FDE Consultants Protocoles Skill

Forward Deployed Engineering (FDE) methodology for AI coding agents. Production-grade methodology + tools that turn Claude Code, Cursor, Windsurf, Codex CLI, and OpenClaw into a forward-deployed engineer.

What is FDE Consultants Protocoles?

FDE Consultants Protocoles is an industrial-grade Skill that teaches any AI coding agent to apply the Forward Deployed Engineering methodology — the discipline pioneered at Palantir (~2010) and now common across enterprise AI engineering teams.

In 2026, FDE postings grew 729% year-over-year (source). This Skill encodes that workflow into:

• A reusable methodology with 14 Operating Principles, 10 Anti-Patterns, and a 6-Q decomposition
• A 4-stage loop (Scoping → Prototyping → Production → Feedback)
• 7 MCP tools exposed via stdio transport
• A multi-agent runtime (Modex) with 8 agents: 4 DeepSCR roles + 4 FDE domain specialists
• The Frozen Arbiters: opt-in full-autonomy governance — a sealed, tamper-evident Contract plus frozen mutation-tested oracles (with a measured baseline, never a declared one) that override the Certifier's optimism at ship time (--governed)
• A FDE Assurance Score Registry with SHA-256 chain (verifiable by anyone with shasum)

Who is it for?

• AI Labs that want a verifiable FDE Assurance Score badge on every agent output
• Enterprise Engineering Teams that need a standard intake form for agent work
• Solo Engineers & Consultants who need an agent that scopes before coding
• Researchers exploring hypothesis-driven agent workflows (DeepSCR protocol)

Quick Start (60 seconds)

Option 1 — Zero-Install (any web LLM)

Copy-paste skill/ZERO-INSTALL.md into ChatGPT, Claude.ai, or Gemini. Works immediately, no install.

Option 2 — One-Command Install (CLI agents)

git clone https://github.com/selectess/fde-consultants-protocoles.git
cd fde-consultants-protocoles
bash install.sh

Auto-installs into Claude Code, Codex CLI, Hermes, and OpenClaw. Cursor and Windsurf use the included one-file rule (see skills/).

Option 3 — Manual install

See skill/INSTALL.md for step-by-step setup on any runtime.

What's Included

1. Methodology (skill/SKILL.md, Apache-2.0)

The core deliverable. Encodes:

• 14 Operating Principles (e.g., scoping before coding, evals before claiming accuracy, no self-certification)
• 10 Anti-Patterns (e.g., "just use AI/ML", trust me bro, magic number default)
• 6-Q decomposition (process → decision → data → cost → current state → success metric)
• 4-stage loop (Scoping → Prototyping → Production → Feedback)
• FDE Assurance Score section required on every deliverable

2. MCP Server (skill/mcp_server/, Apache-2.0)

7 tools exposed via stdio transport, zero external dependencies (Python stdlib only):

• fde_recon — Stage 0 Reconnaissance: scan the real codebase before scoping
• fde_decompose — validate 6-Q decomposition, reject vague inputs
• fde_roi — compute ROI, fail if below threshold
• fde_scientific_search — held-out promotion gate (DeepSCR)
• fde_evals — generate eval rubric + golden set
• fde_ontology — build domain ontology from case studies
• fde_trust_score — compute FDE Assurance Score

3. Modex Multi-Agent Runtime (modex/, MIT)

A 1-agent local runtime (free, MIT) + Plugin with ed25519 license ($6 lifetime, BSL). 4 DeepSCR roles with separation of powers:

• Lead — decomposes the problem into the 6-Q spec
• Researcher — gathers scientific + market evidence
• Builder — produces the deliverable, consulting 4 FDE domain specialists in parallel (scoping, architecture, agent engineering, production readiness)
• Certifier — independently re-derives the FDE Assurance Score

Optionally governed by the Frozen Arbiters (modex/arbiters.py): a sealed Contract maps every stage to a clause (uncovered actions are mechanically rejected), frozen mutation-tested oracles re-measure the shipped candidate against the best measured alternative at ship time, verdicts must cite their evidence run (uncited = null), and certified trajectories distill into persisted lessons. python3 -m modex.engage --project <path> --governed

4. FDE Assurance Score Registry (registry/, Apache-2.0)

Public, append-only, SHA-256-chained. Every certified deliverable gets an entry:

• genesis.json — chain origin
• 3 case studies (SaaS churn 93/100, retail forecasting 91/100, fintech fraud 93/100)
• Each entry links back to the previous via prev_hash (git-backed, no SaaS)

5. Independent Certifier (modex/certify_skill.py)

Re-derives the Skill's FDE Assurance Score from raw evidence. Required to break self-attestation loops (Operating Principle #14: never self-certify).

