Context Development Lifecycle
Context Development Lifecycle
Patrick Debois's framing: if context is the new code, then the loop you wrap around it should look like a software development lifecycle. Four phases, drawn as an infinity loop in the talk:
Generate → Evaluate → Distribute → Observe → (back to Generate)
The whole idea is that every piece of context — prompts, agents.md/CLAUDE.md, skills, library docs, tickets, specs — should pass through this loop with the same rigor a code change passes through SDLC.
The four phases
1. Generate
Where most teams already are. Methods:
- Direct prompting
- Reusable instruction files (
agents.md,CLAUDE.md, etc — a soft standardization is emerging) - Pulling library/API docs (avoid model hallucination on stale versions)
- MCP-style pulls from GitLab/GitHub/Slack/Jira
- Spec-driven development (write the spec; the agent breaks it into a plan)
2. Evaluate
The least practiced phase. Patrick's testing layers, by analogy to code testing:
- Linting — schema validation (e.g. a skill must have a description; under N chars)
- Grammarly-style checks — ask an LLM "is this context understandable?" Get back "not verbose enough" / "missing pieces"
- LLM as Judge — given the prompt and a piece of generated code, did it follow the rule? (e.g. "every API endpoint must use
/awesomeprefix" → judge checks) - Sandboxed agent tests — bind the judge to tools, let it execute the generated code (curl an endpoint, run the test)
- Error budgets, not pass/fail. Evals are non-deterministic — run 5 times, count successes. "You cannot do exact testing all the time."
CI/CD-friendly with the error-budget caveat. This is the part Patrick says people most underestimate the work of: "you thought you'd save time by writing context instead of code; you'll spend that time writing the right evals."
3. Distribute
How a context piece moves from "in my repo" to "shareable across teams":
- Repo check-in (zero friction; teammate clones, has it)
- Packaging — bundle context as a library (skills are a package format that all the major coding agents now recognize)
- Registries — a marketplace to discover packages (Tessl, Anthropic skill registry, etc)
- Dependency hell — yes, context will get this too. React-context-package conflicts with front-end-package
- Security scanning — Snyk and others now scan context for credentials, third-party exposure
- AI SBOM — provenance: who built this skill, with what model, on which dependencies
4. Observe
Closing the loop:
- Read agent logs (Agent ND standard is emerging) to see when agents say "I'm missing context X" — surface that across the org
- PR feedback as context signal: a PR comment "this isn't right" is feedback on the context that produced the PR. Improve the context, not just the patch.
- Production failures: instrument code, push to prod, capture failures with input/output, propose a test case. Self-healing context loop.
- Sandbox + context filter — agents are very resourceful at finding env vars / secrets. Sandboxes don't filter context loaded into the agent (agents.md, skills auto-load); you need a "context filter" — Patrick's analog of a web application firewall — to strip prompt injections and unsafe patterns before they reach the agent.
The context flywheel
Patrick's coda: there are three nested loops:
- Solo loop — you crafting your own markdown
- Team loop — make it a reflex: missing context → add context. Library author / improvement loop
- Org-of-teams loop — fix it once, every other team benefits. The flywheel.
Better context → better agent output → better observations → better context.
Why this matters
- Most teams are doing all four phases ad-hoc. Patrick's contribution is to name them and demand rigor — same move he made with DevOps in 2009.
- Maps cleanly to existing engineering muscle: anyone who can run a CI/CD pipeline can stand up evals; anyone who maintains a library can package skills.
- Naming the lifecycle creates a vocabulary for the gaps. "We're strong on Generate, weak on Observe" is a sentence a team can act on.
How it relates to other framings in this wiki
- Context Engineering (Martin Keen, IBM) is the four-pillar architectural view (connected access, knowledge layer, precision retrieval, runtime governance). Patrick's CDLC is the process view alongside Keen's architecture view. Same field, complementary lenses.
- Harness (LLM Agents) is what you're managing the lifecycle of. Patrick name-checked "harness engineering" on stage as a parallel discipline — see Harness Engineering (Ryan Lopopolo, AI Engineer) for the operational counterpart.
- AWARE Framework is the enterprise-governance framework; CDLC is more developer-loop oriented. AWARE is what runs in production; CDLC is how you build what runs in production.
- What an Enterprise Context Layer Is (Prukalpa) lifts the CDLC out of code-context into business-context: metrics, glossaries, sales playbooks. Prukalpa's "Development Lifecycle" capability is essentially CDLC, applied to the whole enterprise artefact set rather than only to prompts/skills. See Data Democratisation in Sales — Governed Context Layer, Not Dashboard Access for the worked sales example.
Sources
- Context Is the New Code (Patrick Debois, AI Engineer)
- What an Enterprise Context Layer Is (Prukalpa) — extends CDLC to enterprise business context