It reaches for humanity's best thinking and widens your own, first — then solves the problem in front of you.
🌐 Language: English (this page) · 中文
wisdom-first is an Agent Skill — a portable capability that any AI agent can load. Before your assistant improvises an answer to a question that deserves real thought, it reaches for the best thinking humanity has already done on that kind of problem: it names two or three complementary frameworks, recommends the best books, distills them, and only then solves your problem through those lenses.
The easy path with AI is to hand over your judgment and take the answer on faith. You get output — but you're left knowing what without knowing why, a little more dependent each time.
wisdom-first does the opposite. It names the actual frameworks, distills the moves that matter most, and points you back to the source books — so every answer doubles as a lesson, not just a verdict for today. The AI does the retrieval and the synthesis; you keep the wisdom.
That's the whole bet — an AI that makes you wiser, not one that makes you dependent.
Easiest: just ask your agent to install it. Send it one line:
Install this skill for me: https://github.com/seacen/wisdom-first
Works with any agent — including the ones the npx route below can't reach. Then start a fresh session.
Or use the command line:
npx skills add seacen/wisdom-firstWhen you ask for advice, or take on a substantive task that rewards real thinking — strategy, a speech, high-stakes writing, a negotiation, a conflict, a hard decision, a design — wisdom-first:
- Surfaces two or three complementary frameworks + the best books for your specific need;
- Distills their essence and recommends them for deeper reading;
- Solves your problem by weaving those lenses together, so the answer visibly comes from the wisdom, not from improvisation.
It triggers on its own for advisory / long-horizon work, or call it explicitly with /wisdom-first.
The skill contains no lookup table of "topic → book." Hard-coding titles would make it brittle and impossible to generalize. Instead, its core (references/the-taste.md) is pure, general taste — a handful of domain-blind judgments about what makes one body of thought better than another for a given need:
- diagnose the live need (and notice when the real block is nerve and state, not content);
- match the scope of the wisdom to the scope of the problem;
- prefer the work that defines its field, gives a method you can practice, comes from the source, and installs a durable model;
- offer two or three complementary works and weave them, rather than name-drop one.
From those judgments alone, the right canonical works fall out of the model's own knowledge — across domains the skill was never told about.
A few excellent books sit too quietly in a model's "prior" to be reached by even perfect general taste. references/authors-shelf.md lets the author hand-seed a few personal favorites as a tie-break only — never overriding fit, never adding a domain. It's a way to pass a point of view about what's good along to whoever installs the skill. Fork it and make it your own shelf.
The skill was never given a "topic → book" list — these picks fall out of its general taste. A few of the kinds of works it surfaces on its own:
| When you're working on… | It tends to reach for… |
|---|---|
| A presentation, and you're nervous about it | a delivery-and-presence classic (the author's shelf seeds As We Speak), alongside a "make it memorable" framework |
| A conflict with your partner, friend, or colleague | Nonviolent Communication (Marshall Rosenberg) |
| A report or email your boss must grasp fast | The Pyramid Principle (Barbara Minto) + MECE |
| Company strategy, spread too thin across products | Playing to Win (Lafley & Martin) and the strategy-as-choices cascade |
| Feeling stuck and adrift in work and life | The 7 Habits of Highly Effective People (Stephen Covey) |
| A negotiation where you fear it'll blow up | Getting to Yes (Fisher & Ury) and BATNA |
…then it distills the few moves that matter and weaves them into your actual answer.
wisdom-first was developed and validated with multi-agent evaluation workflows: each candidate version ran with-skill vs. baseline across a battery of real scenarios (speaking, strategy, a life rut, interpersonal conflict, business writing, negotiation, habit-building, a zero-downtime technical migration), each scenario 3× for stability, graded on whether it recalled the right canonical work, sized the wisdom to the problem, and actually applied the frameworks instead of name-dropping them. Rules that sounded wise but hurt the data (e.g. "always prefer the work that re-frames the problem") were caught and reverted. The taste you see is what survived the evidence.
skills/wisdom-first/
├── SKILL.md # trigger + the 3-step sequence + output shape
└── references/
├── the-taste.md # the core: 8 pure, domain-blind selection judgments
├── authors-shelf.md # optional personal-taste tie-break layer
└── applying-via-workflow.md # escalating the "solve" step to a multi-agent Workflow
.claude-plugin/plugin.json # Claude Code plugin manifest
brand.md # brand definition — mark, color, type, voice
assets/logo/ # logo assets (SVG + PNG) + usage spec
evals/
└── evals.json # acceptance scenarios
MIT © 2026 Seacen Zhao
- Built as an Agent Skill — the open format for extending AI agents.
- Installable through the open
skillsCLI by Vercel Labs.