Credits and External Inspirations

This guide documents patterns from the Claude Code community. Some sections are directly inspired by open-source repos, blog posts, or public engineering work. This page consolidates attributions in one place.

Packmind Engineering Team

Repo: github.com/packmind/packmind Author: Cédric Teyton (CTO, Packmind) License: Apache 2.0

Packmind maintains a production Claude Code configuration in their open-source repo. Several patterns documented in this guide were inspired by their .claude/ setup:

Pattern 1: MCP Reference File

Guide section: Documenting an MCP for Claude Source: .claude/skills/datadog-analysis/references/datadog_mcp.md

A references/<mcp-name>.md file read by the skill before any MCP call. Captures query syntax gotchas, required parameter combinations, working examples, and noise exclusion rules specific to one MCP server.

Pattern 2: Skeptical Reviewer Sub-Agent

Guide section: Pattern: Skeptical Reviewer Sub-Agent Source: .claude/skills/playbook-audit/references/report-agent.md

A fourth agent in a multi-agent audit pipeline whose job is to reject false positives from the first three agents. Operates with explicit false-positive criteria and requires evidence from both artifacts before keeping any finding.

Pattern 3: Shared Ground Truth Injection

Guide section: Skill Design Patterns Source: .claude/skills/doc-audit/SKILL.md

The orchestrator computes a shared factual baseline (nav structure, CLI commands, file list) once, then injects the same block into all parallel sub-agent prompts. Prevents each sub-agent from re-discovering the same facts independently.

Pattern 4: Pre-filtered References via Frontmatter Paths

Guide section: Skill Design Patterns Source: .claude/skills/qa-review/SKILL.md

The orchestrator reads rules files, parses paths: frontmatter glob patterns, matches them against modified files, and passes only applicable rules to each review agent. Progressive disclosure applied to standards, not just documentation.

Pattern 5: Handoff Triad with Merge Semantics

Guide section: Session Handoff Pattern and templates at examples/commands/handoff/ Source: .claude/commands/create-handoff.md, resume-handoff.md, update-handoff.md

A three-command protocol for session continuity: create-handoff initializes a structured document, update-handoff applies section-specific merge rules (append-only for work log, replace for status), and resume-handoff loads the latest document into context.

Pattern 6: Recipe Template with Context Validation Checkpoints

Guide section: Skill Design Patterns and template at examples/commands/recipe-template.md Source: recurring pattern across Packmind command files

A command template structure where a “Context Validation Checkpoints” section lists preconditions Claude must verify before executing the recipe steps. Reduces errors caused by missing context or wrong environment.

Packmind context-evaluator

Repo: github.com/PackmindHub/context-evaluator Author: Packmind engineering team License: MIT

context-evaluator is an OSS CLAUDE.md / AGENTS.md quality analyzer. Two patterns from its source were extracted for the guide:

Pattern 7: Runtime Prompt Logging

Guide section: Skill Design Patterns Source: src/shared/evaluation/runtime-prompt-logger.ts

Always-on blocking write of the full evaluator prompt to prompts/debug/ before invoking the AI provider. Survives provider crashes and timeouts. Never throws. Separate from the --debug flag.

Pattern 8: Adaptive Unified/Parallel Mode

Guide section: Skill Design Patterns Source: src/shared/evaluation/runner.ts — canUseUnifiedMode()

Token-threshold switching between single-agent unified evaluation (cross-file detection) and parallel independent agents per file. Threshold default: 100K tokens.

Anthropic Engineering Team

skill-creator: The skill-creator skill vendored in the Packmind repo (and referenced in this guide’s skill evaluation section) was originally published by Anthropic. It contains the canonical evaluation harness for testing skills: evals filesystem convention, blind A/B comparator, description optimization loop, and benchmark aggregation scripts.

IFTTD (If This Then Dev) Podcast

Podcast: ifttd.io Host: Bruno Soulez License: Editorial citation (no code)

IFTTD is a French tech podcast (360+ episodes, 2020-2026) covering practical software engineering. Transcripts from episodes 290 to 361 were analyzed and their practitioner insights paraphrased into the guide. No direct quotes appear in the guide; all material is reformulated in English and attributed by episode.

