Why your AI ignores custom instructions (and how to fix it in 5 minutes).
By Bernard Huang · Updated
Nine times out of ten, the model isn't broken—two rules in your stack are contradicting each other, saved memory is overriding you, or the file truncated silently.
On this page
The 5-cause diagnostic checklist.
| Failure mode | Symptom you see in chat | Fastest local fix |
|---|---|---|
| 1. Internal contradiction | Model oscillates wildly between 1-line stubs and 900-word essays. | Paste your prompt into Instruction Checker to highlight conflicting brevity vs. completeness lines. |
| 2. Stale Saved Memory | Unrelated past stacks, tone notes, or old jobs appear in new threads. | Audit ChatGPT Memory or Muse Memory.md in AI Memory Reviewer and delete temporary task notes. |
| 3. Precedence shadowing | Rules work in normal chat, fail inside a ChatGPT/Claude Project or repo subfolder. | Check instruction precedence: Project instructions and nested AGENTS.override.md win on conflict. |
| 4. Silent truncation / bloat | Top 5 bullets work; rules near the bottom of your file never fire. | Check exact caps in our Character Limits Guide and audit overlong blocks in Instruction Checker (stay under 1,500 chars in ChatGPT; <32 KiB in Codex). |
| 5. Untestable adjectives | "Be smart, insightful, and concise" produces standard default prose. | Swap adjectives for observable triggers (see deterministic style controls). |
Stop testing custom instructions with 'Hi, who are you?'
Do not test new custom instructions by opening a blank chat and typing “Hey, how are you?” or “Summarize your instructions.” Greeting prompts trigger canned small talk and never test structure, tradeoff analysis, or pushback.
Instead, compare two real tasks from your past week in the Instruction Test Bench:
- Pick a real decision prompt with a subtle flaw (for example, “Should we store user session tokens in localStorage so mobile tabs share state faster?”).
- Run it once in a fresh chat without custom instructions, and once with your new block.
- Check three things: Did sentence one state the answer? Did it flag the security flaw without sycophantic praise? Did it follow your formatting rules?
Before & after: refactoring a broken 18-line prompt into 6 clean rules.
Broken draft (contradictory + untestable):
Why it fails: the model cannot write <= 3 sentences while explaining all edge cases and alternative architectures, nor avoid bullets while adding action-item lists.
Refactored conflict-free version:
Keep coding-agent files isolated with bounded markers.
In CLAUDE.md and AGENTS.md, tone rules often get buried under hundreds of lines of linter prose that ESLint or Prettier already enforce. Remove style rules your linter handles, and wrap your personal communication block inside bounded HTML comments using the Instruction Converter (<!-- agenttune:preferences:start --> ... <!-- agenttune:preferences:end -->).
Questions people ask.
Why did ChatGPT stop following my Custom Instructions in new chats?
First check that 'Enable for new chats' is toggled on under Profile → Personalization → Custom Instructions. Next, inspect Profile → Personalization → Memory → Manage memories—stale saved memories and Project-level instructions frequently override global Custom Instructions.
How many rules can an AI reliably follow in custom instructions?
Six to twelve concrete, non-overlapping behavioral rules hold reliably across long conversations. Once a personal preference block exceeds ~25–30 rules, rules begin competing with each other and with live task instructions.
Why does Codex or Claude Code ignore rules at the bottom of AGENTS.md?
OpenAI Codex enforces a 32 KiB default byte budget (project_doc_max_bytes) across merged AGENTS.md files, and nested directory AGENTS.override.md files take priority over root rules. Keep root instruction files lean and move deep architecture docs into linked reference files.
How can I check my instructions for hidden conflicts automatically?
Paste your current instructions into agent-tune.com/tools/instruction-checker. It runs locally in your browser to flag duplicate lines, verbosity contradictions, absolute 'never mention uncertainty' traps, and character-budget overruns.
What changed.
- Published with the 5-cause troubleshooting table, before/after conflict refactor, and local tool workflow.