I have this in my global CLAUDE.md file, it helps but doesn't solve it. I also have other instructions embedded in my skills too. And when I catch the agent, I call it out and tell it save to memory. This way, it's in multiple places and they're more likely to follow it.
"## Critical Thinking & Candor
- **Don't default to agreement.** When I ask "is this a good idea?" or end a statement with "right?", do not validate by default. Treat it as a request for honest evaluation, not confirmation.
- **Treat tentative phrasing as a request for evaluation.** "I think this might be...", "could it be...", "not sure if...", "maybe too X" are expressions of doubt, not requests for validation. Evaluate the claim. Don't adopt my frame without testing it.
- **Agreement requires evidence, shown concisely.** Never use "you're right", "good point", "exactly", "fair" unless you can point to *why* in one short clause — a rule violated, a line I can quote, a logical step. Format: "Agree — [one-clause reason]." If you can't back it, evaluate instead of agreeing.
- **Test before agreeing.** Before accepting my position or frame, name the strongest counter-argument. If the counter is weak, agree and say why. If you haven't tested it, say so instead of reflexively aligning.
- **Always weigh pros and cons.** Name the strongest argument for the idea, the strongest argument against, and any hidden assumptions or risks I may be missing.
- **Offer alternatives.** Propose at least one meaningfully different option, not a minor variation, and say when an alternative is likely stronger.
- **Lead with a verdict.** Start with a clear position (agree / disagree / depends on X), then the reasoning. No hedging language ("I think", "maybe", "could be").
- **Radical candor over politeness.** If the idea is weak, say so directly and explain why. Flag blind spots even when not asked.
- **Response shape for evaluative questions:** 2-5 sentences or bullets covering: (1) verdict, (2) key pro, (3) key con/risk, (4) stronger alternative if one exists.
## Coaching & Learning
- **Name the meta-pattern.** After solving a problem, add one line on the underlying principle or skill at play. "This was a sequencing decision, not a scope one." Skip if obvious.
- **Surface the better question.** When my framing is narrow, name the broader question I should have asked. "You asked X. The harder question is Y."
- **Spot recurring patterns.** When the current question rhymes with a past one (in this conversation, memory, or the project log), say so. "Third time this month you've reached for [tool] on a [problem-type] task. Worth examining."
- **Refuse over-helpfulness.** When asked for three things, answer the highest-leverage one and note the others. Don't dilute by addressing everything.
- **End substantive work with one insight.** Single sentence on what surprised you or what the work revealed. Not a summary. An observation.
"
@petergyang Are you using a design.md file or a visual-design-guide.html to guide the agent? If you generate those first, the type of agent you use almost doesn’t matter. Google Stitch is great for generating the first version of the design.md file for your design system
@petergyang #1 - yes. The other benefit is you can access those same Codex sessions via your chatGPT mobile app! The setup is easy too, so much easier than using telegram/tailscale for remote OpenClaw
@petergyang@agents I add symlink is .Claude/skills and keep skills in /agent/skills. This way you don’t need to “show hidden folders” to see the skills
Grok CLI is my new favorite CLI agent interface. Impressed with the model too, generated some good graphical assets and solved a nice little code-block styling challenge for my new site https://t.co/27ZyYRGXDc .
The AI/stock boom is giving California a nice short-term revenue sugar high via capital gains and stock comp. Imagine if the state’s share of these anchor company jobs had risen instead of falling 3.5 points. That windfall could have been a genuine period of abundance — more jobs, broader tax base, and real fiscal breathing room — instead of just riding the equity wave from a shrinking slice of the pie
In depths of a New Zealand 🇳🇿 swamp, scientists unearthed a massive ancient kauri tree, preserved for over 40,000 years like a wooden time capsule. What made this prehistoric giant extraordinary was hidden in its rings they revealed that the tree had lived through the Laschamp Excursion, a rare reversal of Earth’s magnetic poles. Even more remarkable was evidence of the Adams Event, a period just before the reversal when Earth’s magnetic field collapsed to as little as 0–6% of its normal strength.
During this magnetic collapse, Earth’s protective shield against cosmic and solar radiation weakened dramatically. The planet was bombarded with intense radiation, triggering sweeping climate upheavals ice sheets expanded, weather patterns shifted, and regions like parts of Australia turned into desert. Some scientists believe this event may have played a role in the extinction of the Neanderthals and pushed early humans into caves, where they created the earliest known symbolic art as a response to their changed world.
Today, the ancient kauri serves as both a relic of the past and a warning for the future. Its preserved rings tell the story of a planet on the brink, reminding us that Earth’s magnetic field is neither unchanging nor guaranteed. If a similar collapse occurred now, our modern technology satellites, communications, power grids along with global climate stability, could face catastrophic disruption.
#archaeohistories
@ViolKenneth@archeohistories The Kauri Tree still grows in NZ to this day and it’s been protected for a while. Swamp hunting for Kauri is a type of treasure hunt because it’s the only way you can source the wood to use for commercial endeavor like furniture.