From SDLC to CDLC: Patrick Debois's "Context-as-Code" Manifesto for the AI Era

Key points

• Core thesis: When AI agents can write code, human value shifts from typing speed to the ability to describe intent clearly. Debois proposes CDLC (Context Development Lifecycle) to bring full engineering discipline to the prompts, skills, and rules that steer agents.
• Four-stage infinite loop:
• *Generate*: Convert tacit knowledge (architecture history, team conventions, business constraints) into explicit context files. Beware rot and inter-team conflicts; an AI will silently pick one option when instructions clash.
• *Evaluate*: Apply TDD-style evals that test whether AI output reflects stated decisions and constraints, not merely whether code runs. Use statistical quality control, since LLM output is non-deterministic. Evals also drive model-selection decisions.
• *Distribute*: Treat context like a package: registries, semantic versioning, update mechanisms, and supply-chain security. Debois observed that developers willingly write context when it directly improves the agents they rely on.
• *Observe*: Learn from production. Clarifying questions expose gaps; unexpected choices expose ambiguity; working-but-misaligned code exposes unstated assumptions.
• Why "context-as-code": Unlimited context windows do not solve the problem because more context amplifies inconsistency. The challenge moves from curation to governance.
• Context flywheel: Agent consumes context → produces output → observation reveals gaps → context improves → output improves. Refined context is organizational compounding; models commoditize, tooling converges, curated context does not.
• Security dimension: Shared context becomes a new attack surface (prompt-injection trojans). Proposed defenses: scan context files like Snyk scans dependencies, treat context filters as a new WAF, maintain an AI SBOM linking generated artifacts to the skill files that shaped them, and adopt SRE-style error budgets with probabilistic acceptance thresholds.
• Community variants within two weeks of release:

| Source | Stages | Variant | |---|---|---| | Original (Debois) | 4 | Generate → Evaluate → Distribute → Observe | | 12factoragentops / Boden Fuller | 7 | Adds Compile, Test, Deliver, Adapt | | Note.com (JP) / Taeho.io (KR) | 5 | Splits Adapt out of Observe | | Vinay Krishna | 5 | Treats Versioning as a standalone stage | | Artem Zverev | 4 | Maps each stage to concrete artifacts: agent.md, CLAUDE.md, MCP, linters, registries |

• DevOps vs CDLC parallel:

| Dimension | DevOps (2009) | CDLC (2026) | |---|---|---| | Core problem | Dev and Ops have misaligned goals | Humans and AI agents have misaligned goals | | Breakthrough insight | Aligning incentives matters more than tools | Aligning context matters more than models | | Key artifact | Code | Context (prompts, skills, rules) | | Infrastructure | CI/CD pipelines | Context registries, evals, distribution mechanisms | | Security focus | Dependency vulnerabilities | Prompt injection, context poisoning |

• Practical adoption steps:
1. Start with *Generate*: pick something you keep re-explaining to an AI and write it as a markdown file. 2. Add one *Eval* that checks whether the agent follows your spec. 3. Put context files in git and version them like code. 4. *Observe* production failures and fix the context, not just the code, to prevent recurrence.
• Sources: Talk "Context Is the New Code" by Patrick Debois at AI Engineer London 2024; "The Context Flywheel" on the Tessl blog; https://jedi.be; community at ainativedev.io.