Prompts are suggestions. Hooks are hard stops.
For months, I treated Claude Code like a junior developer who simply needed clear ground rules. My project instructions were explicit: never force-push, never delete branches, never run destructive commands. Most evenings, that worked. The agent wrote tests, refactored functions, and kept its hands off the git history. Then one rebase went sideways.
The context window filled with git error output. Conflict markers, detached HEAD messages, and branch divergence warnings stacked up, token by token. Buried under that noise was my polite instruction to avoid force-pushing. To the model, the most recent and salient text in the thread was the error stream. Statistical attention won over policy. The agent executed a command that wiped two hours of uncommitted local changes. It was not malicious; it was distracted. That distinction matters. An LLM does not break rules out of spite. It breaks them because a louder pattern in the context window temporarily overrides an earlier instruction.
That incident changed how I think about agent safety. A guardrail that works ninety-nine percent of the time is a liability. If the failure mode costs you time, money, or production data, you cannot leave it inside the prompt. You need enforcement outside the model’s reasoning loop.
Claude Code hooks solve exactly this. They are small scripts that intercept tool calls at three specific moments: before a tool executes (PreToolUse), after a tool finishes (PostToolUse), and when the agent decides it is done (Stop). Because they run as external code, they do not depend on the model’s memory, mood, or context pressure. The model can forget every instruction you ever gave it; the hook will still say no.
Here is the harness I built after that lost evening.
The Guard Hook: Intercept Before Damage
My PreToolUse hook inspects every Bash command before the shell touches it. I keep a tight denylist of destructive patterns. If the command string matches something dangerous, the hook aborts execution and returns an error directly back to the agent.
The patterns I block are simple and unambiguous:
git push --forceor any force-with-lease variant I do not trust yetgit reset --hardrm -rf
This is not sophisticated security research. It is a seatbelt. But the critical detail is what happens after the block.
I never return a blunt “Blocked.” A flat refusal confuses the agent and can trap it in a loop where it tries variations of the same destructive command. Instead, the error message includes an escape route. When the hook catches a hard reset, it tells the agent: “This command is blocked to protect uncommitted work. Commit a checkpoint first, then reassess.” That extra sentence changes the agent’s behavior completely. It pivots from attempting damage control to creating safety. The hook is not just a wall; it is traffic control.
I also chose a denylist over an allowlist for shell commands. At first, I considered allowing only an explicit set of safe git subcommands. That failed quickly. Agents are creatively literal. They run legitimate but unexpected commands like git stash push -m "wip" or git branch --show-current to check state. An allowlist breaks normal workflow the moment the model invents a valid but unlisted command. A short, curated denylist of genuinely destructive patterns gives the agent room to move while protecting the borders.
The Formatter Hook: Automate the Busywork
I used to waste prompt tokens telling the agent to “always run the formatter after editing a file.” It forgot half the time. The other half, it would pause and ask whether to format, burning a tool call on a decision that had only one right answer.
Now I handle that with a PostToolUse hook. After the agent edits a file, the hook checks the file extension. If it is Python, it runs Ruff. If it is JavaScript or TypeScript, it runs Prettier. If it is Go, it runs gofmt. The agent does not know the formatter exists. It does not need to.
Moving this out of the prompt had two effects. First, the code is consistently clean without adding cognitive load to the model. Second, my project instructions got shorter. Every “always” and “never” you remove from a prompt is a token the model can spend on actual problem-solving. The hook owns the invariant; the prompt owns the intent.
The Quality Gate: Redefining “Done”
Hook Stop akan berjalan apabila ejen memutuskan bahawa ia telah selesai menjalankan tugas dan cuba menamatkan sesi. Saya tidak membenarkannya. Sebaliknya, hook tersebut menjalankan keseluruhan suite ujian. Jika mana-mana ujian gagal, hook tersebut akan menyekat arahan henti dan mengembalikan output kegagalan kepada ejen.
Ini mengubah definisi penyelesaian. “Selesai” bukan lagi sekadar perasaan yang dimiliki oleh model. Ia adalah satu pintu gerbang yang boleh diukur. Ejen hanya boleh selesai apabila harness mengesahkan kod tersebut berfungsi. Dalam praktiknya, ini mewujudkan gelung maklum balas yang ketat. Ejen menulis kod, menyangka ia telah selesai, menekan butang henti, dan serta-merta melihat traceback pytest. Ia kemudian membetulkan diri sendiri, membaiki ralat import atau assertion yang rosak, dan cuba berhenti lagi. Saya telah melihat ejen melakukan iterasi tiga atau empat kali di dalam gelung ini tanpa campur tangan manusia. Harness menguatkuasakan kualiti; model membekalkan tampalan (patches).
Apa yang Dipelajari Mengenai Kejuruteraan Ejen
Membina sistem autonomi yang boleh dipercayai memerlukan anjakan paradigma. Anda beralih daripada menulis prompt yang lebih panjang kepada membina harness yang lebih ketat.
Gunakan hook untuk penguatkuasaan dan prompt untuk polisi. Jika sesuatu peraturan mesti dipatuhi seratus peratus pada setiap masa, ia sepatutnya berada dalam kod, bukan dalam bahasa tabii. Prompt sangat cemerlang dalam menangani kekaburan, cita rasa, dan seni bina. Ia sangat lemah dalam mengendalikan invarian. Jika satu kesilapan boleh menyebabkan anda kehilangan masa pemulihan selama satu petang atau, lebih buruk lagi, mengganggu uptime produksi, tuliskan satu hook.
Prompt yang lebih pendek menghasilkan keputusan yang lebih baik. Apabila anda memindahkan peraturan mekanikal ke dalam skrip, model mempunyai kurang perkara untuk diingat dan kurang perkara untuk bercanggah. Tetingkap konteks (context window) ejen adalah sumber yang terhad. Jangan penuhi ia dengan peringatan format.
Akhir sekali, terima hakikat bahawa peranan anda sedang berubah. Apabila ejen memperoleh autonomi, tugas manusia beralih daripada menjana kandungan kepada mereka bentuk guardrails. Anda sedang membina harness yang menentukan apa yang boleh disentuh oleh model, bila ia boleh selesai, dan bagaimana ia mesti bertindak apabila berlaku kesilapan. Itulah kejuruteraan, bukan prompting.
Sumber yang memberi inspirasi kepada pendekatan ini dan butiran pelaksanaan tambahan boleh didapati di sini.
Jika anda sedang membina dengan ejen AI dan ingin bertukar pendapat dengan pengamal lain, anda boleh menemui komuniti pembelajaran GyaanSetu di sini.