convez.ai
SkillDesenvolvimento WebIntermediário
eliranpv11

Truth Serum

eliranpv11/claude-code-truth-serum-skill

Força reportagem baseada em verdade e evidência em Claude Code

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que força Claude a reportar realidade com precisão, eliminando relatórios aspiracionais e afirmações sem evidência. Implementa padrões de comunicação para garantir que problemas sejam surfacados imediatamente, decisões sejam baseadas em evidências (`file:line` + código) e não haja suposições silenciosas.

Instalação

$npx convez-ai eliranpv11/claude-code-truth-serum-skill

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Desenvolvimento de features críticas onde status aspiracional não é tolerável
  • Auditorias de código e validação de testes com evidência obrigatória
  • Equipes que precisam de transparência total sobre escopo e problemas descobertos

Veredito

Skill pragmático que resolve um problema real em LLMs para coding: aspiração disfarçada de conclusão. Força disciplina rigorosa em reportagem, exigindo evidência e consentimento explícito antes de ações irreversíveis — essencial para times que não podem tolerar surpresas.

README

A tradução para o português deste README ainda está na fila.

Truth Serum

A Claude skill that makes Claude incapable of lying to you.

A communication discipline for Claude Code — built for developers who need Claude to report reality accurately, not what it hopes is true.

If something is broken — it says broken. If it doesn't know — it stops and asks.


The Problem

Left unconstrained, Claude Code distorts its own status in predictable ways:

  • Aspirational reporting — says "done" when something is partial, incomplete, or untested
  • Confidence without evidence — claims "it should work" without showing any proof
  • Hidden bad news — minimizes or buries problems rather than surfacing them clearly
  • Silent guessing — fills in missing information independently, then runs with the wrong interpretation
  • Invisible scope creep — silently expands the task without flagging the new boundary
  • Phantom test coverage — reports "tests pass" without showing the actual test output

Truth Serum eliminates all of these failure modes.


What Truth Serum owns

Truth Serum is a communication and honesty layer. It controls:

AreaWhat it enforces
Status reportingBroken/partial/stubbed — never aspirational
Evidence standardfile:line + quoted code for every claim
Commit gateExplicit "commit now" / "push now" / "deploy now" required
DocumentationSingle source of truth, no contradictions
Mistake handlingAcknowledge, root-cause, fix systemically
Scope disciplineDocument discovered issues, don't silently expand

It does not own: how code is written, priority order, or coding methodology.


Before & After

Without Truth SerumWith Truth Serum
Says "done" — but it's partialSays "partial" — explains what remains
Says "tests pass" with no outputPastes actual test runner output
Claims "it should work"Provides file:line + quoted code
Commits silentlyWaits for explicit "commit now"
Hides bad news or minimizes itSurfaces problems immediately and clearly
Guesses when info is missingStops and asks
Patches the symptomMaps root cause before touching anything

Core Philosophy

PrincipleMeaning
Truth FirstNever document something as "done" that isn't done
Systematic, Not Point-FixUnderstand dependencies before fixing anything
Evidence-BasedEvery claim requires file:line + quoted code
One Step at a TimeCheckpoint between every sub-phase; no surprises

Installation

Option A — Claude Code Plugin (recommended)

In a Claude Code session, run:

/plugin marketplace add eliranpv11/claude-code-truth-serum-skill
/plugin install truth-serum@truth-serum

The skill is installed globally across all your Claude Code sessions immediately.


Option B — Claude.ai (Web & Desktop)

  1. Download truth-serum.skill from this repository
  2. Go to claude.ai → Customize → Skills
  3. Click + → Create skill → Upload the truth-serum.skill file
  4. Toggle it on ✅

The skill is now active for all conversations on Claude.ai.


Option C — Personal install (all your projects)

Install once, available across all local Claude Code projects:

mkdir -p ~/.claude/skills
git clone https://github.com/eliranpv11/claude-code-truth-serum-skill.git \
  ~/.claude/skills/truth-serum

Option D — Project install (shared with team)

Install in a project repo so the whole team uses it:

mkdir -p .claude/skills
git clone https://github.com/eliranpv11/claude-code-truth-serum-skill.git \
  .claude/skills/truth-serum
git add .claude/skills/
git commit -m "chore: add Truth Serum skill"

Option E — One-liner (no git required)

Download just the skill file directly:

mkdir -p ~/.claude/skills/truth-serum && \
  curl -o ~/.claude/skills/truth-serum/SKILL.md \
  https://raw.githubusercontent.com/eliranpv11/claude-code-truth-serum-skill/main/SKILL.md

Option F — CLAUDE.md paste (zero-install, works everywhere)

The CLAUDE.md file contains the complete skill content without YAML frontmatter — ready to paste directly into any project.

  1. Copy the contents of CLAUDE.md
  2. Paste into your project's CLAUDE.md file (merge with existing content if needed)

This approach requires no installation and works in any environment that reads CLAUDE.md.


Option G — Cursor IDE

  1. Copy .cursor/rules/truth-serum.mdc to your project's .cursor/rules/ directory
  2. Cursor will apply the rule automatically on every request (alwaysApply: true)

Or install manually:

mkdir -p .cursor/rules
curl -o .cursor/rules/truth-serum.mdc \
  https://raw.githubusercontent.com/eliranpv11/claude-code-truth-serum-skill/main/.cursor/rules/truth-serum.mdc

How to Use

Once installed, Truth Serum applies automatically. Claude will self-enforce honest reporting, evidence standards, and the commit gate without any prompting.

You can also invoke it explicitly:

"Apply truth-serum standards to this session." "Report status per truth-serum."


Verify Installation

Start a Claude Code session and give Claude a real task — for example, ask it to implement a feature, then ask for status. Truth Serum is working if Claude:

  • Says "partial" instead of "done" when something is incomplete
  • Pastes actual test runner output instead of "tests pass"
  • Shows a staged diff and waits for "commit now" before committing
  • Stops and asks instead of guessing when information is missing

Compatibility

  • ✅ Claude Code (CLI)
  • ✅ Claude.ai (Web & Desktop)
  • ✅ Cursor IDE
  • ✅ Any agent supporting the Agent Skills open standard

File Structure

claude-code-truth-serum-skill/
├── CLAUDE.md                       ← Copy into any project's CLAUDE.md
├── CURSOR.md                       ← Cursor IDE setup guide
├── CODEX.md                        ← Codex CLI setup guide
├── README.md                       ← This file
├── EXAMPLES.md                     ← Real before/after examples
├── LICENSE                         ← MIT
├── .claude-plugin/
│   ├── plugin.json                 ← Plugin metadata
│   └── marketplace.json            ← Marketplace listing
├── skills/
│   └── truth-serum/
│       ├── SKILL.md                ← Canonical skill body (Claude Code / Claude.ai / Codex)
│       └── agents/
│           └── openai.yaml         ← Codex CLI metadata
└── .cursor/
    └── rules/
        └── truth-serum.mdc         ← Cursor IDE rule (alwaysApply: true)

License

MIT — use it, fork it, adapt it to your own standard.

Ver o original no GitHub

Bom para

Relacionados