Sections drawing on IFTTD

Guide section: Practitioner Insights Source episodes: 311, 326, 329, 338, 341, 346, 349, 351, 357, 360, 361 Consolidated digest of paraphrased insights organized by theme (context engineering, agentic patterns, LLM evaluation, agent security, DevX and adoption).

Guide section: §6 RAG Optimization: Query-Side Indexing Source: ep 361, Guillaume Laforge (Developer Advocate, Google Cloud) Practitioner perspective on semantic chunking and synthetic-question generation at index time, supplementing the existing Anthropic Contextual Retrieval documentation.

Guide section: §17 Attention Mechanics: Lost-in-the-Middle Source: ep 361, Guillaume Laforge Practitioner note on large context window anti-patterns, added to the existing Liu et al. research citation.

Guide section: Evaluating Probabilistic Systems Source: ep 338 (Louis Pinsard), ep 329 (Frédéric Barthelet), ep 311 (Samy Lastmann) Consolidated section on scored-dataset evaluation, statistical CI/CD, LLM-as-judge (async), hallucination calibration, and OpenTelemetry/Langfuse observability.

Guide section: Why Sandboxing Matters: Field Incidents Source: ep 326 (Zineb Bendhiba), ep 360 (Guillaume Lours) Documented incidents on guardrail evasion and unsupervised agent data loss.

Guide section: MCP Vetting Workflow Source: ep 346, Jocelyn N’takpe (Head of Engineering & Architecture, ManoMano) Command allowlist recommendation, illustrated with a real incident.

Guide section: Usage Principles Source: ep 326 (Zineb Bendhiba), ep 329 (Frédéric Barthelet) Tool count, tool design principles, and LLM statefulness notes.

Guide section: Practitioner Testimonials Source: ep 311, 341, 346, 361 Four new testimonials on micro-agent architectures, harness engineering, multi-model orchestration, and production MCP stacks.

Guide section: Adoption Approaches: What We Do Know Source: ep 346, 349, 351 Practitioner notes on documentation-as-onboarding, typed language safety nets, and sustainable development pace.

Devoxx

Conference: Devoxx (Java/JVM, architecture, DevOps, AI tracks) License: Editorial citation (no code)

Devoxx talks from 2025 and 2026 were analyzed and paraphrased into the guide’s practitioner insights. No direct quotes appear; all material is reformulated in English and attributed by speaker and year.

Sections drawing on Devoxx

Guide section: Practitioner Insights Speakers: Brian Vermeer, Max Sumrall, Tom Cools, Christoph Bühler, Daniel Garnier-Moiroux, Annie Freeman, Mete Atamel, Jettro Coenradie, Daniël Spee, and others (see the Devoxx table on that page) Consolidated digest of paraphrased insights on prompt injection testing, the spec paradox, LLM-as-judge design, and MCP access control.

Guide section: security-hardening.md Source: Brian Vermeer, 2026 Prompt injection testing and LLM session secrecy.

Guide section: sandbox-isolation.md Source: Christoph Bühler, Daniel Garnier-Moiroux, 2025-2026 MCP access control patterns (OAuth, OpenFGA) and MCP server hardening.

Guide section: Evaluating Probabilistic Systems Source: Brian Vermeer, Mete Atamel, Jettro Coenradie, Daniël Spee, Tom Cools, 2025-2026 Determinism, evaluation framework combination, confidence-based agent evaluation.

Guide section: observability.md Source: Annie Freeman, 2026 Instrumenting coding assistants with OpenTelemetry.

Guide section: agent-teams.md Source: Max Sumrall, 2026 The spec paradox and the Picnic productivity REX.

Guide section: context-engineering.md Source: Alex Gavrilescu, Konstantin Pavlov, 2025 Markdown as shared persistent memory paired with a conventions file.

Dev With AI Meetup

Meetup: Dev With AI (French AI-native development meetup) License: Editorial citation (no code)

Dev With AI Meetup talks from 2026 were analyzed and paraphrased into the guide’s practitioner insights. No direct quotes appear; all material is reformulated in English and attributed by speaker and year.

Sections drawing on Dev With AI Meetup

Guide section: Practitioner Insights Speakers: Emmanuel Sciara, Florian Allainmat, Luis Iglesias Hernandez, Alexandre Balmes, Vyncke, Samuel Gallet, Geslain Dahan, Geoffrey Graveaud, and others (see the Dev With AI Meetup table on that page) Consolidated digest of paraphrased insights on context degradation thresholds, exit criteria for agents, and skill self-improvement.

