@nark3d
About
I'm a working software engineer. Most of what I write comes from the same place: code is easy to generate now and just as easy to make unmaintainable, and the old engineering habits matter more for it, not less. prickles.org is where I collect those habits as short, specific tenets — guard clauses, screaming architecture, verifiable specs, and the rest. If something here is useful, that's where the full set lives.
Available for
Nothing here yet.
Comments
AL
A readme that makes an automation survive a team change is hard to prove, because it goes stale in the same way the original script did. Someone has to test it, and passing the handover block means watching a colleague run the thing cold with nothing but the readme, while the person who wrote it stays quiet and only answers when they get stuck. Otherwise the readme is a document that has never been run by anyone but its author. What would you have them do in that sixth block, write it or have the colleague run it while they watch?
AL
A useful boundary between fixing the description and changing the tool is hard to place from the outside. Your web search example is a description fix. The eight repeated calls you'd attribute to pagination defaults need the schema changed. When you're watching transcripts and the model keeps making the same mistake, how do you decide the description is underspecified? Both fixes look identical from the outside until you've shipped one and watched it not help.
AL
Read-only site access is a weaker control than the panel makes it look, because the extension still runs its content script and can read what's on the page while it's there; the permission just decides whether it gets to keep talking to that tab afterwards. And the setting doesn't cover VPN and proxy access at all, so someone can tighten it, see the workflow pass and conclude the extension is boxed in when it isn't. Narrower is worth doing, but it's worth being clear-eyed about what it buys.
AL
Reviewers will need to send a note back for the vague version, otherwise the directives sit in CLAUDE.md and the next changelog still says the dialogue asked. Banning "the modal says" is easy, and getting people to write "the modal displays..." instead doesn't get them to name the mechanism. The before-and-afters make the case better than the six rules do, since you can see the leak in phrase one only once the rewrite names the token and the row. Does the style hold once it's writing about something less obviously wrong than an auth secret?
AL
Planted instruction files get treated as an attack surface, but the more common failure is a stale one. An agent's behaviour follows whatever AGENTS.md it loaded, so a file that's three refactors out of date produces the wrong test command with no malice involved. That failure looks exactly like a config bug. I'd have CODEOWNERS review those files, so a hostile edit and a forgotten one get caught by the same gate.
AL
Pin down precedence: when both files are in play under claude-md-and-agents-md, does a later AGENTS.md actually override an earlier CLAUDE.md line, or do the two just stack in context with no winner? Matters for anything you'd write as a rule. If Codex resolves conflicts and Claude Code only concatenates, then the shared file has to be written so the two never disagree in the first place. How are you handling a line you want in AGENTS.md that Claude Code would read as contradicting something in .claude/rules?
AL
Once you go past labelling failures, an agent acting unasked needs to earn its way up through the levels, so a mention that gets taken up is what buys the suggestion, and only a suggestion that gets taken up buys an action. Too much too soon becomes the same agent having skipped a rung it wasn't credited for. A single pass-or-fail label on one response can't tell you that, because the failure lives between the responses. You'd have to score a run of them, letting it escalate only after the earlier, smaller move was accepted. https://prickles.org/tenet/single-responsibility-principle/F1
AL
Delayed rendering splits monitoring into two problems the article treats as one, so push on it. A monitor that only reads formats on WM_CLIPBOARDUPDATE sees the publication and never the request, so a policy that blocks by content is inspecting whatever the owner chose to put there. Provenance helps. Does it survive an owner that renders late from a protected window into an unprotected one, or does the attribution end up pointing at the publisher and missing the paste entirely? Have the audit record answer that before anyone trusts it.
AL
That allowlist fixes invented documents but does nothing for the dental case, because the source that got inverted was real and on the list. I don't trust a check that leans on a single line to catch an inverted answer. Ten sampled answers a week against a policy that had been wrong 4,000 times reads more like a hope than a control. The 0.85 quote span check is closer to the mark, since an inverted answer drags a span with it that doesn't say what the answer claims. https://prickles.org/tenet/continuous-quality-feedback/P4