Receber por email
← Voltar para o catálogo

fp-brief

First-principles briefing from technical documents. Use when: understanding why decisions were made, onboarding to feature reasoning, reviewing decision chains, explaining doc from first principles. Not for: PM/CTO summary (use project-brief), pre-doc analysis (use feasibility-study), code explanation (use codex-explain). Output: structured reasoning chain with sensitivity analysis.

172estrelas
Atualizado há 5 meses

Ver no GitHub ↗ Licença: MIT

Como adicionar

/plugin marketplace add sd0xdev/sd0x-dev-flow

O comando exato pode variar conforme o repositório. Confira o README no GitHub.

Para o autor da skill

Cole no README do seu repo

Mostra que sua skill está catalogada na Skillteca, gera backlink e tráfego rastreável.

[![Skillteca](https://www.skillteca.com.br/api/badge/fp-brief/svg)](https://www.skillteca.com.br/skills/fp-brief)
#ai

Skills relacionadas

Ver todas de Documentos →

Alerta por categoria

Receba novas skills de Documentos toda segunda

1 email curto, só com as skills novas de Documentos. 4 minutos de leitura, sem spam, cancela com 1 clique.

Você confirma o email no primeiro envio. Sem spam. Cancela com 1 clique.

Ler descrição completa

First-Principles Briefing Skill

Trigger

  • Keywords: first principles, fp brief, why was this decided, reasoning chain, decision sensitivity, explain decisions, assumption analysis, onboarding brief

When NOT to Use

| Scenario | Alternative | |----------|------------| | PM/CTO executive summary (strip technical details) | /project-brief | | Pre-doc feasibility analysis (before writing spec) | /feasibility-study | | Code explanation at function/file level | /codex-explain | | Code architecture overview | /code-explore | | Simple document summary | Ask Claude directly |

Command Signature

/fp-brief <doc-path> [--depth brief|normal|deep] [--verify off|codex] [--output <path>] [--no-save]

| Flag | Default | Description | |------|---------|-------------| | <doc-path> | Required | Source markdown document path | | --depth | normal | Output detail level | | --verify | off | Independent Codex reasoning verification | | --output | Same dir, -fp-brief.md suffix | Custom output path | | --no-save | false | Print to stdout instead of file |

Workflow

sequenceDiagram
    participant U as User
    participant S as /fp-brief
    participant D as Source Doc
    participant O as Output File
    participant X as Codex (optional)

    U->>S: /fp-brief <doc-path> [--depth] [--verify]
    Note over S: Phase 1: Input Resolution
    S->>S: Validate path (repo boundary)
    S->>D: Read source document
    S->>S: Redaction scan (fail-safe)
    S->>S: Auto-detect format (hybrid)
    Note over S: Phase 2: First-Principles Extraction
    S->>S: Extract Root Problem (5-Why)
    S->>S: Build Assumptions Register
    S->>S: Build Reasoning Chain
    S->>S: Build Alternative Rejection Log
    S->>S: Build Decision Sensitivity
    S->>S: Identify Open Unknowns
    Note over S: Phase 3: Output Assembly
    S->>O: Write *-fp-brief.md
    alt --verify codex
        S->>X: Independent reasoning verification
        X-->>S: Verification Delta
        S->>O: Append Verification Delta
    end
    S-->>U: Report complete

Phase 1: Input Resolution

  1. Path validation: Normalize, reject .. traversal, enforce repo boundary
  2. Read source document
  3. Redaction scan: High-confidence secret patterns → abort; medium → mask [REDACTED]
  4. Format auto-detection: See references/detection-rules.md
  5. Select extraction template based on detected format

Phase 2: First-Principles Extraction

See references/extraction-guide.md for section-by-section heuristics.

| Section | Core Question | |---------|--------------| | Root Problem | What fundamental truth makes this problem unavoidable? | | Assumptions Register | What are we taking for granted, and why? | | Reasoning Chain | How does each decision trace back to a principle? | | Alternative Rejection Log | Why do other approaches violate our principles? | | Decision Sensitivity | If assumption X breaks, which decisions collapse? | | Open Unknowns | What don't we know, and what should we find out? |

For long documents (>500 lines): split by ## headings, extract per-section, merge + dedup.

Phase 3: Output Assembly

  1. Apply depth filter (section inclusion matrix)
  2. Apply source citations (reference source doc section headings)
  3. Apply Evidence Insufficient Rule — never fabricate content for thin sections
  4. Write output file (or stdout if --no-save)
  5. If --verify codex: dispatch verification per references/codex-verify-prompt.md

Depth Levels

| Level | Description | Sections Included | |-------|-------------|-------------------| | brief | Core reasoning only (~500 words max) | Root Problem (full), Assumptions (top 3), Reasoning Chain (key decisions), Sensitivity (top 3) | | normal | Full reasoning chain (~1500 words max) | All 6 sections with citations | | deep | Full chain + analysis (~2500 words max) | All 6 sections + challenge questions, evidence ratings, counterfactual analysis, risk-weighted unknowns |

Verification Delta (section 7) appears only when --verify codex is used, at any depth level.

Length policy: These are upper bounds, not targets. If source doc is thin, output will be shorter. The Evidence Insufficient Rule applies: [Evidence insufficient — source doc lacks data for this section].

Output

See references/output-template.md for full template.

# First-Principles Briefing: <title>

> Source: <path> | Depth: <level> | Format: <type> | Generated: <timestamp>

## 1. Root Problem
## 2. Assumptions Register
## 3. Reasoning Chain
## 4. Alternative Rejection Log
## 5. Decision Sensitivity
## 6. Open Unknowns
## 7. Verification Delta (optional)

Save Behavior

| Condition | Output Path | |-----------|------------| | Default | Same directory as source, -fp-brief.md suffix | | --output <path> | Specified path | | --no-save | stdout only, no file written |

Example: docs/features/auth/2-tech-spec.md → docs/features/auth/2-tech-spec-fp-brief.md

Verification

  • [ ] Input path validated (repo boundary enforced)
  • [ ] Secret redaction scan executed
  • [ ] Format auto-detection result shown in output header
  • [ ] Each Reasoning Chain decision cites source section (Source: §<ref>)
  • [ ] Each Assumptions Register entry has confidence level
  • [ ] Decision Sensitivity maps assumptions to affected decisions
  • [ ] Evidence Insufficient markers used where source data is thin
  • [ ] Output length within depth-level upper bound
  • [ ] If --verify codex: Codex researched independently (per codex-invocation rules)

References

  • Output template: references/output-template.md
  • Detection rules: references/detection-rules.md
  • Extraction guide: references/extraction-guide.md
  • Codex verification: references/codex-verify-prompt.md

Examples

Input: /fp-brief docs/features/seek-verdict/2-tech-spec.md
Action: Read spec → detect tech-spec → extract 6 sections → write 2-tech-spec-fp-brief.md

Input: /fp-brief docs/features/auth/2-tech-spec.md --depth brief
Action: Read spec → extract Root Problem + top assumptions + key decisions + top sensitivity → brief output

Input: /fp-brief docs/features/auth/2-tech-spec.md --depth deep --verify codex
Action: Read spec → full extraction → Codex independent verification → write with Verification Delta

Input: /fp-brief notes/design-decisions.md --no-save
Action: Read doc → detect unknown format → generic extraction → print to stdout
Compartilhar X LinkedIn