6 Claude Code Skills That Prevent the Most Common Autonomous Agent Failures

Autonomous coding agents keep failing the same six ways: guessing a database column name and crashing, declaring "deployed successfully" without checking the endpoint actually responds, editing production files with no audit trail, restoring a monitor that was deliberately killed, losing 10k tokens re-orienting at the start of every session, and reporting "it seems to work" instead of quoted evidence.

We pulled the fixes for all six into standalone Claude Code skills — .md files you drop into ~/.claude/skills/ and they show up as slash commands. No build step, no dependencies.

Repo: https://github.com/ForeverTools/agent-skills

Install

# One skill
curl -o ~/.claude/skills/fix-errors.md \
  https://raw.githubusercontent.com/ForeverTools/agent-skills/main/skills/fix-errors.md

# All six
mkdir -p ~/.claude/skills && \
curl -L https://github.com/ForeverTools/agent-skills/archive/main.tar.gz | \
  tar xz --strip-components=2 -C ~/.claude/skills "agent-skills-main/skills/"

Restart Claude Code. They appear as /skill-name.

The six skills

Skill	What it does	Pattern it encodes
/fix-errors	Triages an error as benign / transient / real regression before touching anything	Diagnose before you code
/verify-endpoint	Hits an HTTP endpoint, checks status + body, reports PASS/FAIL/ERROR	Verify the deliverable, not the log
/smoke-test	Batch-checks a list of endpoints in parallel, zero setup	Catch regressions before deploy
/db-query-safe	Reads a SQLite table's schema before running a query against it	Never guess column names
/log-change	Writes an audit-trail row before a production edit	Log it before you touch it
/session-checkin	Start-of-session orientation in ~1k tokens: pending work, unverified changes, errors	Orient before you act

Why these six

Each one exists because we watched an autonomous agent fail exactly this way in production, more than once:

1. Diagnose before you code — a leftover monitor is more often the bug than the silence it reports. /fix-errors forces the triage step before any fix.
2. Verify the deliverable, not the log — "deployed successfully" and "the endpoint returns 200" are different claims. /verify-endpoint checks the second one.
3. Never guess column names — a wrong column name against SQLite doesn't warn, it crashes the session. /db-query-safe reads PRAGMA table_info() first.
4. Log before you touch — every production edit gets an audit row up front, so a regression is traceable to the commit that caused it.
5. Orient before you act — re-discovering "what's pending, what broke, what's unverified" from scratch every session burns tokens for no reason. /session-checkin compresses it.
6. Report evidence, not prose — each skill returns a structured pass/fail with a quoted excerpt, never "it seems to be working."

They're plain markdown with YAML frontmatter — read one before installing, they're short:

---
description: What this skill does — shown in autocomplete
argument-hint: "<url> [expected-status]"
allowed-tools:
  - Bash(curl *)
  - Read
---

# Instructions Claude Code follows when you run /skill-name
...

If you run Claude Code (or any agent with a similar skills/slash-command mechanism) unattended for any length of time, these six patterns are the ones that keep paying for themselves. Repo's MIT-licensed — stars, forks, and PRs adding your own failure-pattern skills all welcome:

https://github.com/ForeverTools/agent-skills