Key terms (glossary)

This project is the primary source for the terms it coins. Canonical definitions: glossary.

• DeepSCR (Deep Sceptical Contextual Research) — a governance protocol (separation of powers + Hypothesis → Contradiction → Verification → Certification) created by Mehdi Wehbi. The proposer never certifies; evidence must resolve on disk.
• Frozen Arbiters — full-autonomy governance: a sealed tamper-evident Contract, mutation-tested Oracles, and a measured baseline, all frozen before launch. The agent pleads its case; it never grades itself. (modex/arbiters.py)
• FDE Assurance Score — a 0–100 certification score (claim 25 / contradiction 25 / evidence 30 / anti-patterns 20) with an independent veto; certified only above 85.
• Measured Baseline Oracle — a frozen oracle that re-runs the candidate search at ship time; the baseline is measured, never declared. (modex/engage.py)
• Modex Collective — an 8-agent runtime: 4 DeepSCR powers governing 4 FDE specialists, with a self-prompting loop. (modex/collective.py)

Read more: white paper · governed autonomy · 6 governance approaches compared

FDE Assurance Score (this Skill)

The Skill self-attests 94/100 — that is the honest headline number. The bundled Certifier (modex/certify_skill.py) re-derives 100/100 from raw evidence; since it ships in this repo, treat that as a reproducibility check, not third-party certification. For an independent score, run the certifier yourself or commission an external audit.

Verify yourself:

python3 -m pytest skill/tests/ modex/tests/
# Expected: 154 passed in ~5s

python3 -m modex.certify_skill --skill-path ./skill --output ./cert.json
# Expected: FDE Assurance Score 100/100, verdict certified

Architecture (3 layers)

Distribution (Multi-Platform)

The Skill works on every major AI coding agent:

Repository Map

Pricing (the public catalog)

See modex/PRICING.md and modex/LICENSE-SYSTEM.md for the ed25519 license mechanism.

Status (this Skill is V1.0.0)

Contributing

This project follows its own Operating Principles and Anti-Patterns. Read CONTRIBUTING.md for the four-checklist PR process:

1. Falsifiable claim (what changes)
2. 3 failure modes (what could go wrong)
3. Evidence trail (file:line + test commands)
4. Anti-pattern check (no buzzword inflation, no self-certification)

Security

See SECURITY.md. Report vulnerabilities to [email protected].

Before self-hosting the paid Modex Plugin in production, complete its configuration:

• Set your own ed25519 license key in modex/license.py.
• Configure your Stripe credentials in modex/stripe_webhook.py.

License (multi-tier)

• skill/ — Apache-2.0 (open-source methodology + tools)
• modex/ (core runtime) — MIT (free for any use)
• modex/plugin/ (Plugin) — BSL 1.1 ($6 lifetime, see LICENSE-SYSTEM.md)
• portal/landing/assets/ — CC-BY-4.0

See LICENSE, THIRD_PARTY_NOTICES.md.

Compatibility with the AI Lab Skills Ecosystem

FDE Consultants Protocoles is complementary to official Skills ecosystems:

Use FDE Consultants Protocoles when your problem is "ship production code with proof it works". Use the official Skills when it's "do a single document task well".

Links

• What are skills? — Anthropic support
• Model Context Protocol — MCP specification
• DeepSCR FDE Assurance Score — Methodology inspiration
• Palantir FDE — Origin of the FDE role
• FDE job growth — Market data

Last updated: 2026-07-02 · Built with industrial rigor · Verified by 154/154 tests · FDE Assurance Score 94/100 (self-assessed) · Apache-2.0 (Skill) + MIT (Modex core) + BSL (Plugin)