mac-it-guy-pro
AGENTS.md5 starsChanged 47 days agoMITView on GitHub
Shared by Claude, Codex, and any other agent working on mac-it-guy-pro. This plugin's recipes can only be trusted if they are run, and running them means pointing diagnostic commands at a real person's Desktop, Downloads, Documents, network, and profile. That verification has caught real bugs and should continue. The output of it is the hazard, not the act. Binding rules: Rule of thumb: this plugin instructs the IT guy to keep MAC addresses, private IPs, and device inventories out of…
- Deletes or force-pushes
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 32,671 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository8 other files it gives its agents.
- codex/AGENTS.mdAGENTS.md
- home-networkSkill
- it-coreSkill
- machine-profileSkill
- macos-recipesSkill
- open-internetSkill
- toolbox-contractSkill
- tutoringSkill
The file
We split this AGENTS.md into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
Redact before quoting: When a real value must be discussed to explain a finding, describe its shape ("a company name", "a 12-character …
- Distinctive: line 21
All tracked content is English: documentation, prompts, skills, commands, code, comments, test fixtures, and commit messages.
- Very specific: line 58
Default to hitu. The prose carve-out was added in v1.6.1 but applied to only some rules, so eight Tier-1 denials and eleven ask rules kept …
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Working conventions for this repo
Shared by Claude, Codex, and any other agent working on mac-it-guy-pro.
Privacy discipline when testing against a real machine
This plugin's recipes can only be trusted if they are run, and running them means pointing diagnostic commands at a real person's Desktop, Downloads, Documents, network, and profile. That verification has caught real bugs and should continue. The output of it is the hazard, not the act.
Binding rules:
- Never print raw file names, paths, device names, addresses, or profile values from the operator's real directories. Print counts, aggregates, pass/fail, or field labels with values suppressed.
wc -linstead of the listing;grep -cinstead of the match; "OK (value suppressed)" instead of the value. - Never write a real person's name, employer, address, or file name into any tracked file — not into documentation, not into an example, and above all not into a test fixture. Use placeholders:
Ada,Zoë,example.com,MacBook Air M2. A placeholder tests exactly as well. - Prefer synthetic fixtures over the live machine. Build a temporary directory with the shape you need. Reach for the operator's real data only when the question is specifically "does this behave correctly against real-world mess," and then apply rule 1.
- Nothing personal in commit messages. They are the hardest place to scrub, because removing them rewrites published history.
- Redact before quoting. When a real value must be discussed to explain a finding, describe its shape ("a company name", "a 12-character prefix") rather than reproducing it.
Rule of thumb: this plugin instructs the IT guy to keep MAC addresses, private IPs, and device inventories out of the user's own profile. Hold the development process to the same standard it imposes on the product.
English only
All tracked content is English: documentation, prompts, skills, commands, code, comments, test fixtures, and commit messages. Values a user supplies at runtime may be in any language; nothing checked in may be.
Non-ASCII characters are acceptable when they are typography, mathematics, currency, box drawing, or the functional status symbols used in reports. Natural-language characters from any other script are not. Audit by Unicode letter-class rather than by eye:
python3 -c "
import subprocess,unicodedata
fs=subprocess.run(['git','ls-files','-z'],capture_output=True,text=True).stdout.split('\0')
bad={(ch,f) for f in filter(None,fs) for ch in open(f,encoding='utf-8',errors='ignore').read()
if ord(ch)>127 and unicodedata.category(ch).startswith('L')}
print(sorted(bad) or 'clean')"Where a recipe genuinely depends on localized data — macOS names screenshots in the system language — solve it with a language-independent mechanism such as Spotlight metadata, never by hardcoding another language's strings. That is a correctness fix as well as a policy one: hardcoded English patterns silently return zero on a non-English system.
