菜单
Treat AI context as code with a full development lifecycle — Generate, Test, Distribute, Observe, Adapt. From the founder of DevOps applied to AI engineering.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/Farrice/mes-knowledge-base --skill patrick-debois-cdlc命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
The /generate front door — in-house, pay-as-you-go creative generation (Higgsfield replacement). One command for image, video, and audio: routes each request to the right model recipe (skills/generate/models/*.json), defers to binding creative_router lanes (people → fal-people, style-family → fantastic-posters; Higgsfield fully retired 2026-08-06), quotes paid video before running, honors prompt-level budgets ("total budget $3"), runs comparison batches across models, logs every prompt beside its asset, and auto-refreshes the Asset Command Center (/assets-board). Use when the user says "generate", "make an image/video/voiceover", "create ads/assets/creative", names a model (recraft, kling, seedance, nano banana, gpt image), or asks for a multi-model comparison. Source: RoboNuggets 9C4TRbucmhQ, rebuilt on our fal-first stack.
The Edit Bay — in-house conversational video studio (VOX-style explainers, documentaries, shorts, LinkedIn video). Transcript-driven cutting, B-roll ladder, motion graphics as code, captions, music, publish-ready packages — all agent-operated, free-first. Use when Farrice says "edit this video/footage", "make an explainer/documentary/short", "cut this recording", "add B-roll/captions/music", "turn this essay/post into a video", or any video production ask. Front door for three modes: talking-head, VO-only, zero-camera (his recorded VO + assembled visuals — NO voice clone, NO TTS narration, his ruling 2026-08-06). Source: Brad Bonanno agentic edit pipeline (extractions/brad-bonanno-edit-bay/), rebuilt on the house stack: WhisperX local + edit_bay.py ffmpeg assembly + HyperFrames/Remotion graphics + /generate (fal) — NO Higgsfield (retired 2026-08-06).
Architects technical explainer videos using Brad Bonanno's structural patterns — survives short attention spans, communicates complete systems (not just teasers), and functions as a channel-building artifact (one video pulls subscribers, not just views). Use when planning a long-form technical explainer (5-20 min) for YouTube, structuring an explanation of a complex system the user knows but can't teach yet, balancing depth vs accessibility for a technical audience, or building an evergreen video meant to compound subscriber growth over years rather than spike then die. Trigger proactively whenever the user says "explainer video", "technical explainer", "how do I explain X", "long-form YouTube", "evergreen video", or has a complex system that needs to land with a non-expert. For Instagram-format shareworthy content use brock-johnson-shareworthy-content; for short-form video patterns use a TikTok/Reels-specific skill.
基于 SOC 职业分类
正在显示 SKILL.md
| name | patrick-debois-cdlc |
| description | Treat AI context as code with a full development lifecycle — Generate, Test, Distribute, Observe, Adapt. From the founder of DevOps applied to AI engineering. |
| version | 2.0 |
| format | completion-engine |
| workflows | 4 |
| expert | Patrick Debois |
| domain | AI-augmented engineering / context engineering |
| source | AI Engineering Summit keynote, 2026 — 'Context is the new code' |
| source_url | https://www.youtube.com/watch?v=bSG9wUYaHWU |
| extracted | 2026-05-03 |
| extracted_with | claude-opus-4-7[1m] |
The founder of DevOps applied his 2009 question ("what if ops looked more like dev?") to AI engineering in 2025: "what if context is the code?" Result: the CDLC — a 5-stage lifecycle (Generate → Test → Distribute → Observe → Adapt) that treats AI context as code with full systems-engineering discipline.
Use this skill when you need to diagnose, build, or evolve any AI context system — skills, agents, directives, prompts, multi-agent orchestrations.
Patrick Debois is the only practitioner with the credentials to extend the DevOps mental model into AI-augmented engineering at the lifecycle level. His insight: AI engineering teams in 2025 are 3-5 years behind where DevOps was in 2014. He's reapplying the exact maturity arc DevOps traveled (2009-2015) to context engineering (2025-2028) — and the entire upcoming tooling wave is predictable from this analog: context linters, context eval frameworks, context registries with semver, context security scanners, context observability platforms.