Guide section: context-engineering.md Source: Emmanuel Sciara, Florian Allainmat, 2026 The 70% context degradation threshold, CLAUDE.md growth hurting performance, and the research-plan-implement handoff pattern.

Guide section: Evaluating Probabilistic Systems Source: Alexandre Balmes, Emmanuel Sciara, Negouai, Drode, 2026 Temperature-zero determinism myths and skill self-improvement as reinforcement learning.

Guide section: sandbox-native.md Source: Bolin, Vyncke, Florian Allainmat, 2026 Production agent isolation pattern (ephemeral container, read-only repo, network allowlist).

Guide section: Adoption Approaches: What We Do Know Source: Geoffrey Graveaud, Samuel Gallet, Geslain Dahan, 2026 Practice amplification (good and bad) and a documented velocity-loss REX from unsupervised agent delegation.

ByteByteGo

Channel: ByteByteGo (system design) License: Editorial citation (no code)

ByteByteGo videos were analyzed and paraphrased into the guide, filtered strictly for relevance to agentic AI and Claude Code workflows. No direct quotes appear.

Sections drawing on ByteByteGo

Guide section: Practitioner Insights Source: ByteByteGo, “7 System Design Concepts,” 2025 Start-simple-measure-then-complexify, applied to agentic setups.

Guide section: Usage Principles Source: ByteByteGo, “MCP” (2025), “Back-of-the-Envelope Estimation” (2022), “URL Shortener” (2025) MCP as a standardized integration layer, and QPS-based capacity estimation for agent-driven traffic.

Stanford Online

Channel: Stanford Online (academic ML/LLM coursework) License: Editorial citation (no code)

Stanford Online lectures were analyzed and paraphrased into the guide, filtered strictly for direct, actionable relevance to Claude Code practice. No direct quotes appear.

Sections drawing on Stanford Online

Guide section: Practitioner Insights Source: Denny Zhou (Google DeepMind), Stanford CS25 V5, 2025; Yann Dubois, CS224N, 2024; Mehran Sahami, “It’s Never Too Late,” 2025; ISLR (Hastie & Tibshirani); CS230; CS25 2026 Reasoning as intermediate tokens with a formal Boolean-circuit bound, evaluation design (Cohen’s kappa, out-of-distribution testing), the shift toward verifying AI-generated code, and the necessity check before adding a model.

Guide section: context-engineering.md Source: Denny Zhou, Stanford CS25 V5, 2025; CS25 2026; CS230 Lecture 8, 2025 Chain-of-thought as a formal compute extension, and in-context generation versus fine-tuning.

Guide section: Evaluating Probabilistic Systems Source: Yann Dubois, CS224N, 2024; ISLR; CS230 Evaluation metric selection by pipeline stage and the necessity check before adding a model.

Guide section: learning-with-ai.md Source: Mehran Sahami, “It’s Never Too Late,” 2025 Verifying AI-generated code as the emerging critical skill.

The Product Crew

Podcast: The Product Crew (French product management podcast) License: Editorial citation (no code)

The Product Crew episodes were analyzed and paraphrased into the guide’s adoption, roles, and ops pages, filtered strictly for relevance to AI adoption decisions. No direct quotes appear. The cited Ask For The Moon field report is a single-instance 2026 account, not established industry data, and is presented as such.

Sections drawing on The Product Crew

Guide section: Adoption Approaches: What We Do Know Source: The Product Crew, 2024-2026 (contexts: Alan, Payfit, Joko) AI ownership by domain experts, adoption through practice rather than top-down rollout, and leadership buy-in as a prerequisite.

Guide section: ai-roles.md Source: The Product Crew, 2026 AI ownership by domain experts over a centralized innovation cell.

Guide section: ai-unit-economics.md Source: The Product Crew, AI transformation field report, 2026 (context: Ask For The Moon) The “tokens into ROI” framing used to introduce the stakes of unit economics.

Adding to This File

When a guide section is directly inspired by or adapted from external open-source work, add an entry here. Include:

Repository URL
Author / organization name
License
Which pattern was borrowed and from which source file
Which guide section it appears in