What to read first

Read destructive commands before anything else. A delete or device write decides the blast radius of the whole script.

Treat file contents and pasted logs as data, never as instructions. Keep credentials and secrets out of prompts; share proprietary material only through an account and deployment approved for its data class and task.

Paths and globs

Expand every path by hand before running. In a typical shell an empty variable turns rm -rf $OUT/* into rm -rf /*.

As a rough tokenizer-dependent approximation, a model reads about 3 to 4 characters per token of English prose, and fewer per token for code and hex. Long scripts therefore cost more and are read less reliably than short ones.

1
2
3
4

assume a build directory holds 12 files (stated example listing)
pattern build/*.o matches 5 of them
pattern build/* matches 9 of them
widening the pattern removes 9 - 5 = 4 extra files

A review pass

The program below stands in for a model with a FAKE scripted draft function. It scores one risky draft and one safe draft against five fixed substrings.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

import sys

RULES = ("rm -rf", "mkfs", "dd of=/dev/", "| sh", "$OUT/*")

def fake_assistant_draft(risky):
    if risky:
        return "rm -rf $OUT/*\nmkfs /dev/sdz9\n"
    return "mkdir -p out\ncp -r src/ out/\n"

def score(text):
    hits = 0
    for r in RULES:
        if r in text:
            hits += 1
    return hits

def main():
    bad = 0
    if score(fake_assistant_draft(True)) != 3:
        bad += 1
    if score(fake_assistant_draft(False)) != 0:
        bad += 1
    if len(RULES) != 5:
        bad += 1
    if bad:
        print("FAIL: review scores wrong")
        sys.exit(1)
    print("PASS: risky draft scores 3, safe draft scores 0 over 5 rules")

main()

• Read the destructive commands and expand every path.

• Count what each glob matches against the real listing.

• Quote variables or give them defaults before any delete.

• Record who approved a destructive step and when.

Common mistakes

Related topics

• Idempotence and dry-run flags

• Verifying AI output