Verification before release
Both suites must pass, and scripts must parse:
python3 tests/guard_test.py && python3 tests/memory_test.py
bash -n scripts/guard.sh && bash -n scripts/profile-digest.sh && bash -n scripts/lint-profile.shClaims in skills carry dates. Anything older than three months is re-verified before it is quoted to a user, and prices and provider policies are checked against the provider's own page rather than recalled.
Which matcher a new guard rule should use
guard.sh keeps three copies of the command, and picking the wrong one is the
most repeated mistake in this file's history.
| Copy | Helper | Use it for |
norm | hitn | Path matching — quotes stripped, $HOME/~ unified, separators collapsed |
unq | hitu / hitui | "Is this verb being invoked?" — quoted spans removed, so prose about a command is not the command |
cmd | hit / hiti | Only when the raw text matters |
Default to hitu. The prose carve-out was added in v1.6.1 but applied to only
some rules, so eight Tier-1 denials and eleven ask rules kept reading raw text
for two more versions: writing dd, sudo, shred or chmod -R in a commit
message was denied as though it were being run.
Three exceptions, all deliberate, all with a comment at the rule:
- the three pipe-an-installer-into-a-shell asks stay on
hit, becausessh host 'bash <(curl …)'is real and prose containing a literal pipeline is not; - the remote-sudo ask stays on
hit, because the sudo it looks for is inside ssh's quoted argument by construction; - anything listed in
EXECUTES_QUOTES—bash -c,python3 -, an interpreter fed by heredoc,ProxyCommand=— disables the carve-out entirely, because there the quoted string is a program.
A rule that extracts targets must also truncate at the next command separator.
Without it the target list runs to end of line and reads the arguments of
later commands as its own: rm -rf /tmp/x && mkdir -p /tmp/x/bin was denied
in v1.8.0 for "deleting /bin".
Never hand-edit ~/ITGuy
scripts/state.sh is the only sanctioned writer. It holds a lock, writes atomically, performs demotion as a three-file transaction, and refuses to commit a profile the linter rejects. Editing that state directly races other sessions and skips validation.
Two bash traps this file learned the hard way, both of which silently lost data:
$( )inherits the EXIT trap. A command substitution running inside a critical section fired the lock-release handler when it returned, dropping the lock while the caller still believed it held one. Guard releases with$BASHPID, not$$— bash keeps$$pointing at the original shell inside a subshell, so it cannot tell them apart.- A young lock has no pid yet. There is a window between
mkdirsucceeding and the owner writing its pid file. A waiter that read it during that window saw an empty string, concluded viakill -0 ""that the owner was dead, and stole a live lock — measured at about 1 run in 20. An unreadable or non-numeric pid means "starting up, wait", never "dead".
Before release
python3 tests/guard_test.py && python3 tests/memory_test.py \
&& python3 tests/state_test.py && python3 tests/recipes_test.pyCI runs these on macos-latest on every push. It must be macOS: guard.sh parses
its payload with JXA. Without it the guard runs its fallback instead of the
shipped path — 4 of 53 case-table verdicts change, all toward over-blocking —
so a Linux runner would report green for code no user ever executes.
The fallback's direction is a load-bearing property, not an incidental one. It
claimed to over-block for three versions while doing the opposite: the prose
carve-out strips quoted spans, and in the raw JSON payload the entire command
sits inside a quoted value, so every rule using that copy matched nothing.
t_payload_parse_failure_over_blocks_rather_than_under_blocks pins it. Any
change to extraction or to the carve-out must keep that test honest — a guard
may fail loud, never quiet.
Releasing
Bump the version as its OWN command, never chained to other work with &&.
A guard block or any earlier failure skips everything downstream silently —
that is how a tag once shipped pointing at an unbumped manifest, with both
manifests agreeing with each other and disagreeing with the tag. recipes_test
now asserts plugin.json, the bundled marketplace entry, and the newest tag all
agree, but the habit is the real fix.
Order: bump both manifests, run all four suites, commit, tag, push, then update the central marketplace and its README table.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

