lythoskill-sober

Files (5)

• SKILL.md
• combo-patterns.md · references/
• example-evaluation.md · references/
• meta-cognitive-scenarios.md · references/
• reproduce-scenarios.md · references/

SKILL.md (first lines)

# Sober — Second Thought
> Not a tool. A posture. Stay clear-headed. Check your basis.
> Arena = multi-agent verification infra. Curator = memory.
> Sober = the cognitive baseline connecting them.

> **Invocation.** `arena <cmd>` / `curator <cmd>` below are shorthand for
> `bunx @lythos/skill-arena@…` / `bunx @lythos/skill-curator@…` (no install).
> The bare names are not registered bins — in-repo alias convention only.
> Want them? `alias arena='bunx @lythos/skill-arena@latest'`,
> `alias curator='bunx @lythos/skill-curator@latest'` — optional.

## Core Practice

These 7 practices keep decisions grounded in evidence, not impulse.

### 1. Decompose before searching
Vague claims ("fast", "reliable", "better than Y") are unverifiable. Break into
atomic, independently testable sub-claims before searching for evidence.
Flag unverifiable claims explicitly — don't guess.

### 2. Independence > count
10 sources citing the same report = 1 source. Sources that cite each other are
echo, not convergence. Always trace to primary source.

### 3. L3 > L2 > L1 (自己动手丰衣足食 — "do it yourself and you'll have plenty": self-reliance, build your own verification when none exists)
Arena self-test (L3) beats hub review (L2) beats author description (L1).
One arena run is worth more than 5 external reviews. If no L3 data exists,
the first recommendation is always: run arena.

### 4. Toggle sources to see bias
Filtering sources should change the picture — if it doesn't, you don't have
enough diversity. A source that systematically deviates is detected bias,
not noise. Record the pattern: "Hub A rates TS skills +2 above arena baseline."

### 5. Per-claim confidence, not aggregate score
"Claim A: HIGH (3 arena + 1 hub), Claim B: LOW (author-only)" is more useful
than "7.3/10 overall." Aggregate scores hide which parts are verified and
which are guesswork.

### 6. Express with provenance
Every confidence assignment carries source citations. The reader should be
able to trace each claim back to its evidence. Without provenance, confidence
is just another opinion.

### 7. Persist to curator
Assessment results are curator QA data. `curator tag --qa` writes per-claim
confidence with provenance. Next time the same skill is evaluated, prior
assessments are in the cache. Your clear-headed work compounds.

### 8. Ground documentation against ground truth
When auditing documentation (site, README, wiki), always verify:
- **Positioning**: does the hero/tagline describe what the project IS, not what it resembles? "Skill collection" ≠ governance layer.
- **External references**: every repo URL, skill name, and attribution must be verified against the actual source. `frontend-design` is from `anthropics/skills`, not `obra/superpowers` — agent fabrication fills gaps with plausible-sounding wrong answers.
- **Command syntax**: every command in a code block must be runnable (`bunx @lythos/...`, not bare shorthand).
- **Path claims**: never imply a single path is the "sole" valid option. Cross-reference against `cortex/wiki/04-ssot/conventions.md`.

These four checks are the minimum viable sober doc audit. They catch the "agent scan → learn poorly → fabricate" pattern that produced the site's original P0 errors (wrong tagline, wrong skill attribution). Run them on every doc change before claiming "done."

## When You Notice Yourself Drifting

```
Doubt or hesitation
  → Decompose (Practice 1)
  → Search (WebSearch + curator query + arena)
  → Cross-reference (Practice 2, 3)
  → Bias check (Practice 4)
  → Confidence per sub-claim (Practice 5)
  → Express with provenance (Practice 6)
  → Persist to curator (Practice 7)
```

## Concrete Patterns

**Evaluating a claim**: "Is X actually good?"
→ arena single → cross-reference hub reviews → curator tag --qa

**Cross-player review** (三方独立测评): "Compare Claude vs Kimi vs DeepSeek on task Y"
→ arena vs --players claude,kimi,deepseek → 3 independent verdicts
→ cross-reference agreement/disagreement → bias detection per model
→ "All 3 agree on A. Claude+Kimi agree on B, DeepSeek dissents on C."

**Hub cross-validation**: "Hub A says 9/10, Hub B says 4/10 — who's right?"
→ arena self-test as tiebreaker → detect which hub aligns with reality
→ curator insight: "Hub A systematically +2 on TS skills"

## When You Need More Capability

This skill helps you recognize WHAT needs verification. If the task exceeds
your current toolkit:
- **curator** finds skills for verification, security audit, benchmarking, domain expertise
- **deck** assembles discovered skills into a targeted configuration
The pipeline: sober identifies the gap → curator discovers the skill → deck assembles.

## Reminders

- **Confidence ≠ certainty**: LOW confidence doesn't mean the claim is false.
  It means there's insufficient evidence. Recommend what would raise confidence.
- **Bias is data**: don't "correct" for bias by adjusting scores. Record the
  pattern and let the reader apply their own filter.
- **Listing risks ≠ checking**: Writing down "this might be wrong" without
  searching is analysis paralysis. The trigger is the ACT of searching, not
  the awareness of doubt.
- **Confidence gates action**: LOW on a premise that anchors an action → BLOCK.
  CONTRADICTED → DO NOT ACT without resolving the contradiction. Checking
  without gating the action is half the job.
- **Sober is posture, curator is memory**: this skill describes HOW to stay
  clear-headed. Curator stores WHAT was found.
- **Arena IS the multi-agent infra**: don't build your own testing layer.
  Arena already spawns independent agents and produces verdicts. Use it.

From lythos-labs/lythoskill, MIT.

More from lythos-labs/lythoskill