an entire evening's architecture decisions when a session crashed. Next session, Claude had no idea what we'd decided. LAST_SESSION.md means the next session starts with full context.
Add this to your CLAUDE.md
After completing a feature or major decision, commit all changes and write a summary to LAST_SESSION.md with what was built, decisions made, and open threads.
3. File organization enforcement
What it does: A rule that stops Claude from dumping .md files in the project root.
Why I added it: After 20 sessions on one project, I had 47 markdown files in root. Plans, research, notes, reports — all flat in one directory. Took an hour to sort. Now Claude uses subfolders automatically.
Add this to your CLAUDE.md
Keep project roots clean. Don't dump .md files in root. Use subfolders: plans/, research/, docs/. Name for searchability, not convenience.
4. Decision tracking
What it does: When I make a decision — reject an approach, pick a direction, set a constraint — Claude saves it immediately as a decision file.
Why I added it: I rejected a database schema three sessions ago. Next week, Claude proposed the exact same schema. It had no memory of why we rejected it. Decision files fix this permanently.
Add this to your CLAUDE.md
Save decisions immediately. When the user makes a clear decision, create a decision file with what was decided, what was rejected, and why.
5. Security self-review
What it does: Before any deploy, Claude runs a security check automatically.
Why I added it: I deployed a Supabase project with a SECURITY DEFINER function that was world-callable. A fresh-context security agent caught it in 2 minutes. I now run this before every staging deploy.
Add this to your CLAUDE.md
Before any staging/production deploy, run a security review. If the session involved 5+ iterations on the same files, spawn a separate review agent with fresh context.
6. No deploy without consent
What it does: Claude never deploys to production without me explicitly saying 'deploy.'
Why I added it: Claude helpfully ran vercel --prod during a debugging session. On a Saturday. With broken auth middleware. Staging first, always, and only with verbal confirmation.
Add this to your CLAUDE.md
Never deploy to production without explicit user consent. Staging/preview first, always.
7. Try before delegating
What it does: Before telling me I need to do something manually, Claude checks if it can do it itself.
Why I added it: Claude kept saying "you'll need to go to the Supabase dashboard to do this." Half the time, the Supabase CLI could do it. Now it tries first, delegates only when genuinely blocked.
Add this to your CLAUDE.md
Try before delegating. Before telling the user "you need to manually do X", check whether you can do it yourself with Bash, CLI tools, or MCP servers.
8. Wisdom extraction
What it does: When a mistake is corrected, Claude captures the lesson as a reusable rule with a trigger condition.
Why I added it: I fixed the same Supabase migration numbering bug three times across different projects. Now there's a wisdom file that triggers whenever Claude pushes migrations. Never happened again.
Add this to your CLAUDE.md
When a mistake is corrected, ask: is this specific to this project, or a general pattern? If general, create a wisdom entry with trigger condition, rule, and the incident that taught it.
9. Context loading on session start
What it does: First thing every session: Claude reads the project's context bundle and STATE.md before responding.
Why I added it: Without this, the first 10 minutes of every session is Claude rediscovering the project. With it, Claude starts with full context — architecture, decisions, current state, open threads.
Add this to your CLAUDE.md
On first action in a session, read the project's context file and STATE.md before responding. This ensures you start with both durable architecture knowledge and current state.
10. Append-only state files
What it does: STATE.md is never overwritten — only specific sections are updated.
Why I added it: A background agent deleted half of STATE.md by rewriting the whole file. It destroyed deployment records and pending action lists that other agents were reading. Append-only prevents this.
Add this to your CLAUDE.md
STATE.md is append-only. Use Edit to add/update specific sections — never Write/replace. Shared inter-agent state; deleting content breaks coordination.
Want all 10 pre-configured?
The Blueprint skill installs all of these plus memory persistence, auto-checkpoints, and wisdom extraction in one command.
Get the Blueprint