Direct relevance to Antigravity: The system operates at registry-tier complexity (210 skills, 117 agents, 58 directives) but at copy-paste-tier hygiene. The 2026-04-24 system audit identified this gap empirically. CDLC provides the lifecycle vocabulary to organize what's already being built (eval_harness.py, routing_enforcer.py, skill_auditor.py, evolution_orchestrator.py) and what needs to be built next.
| # | Slash Command | What It Produces | When to Deploy |
|---|---|---|---|
| 1 | /cdlc-audit | Diagnostic scoring of any AI context system across 5 CDLC stages with bottleneck identification + per-stage upgrade prescriptions | When a system has 5+ context artifacts and the user wants a systems diagnosis (not a single-artifact review) |
| 2 | /context-evals | Complete eval suite for one target artifact: lint + Grammarly + LLM-as-judge unit tests + e2e tests, all with N-run success rates and calibrated error budgets | When a load-bearing artifact has no eval coverage or only lint-tier |
| 3 | /context-library | Migration plan from current distribution tier to next: version pins, dependency declarations, security scans, SBOMs, predicted failure modes | When a collection has 10+ artifacts being shared across projects/teammates and the user is hitting version-drift or dependency friction |
| 4 | /context-observe | Observability loop closure: three feedback channels (agent logs / PR comments / prod failures) wired to context-improvement candidates with explicit trust budget and review cadence | When the system has ≥2 weeks of operational data, recurring failures are appearing, and existing context is mature enough to diagnose vs rebuild |
The 4 workflows form a sequence:
/cdlc-audit identifies which stage is the bottleneck/context-evals upgrades the Test stage (most common bottleneck)/context-library upgrades the Distribute stage/context-observe upgrades the Observe + Adapt stagesRun /cdlc-audit first. The other 3 are prescriptions tied to specific stage scores.
| Patrick's Workflow | Stacks With | Why |
|---|---|---|
/cdlc-audit | /system-audit, skill_auditor.py audit, eval_harness.py status | Patrick's audit gives the lifecycle frame; existing tools give empirical evidence to score against |
/context-evals | eval_harness.py (runner), chain_runner.finalize (production-failure feed) | This workflow produces the suite; the harness executes it; finalize feeds prod failures into pending tests |
/context-library | skill_auditor.py, evolution_orchestrator.py | Auditor's A/B/C/REVIEW classification feeds quality gating; orchestrator can monitor version drift |
/context-observe | recall_logger.py, routing_enforcer.py, evolution_orchestrator.py | Existing loggers ARE Patrick's Channel 1; orchestrator's daily/weekly/monthly cadence IS Patrick's review loop |
| Patrick's Workflow | Pairs With | Compound Output |
|---|---|---|
/cdlc-audit | /system-hygiene | Lifecycle diagnosis + tactical hygiene fixes |
/context-evals | Lara Acosta / Luke Iha / Sean Macintyre voice skills | Author voice-specific unfakeability tests (Pattern 20 detection, AI-tells-banned tests, etc.) |
/context-library | evolution-direction.md | Library curation aligned with current evolution priorities |
/context-observe | /skill-evolution, /evolution-status | Observe loop feeds skill-evolution review queue |
| # | Move | Trigger |
|---|---|---|
| 1 | Force the Parallel First — name the established discipline this resembles before proposing anything new | Introducing infrastructure for an emerging domain |
| 2 | Demand the Test Tier — "what tier is your eval — lint, Grammarly, unit, or e2e?" | Auditing any context artifact |
| 3 | Run-It-Five-Times Reflex — never single-run an LLM eval; always N-run with error budget | Any LLM-as-judge evaluation |
| 4 | Convert PR Argument to Context Upgrade — instead of arguing the comment, ask what context would have prevented it | PR review friction with AI-generated code |
| 5 | Default to "Crap Until Proven Otherwise" — baseline classification is REVIEW, eval evidence required to elevate | Auditing third-party or legacy context library |
Every output from this skill should pass Patrick's unfakeability test:
"If this artifact were not loaded, would the output still pass the test?" If yes → the test is theatre. Re-author so passing PROVES context loaded.
┌─────────────────────────────────┐
│ │
▼ │ │ ▲
GENERATE → TEST → DISTRIBUTE → OBSERVE → ADAPT
▲ │ ▼
│ │
└─────────────────────────────────┘
Loops are paced by their weakest stage. Improving any stage other than the bottleneck wastes effort. (Goldratt's Theory of Constraints, applied to context engineering.)
All extraction from: Patrick Debois, "Context is the new code / Context Development Life Cycle" — AI Engineering Summit architect track keynote, ~25 minutes, 4,276 words. Source URL: https://www.youtube.com/watch?v=bSG9wUYaHWU
Patrick is currently founder/CTO at Tessl, building tooling for the CDLC he describes. His next conference: AI DevCon London, June 1-2, 2026 (he curates the content).
4 deterministic practitioner prompts — each carries an Output Contract, Output Skeleton, and Quality Gate. When a deliverable matches one, Read it and honor its contract instead of improvising the output shape.
skills/patrick-debois-cdlc/references/prompts-v2/cdlc-audit.mdskills/patrick-debois-cdlc/references/prompts-v2/context-eval-suite.mdskills/patrick-debois-cdlc/references/prompts-v2/context-library-migration.mdskills/patrick-debois-cdlc/references/prompts-v2/observability-loop-closure.md