The piece published by Archwise sets out a diagnosis we share word for word: "Most organisations think they have a prompt problem. In reality, they have a context problem." Prompt engineering, the consultancy writes, was useful and necessary - but it optimises interactions, not systems. Its benefits are "local and fragile". And as AI becomes embedded in critical systems, that fragility has a price.
This echoes what we recently drew from Wharton's research: prompt tricks do not scale. Archwise goes further and names the remedy - context engineering - and above all the principle underneath it: treat context as infrastructure. That is exactly the conviction Survol is built on. Let us walk through the article's stages.
"Prompt debt": an invisible liability
Archwise describes something many teams live without naming it: prompt debt. By collecting prompts "like strategic assets" - with no context, no versioning, no traceability - the library becomes unmanageable. "Prompt experts" turn into bottlenecks; when one leaves, the team loses capability and the old mistakes come back. The knowledge is tacit, tribal, and transferring it is "slow and fragile".
"Like any technical debt, prompt debt generates friction, errors and rework, and limits the organisation's ability to scale." - Archwise
Agentic development makes that debt worse. Because a coding agent needs, in every session, a context: the team's conventions, the decisions already made, the files concerned. If that context lives in a lead's head or in a CLAUDE.md nobody maintains, every session starts from approximate knowledge - and produces results that are inconsistent from one run to the next.
Context as infrastructure: the heart of the article, the heart of Survol
The conceptual leap Archwise calls context engineering consists of no longer treating context as a by-product, and making it "critical infrastructure, on the same level as code or data". Concretely, the consultancy lists four requirements. Let us take them one by one - because Survol turns each of them into a feature.
Context as infrastructure, per Archwise: capture it systematically, version and audit it like code, embed it in onboarding and governance processes, and make it accessible to humans and AI systems alike. Point by point, that is Survol's specification.
- Capture context instead of leaving it tacit. In Survol, a product's context is not scattered: each feature's spec, the settled product decisions (with their why), and the agent's guidance files (
CLAUDE.md-style, reusable procedures) are gathered and managed from a single place. Knowledge stops being tribal: it is written down and located. - Version and audit. Every version of a feature, every dated decision, every applied constraint is traced. When an organisation constraint is applied - or challenged by the agent - it is logged. Context has a history, just like code.
- Embed it in governance - proactive, not bureaucratic. This is where Archwise is sharpest: "bureaucratic" prompt governance (committees, approvals) fails, because it controls form without addressing substance. Survol does the opposite: our organisation technical constraints - security, libraries, conventions - are injected automatically at the start of every session, at three levels (mandatory, strong, preference). The rule is not a document you hope someone reads; it is infrastructure that applies itself, and that a human can evolve.
- Make it accessible to humans and agents. The same context serves as the source of truth for the team (who read the product map) and for the agent (which receives the approved spec and the decisions as its instruction). A principle we sum up like this: the approved spec IS the prompt.
Onboarding, transfer, maintainability: the gains Archwise measures
The article documents three success stories, all driven by the same shift. They describe, in outline, the value produced by context treated as infrastructure - and they are exactly the benefits Survol aims for:
- Faster onboarding - "from weeks to days" thanks to a central living document. In Survol, a new joiner (or a new agent) inherits the product map, the decisions and the constraints without having to interrupt an expert.
- Systematic knowledge transfer - instead of depending on collective memory. Capitalisation is native: nothing that gets decided is lost.
- Proactive governance and versioned context - to "evolve AI sustainably, with resilient systems". That is our pillar: supervise without owning, trace without slowing down.
There is even a gain Archwise does not cover that well-kept context makes possible: cost control. A structured context is a context you can trim - passing the agent only the spec, the decisions and the files concerned, rather than a soup of prompts. That is a third fewer tokens for the same feature. The discipline of context is not only about quality: it is also good housekeeping.
The takeaway
- The real problem is almost never the prompt: it is the context, invisible and tacit, living in people's heads and in scattered documents.
- Prompt debt is paid in inconsistency, expert dependency and rework - and agentic development amplifies it.
- The remedy is treating context as infrastructure: captured, versioned, audited, governed, accessible to humans and agents alike.
- That is precisely what Survol does with your agents' context - spec, decisions, constraints injected and traced - for fast onboarding, proactive governance and controlled cost.
Archwise closes with a question we happily pass on to our readers: does your organisation invest more in polishing its prompts than in improving the context available to its teams and its agents? If the answer leans towards prompts, a debt is quietly accumulating. The good news is that context can be built - and that a cockpit like Survol exists to turn it into infrastructure, from the very first session.
Source: "Pourquoi le Prompt Engineering ne suffit pas pour construire des systèmes avec l'IA", Archwise. The concepts of prompt debt and context engineering are developed there by the consultancy.
