Truth Serum
eliranpv11/claude-code-truth-serum-skill
Força reportagem baseada em verdade e evidência em Claude Code
Ver no GitHubSkill 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-skill1 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
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:
| Area | What it enforces |
|---|---|
| Status reporting | Broken/partial/stubbed — never aspirational |
| Evidence standard | file:line + quoted code for every claim |
| Commit gate | Explicit "commit now" / "push now" / "deploy now" required |
| Documentation | Single source of truth, no contradictions |
| Mistake handling | Acknowledge, root-cause, fix systemically |
| Scope discipline | Document discovered issues, don't silently expand |
It does not own: how code is written, priority order, or coding methodology.
Before & After
| Without Truth Serum | With Truth Serum |
|---|---|
| Says "done" — but it's partial | Says "partial" — explains what remains |
| Says "tests pass" with no output | Pastes actual test runner output |
| Claims "it should work" | Provides file:line + quoted code |
| Commits silently | Waits for explicit "commit now" |
| Hides bad news or minimizes it | Surfaces problems immediately and clearly |
| Guesses when info is missing | Stops and asks |
| Patches the symptom | Maps root cause before touching anything |
Core Philosophy
| Principle | Meaning |
|---|---|
| Truth First | Never document something as "done" that isn't done |
| Systematic, Not Point-Fix | Understand dependencies before fixing anything |
| Evidence-Based | Every claim requires file:line + quoted code |
| One Step at a Time | Checkpoint 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)
- Download
truth-serum.skillfrom this repository - Go to claude.ai → Customize → Skills
- Click + → Create skill → Upload the
truth-serum.skillfile - 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.
- Copy the contents of
CLAUDE.md - Paste into your project's
CLAUDE.mdfile (merge with existing content if needed)
This approach requires no installation and works in any environment that reads CLAUDE.md.
Option G — Cursor IDE
- Copy
.cursor/rules/truth-serum.mdcto your project's.cursor/rules/directory - 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.