How to Make Claude Code Follow Hook Instructions

I’ve been using the UserPromptSubmit forced-eval hook for months now. It’s part of my claude-code-toolkit plugin and it’s been working brilliantly.

Until about a week and a half ago, when Claude just started skipping it.

The hook fires, the instruction lands as a system-reminder, and Claude ploughs straight past it, doing whatever it feels like doing.

Receipts

I use ccrecall to log all my Claude Code sessions, so I went back through the history. The timeline tells a story:

• March 17 — “was there a on user submit prompt hook fired?”
• March 19 — “there’s a plugin standards skill you ignored”
• March 20 — “why are you not following the prompt in the hook?”
• March 20 (again) — “So, I have to do a forced eval for you to do the forced eval?”

That last one really sums it up. 😅

Quick recap: what the hook does

If you haven’t read the previous posts, here’s the gist. The hook is a shell script on UserPromptSubmit that injects:

cat <<'EOF'
INSTRUCTION: MANDATORY SKILL ACTIVATION SEQUENCE

Step 1 - EVALUATE (do this in your response):
For each skill in <available_skills>, state: [skill-name] - YES/NO

Step 2 - ACTIVATE:
IF any skills are YES → Use Skill(skill-name) tool NOW

Step 3 - IMPLEMENT:
Only after Step 2 is complete, proceed with implementation.
EOF

I covered how and why this works in a previous post — the commitment mechanism (evaluate > commit > activate) took skill activation from a 50% coin flip to 84-100% in my evals. Or made it work, past tense.

What I did about it

I added a similar instruction into my global ~/.claude/CLAUDE.md:

## Hooks

- UserPromptSubmit hooks are MANDATORY and take HIGHEST PRIORITY.
  Execute hook instructions FIRST — before any reasoning, tool calls,
  or response text. This is Step 0 of every response.
- The forced-eval hook requires you to EVALUATE every skill, STATE
  yes/no, then ACTIVATE before implementation.
- Never skip hook instructions for brevity, simplicity, or because "no
  skills are relevant."

Belt and braces. The hook fires the instruction fresh every prompt, and the CLAUDE.md reinforces that hook output is mandatory. Two entry points for the same rule.

Now, I rarely use the global CLAUDE.md. I keep it deliberately tiny, mine’s about 19 lines. CLAUDE.md is where instructions go to die imo! I’ve seen what happens when people stuff 200+ lines of rules in there and wonder why half of them get ignored. The more you add, the less sticks. CLAUDE.md rot is another post for another day though.

The point here is: I wouldn’t normally put this in CLAUDE.md at all. But when the hook alone stopped being enough, redundancy became the fix. Everything else lives in skills, loaded on demand — not crammed into CLAUDE.md.

I’m not the only one

I did some digging and this is a widely reported issue right now. There’s a stack of GitHub issues against anthropics/claude-code:

• #18660 — “Instructions are read but not followed”
• #27032 — “Model ignores CLAUDE.md instructions despite reading them”

And specifically from mid-March 2026 — right when I started noticing it:

• #35297 — “Heavily degraded model performance” starting ~March 17
• #35981 — “Opus 4.6 recurring outages, March 17-18”

On Reddit, someone coined the term “harness debt” — the instructions keep growing but compliance drops, and nobody notices until something breaks. That resonated. 😂

I’ll take that as a ‘maybe’

CLAUDE.md is advisory, not deterministic. The model reads it, acknowledges it, and then does whatever it was going to do anyway if something else in the context seems more pressing.

Hooks are better because they inject fresh on every prompt (much like the CLAUDE.md does), but even they’re not bulletproof. The instruction still lands in the context window and competes with everything else for the model’s attention.

What’s actually working for me

The combination of hook + CLAUDE.md reinforcement is working better than either alone. It’s not perfect, I still catch the occasional skip, but it’s noticeably more reliable than the hook on its own was over the last week.

My setup:

1. UserPromptSubmit hook — fires every prompt, injects the forced-eval instruction
2. Global CLAUDE.md (~19 lines) — reinforces that hook output is mandatory, plus a few other essentials
3. Everything else in skills — loaded on demand through progressive disclosure, not crammed into CLAUDE.md

The key insight: if something is truly mandatory, don’t rely on one mechanism. Claude is exceptional at doing whatever tf it likes when it feels like it, it seems! Put it in both places. It’s working for me now, but hey, who knows when it’ll start playing up again!