An open development pattern for software that uses AI
Build with AI. Keep people able to take over. Keep The Lights On Development (KLOD) is a pattern for software that uses AI. Build in the data, tools and knowledge people need, so your team can finish essential work when AI stops or becomes unsafe.
Start building that way with two free, open-source skills for your coding assistant.
BUSINESS AS USUALEverything works. For now.
NOTICE FROM YOUR AI SUPPLIER
LLM access withdrawn.This service is no longer offered in your region.KLOD
KEEP THE LIGHTS ON
DEVELOPMENT
When the model is gone, who does the work?
Build the manual path in. Practise it. Measure it.
LLM OFFLINE · BUSINESS RUNNING
The story plays when it scrolls into view. Press L to replay it.
A regional access restriction. A prepared human. A business that carries on.
The story: your AI supplier withdraws service in your region and disconnects the AI model (the LLM). A prepared person switches to the manual path built in advance. The lights come back on and the work continues.
Free and open source
Give your coding assistant instructions to help you build and review software people can operate when AI stops.
Use KLOD in your existing workflow. Both integrations are maintained by KLOD.
Add KLOD planning and review commands to your Spec Kit workflow.
Set up the extensionInclude human operation, implementation tasks and verification in your changes.
Set up the schemaWith APM, the Agent Package Manager, installed, run this in your project folder:
apm install AlexandruGirlea/keepthelightson#v0.1.0 Then ask your assistant: Use klod-check to review this project.
klodBuild the featureklod-checkCheck your projectYour team reviews the changes and practises the handover on real work. Measure what people can complete, fix the gaps and repeat.
Read, use and contribute on GitHub01 / This has already happened
Access, infrastructure and prices are decided by your AI supplier. Whether your people can keep working should be decided by you. Three cases, each with its source.
12 JUN 2026 / EXPORT CONTROLS
A US directive restricted Fable 5 and Mythos 5 access for foreign nationals. Anthropic suspended both for all users. Fable access returned on 1 July 2026, after the controls were lifted.
Your business cannot set that timetable.
12 JUN 2025 / CLOUD OUTAGE
A Google Cloud failure spread across regions and products, including AI prediction services. A shared infrastructure failure can interrupt the tools your business needs, even when your own code is healthy.
A second AI supplier can share the same point of failure.
19 FEB 2026 / MODEL RETIREMENT
Claude Haiku 3.5 retired from the Claude API. Its recommended replacement, Haiku 4.5, lists base input/output rates of $1/$5 per million tokens, versus $0.80/$4: 25% higher unit prices.
Plan for replacement costs before the old model retires.
Sources checked 10 September 2026.
02 / The six principles
Use these principles when you add AI to a business process. Keep essential work possible, and give people a way to stop AI and take control when something goes wrong.
List the work that depends on your AI supplier, such as answering customers or processing invoices. Name who is responsible and decide what must keep running.
Save the requests, results and business rules people need in a form they can read without your supplier. Show what is finished and what still needs doing.
Provide the screens, tools and permissions people need to finish the job without AI. Name and train the people responsible, and write instructions someone else can follow.
Check what could stop AI and human work together: a shared login, cloud service or access restriction. The backup needs to work through the failure you are preparing for.
Regularly have people complete real work with AI unavailable. Measure time to the first correct result, how much work they can handle, and for how long. Fix what fails and repeat.
Give authorised people an independent way to stop AI from acting when its supplier fails or its behaviour becomes unsafe. Keep the controls and information people need to take over. For devices, machines and infrastructure, define a safe state when human operation is not possible, and preserve safety interlocks. Practise the switch and the handover under an agreed exercise plan.
Example: customer support
Your support AI goes offline. Customer questions stay in a shared inbox. Your team can read each conversation, open the help articles and send replies without the AI supplier. A recent drill tells you how quickly they can start and how many customers they can help.
A human team may handle less work than AI. Decide what gets priority and tell customers what to expect. Follow the customer-support example, then use section 4 of the specification to check the required evidence.
For IoT, medical robotics, nuclear infrastructure, and government or military systems, plan how AI is stopped, how safety controls stay active, and when trained people can take over. Read the guide: Keep people in control
Where regulation matters
EU rules require continuity measures in finance and other covered sectors. Rules for high-risk AI also address human oversight and safe stopping. See which laws already apply, then read what EU law already requires.
03 / The development pattern
For each capability (one piece of work your business must deliver, such as answering customer email) that depends on an AI supplier, record the dependency, build the manual path and test it on real work. Then declare a KLOD level for that capability, from L0 (no manual path) to L3 (people do the work routinely without AI).
Put every capability that depends on an AI supplier in a register: the supplier, the work affected, the level you declare, and the person accountable. Anything not in the register counts as L0.
Specification 4.1 Inventory: the registerGive people a way to open each task, use its inputs and business rules, and save the result without the AI supplier. Document the steps and limitations, then test how much work people can sustain.
Specification 4.2 The manual pathFor L2, run a drill (real work with the AI genuinely switched off) at least every 90 days. Make one drill in four unannounced. Measure Time to Manual (how long until the first correct result without AI) and manual capacity (how much of the normal volume people can sustain, and for how long). Record failures and repeat the drill after fixing them.
Specification 4.3 Measurement and 4.4 DrillsA KLOD level describes one capability, never a whole product. L2 and L3 declarations need current evidence and have expiry rules. Read what the four levels mean
04 / Put the pattern to work
Follow an example, check one capability, then build and test the manual path and the off switch.
Three worked examples: customer support, connected equipment and an optional AI feature in a document editor.
Learn the pattern 02 / CHECKAnswer questions about one capability in your browser. Get a provisional level, the records to check and the next actions. Nothing is sent to us.
Run the self-check 03 / BUILDKeep work accessible, add human controls and prevent late AI results from undoing people's decisions.
Build a usable manual pathOne KLOD level per capability. L1 is not a safe place to stay. L2 and L3 expire.
05 / More incidents and research
These incidents and research findings cover outages, restricted access, changing suppliers and the loss of human skills. Each entry links to its source. If you work in a bank, an insurer or an essential service, see which laws already apply.
Three overlapping infrastructure bugs degraded Claude output quality between 5 August and 18 September 2025. At the worst hour, 16% of Sonnet 4 requests were affected. There was no outage. Customers could not tell a provider regression from their own prompt drift for six weeks.
Anthropic postmortem, 17 Sep 2025
Every OpenAI service was degraded or unavailable from 3:16 PM to 7:38 PM PST on 11 December 2024. The cause was not an attack and not capacity. It was a routine internal telemetry change that overwhelmed their own control plane.
OpenAI incident report, 11 Dec 2024
A database permissions change at Cloudflare on 18 November 2025 crashed the proxy that fronts a large share of the web, and took ChatGPT down with it. Using two AI suppliers does not help when both sit behind the same infrastructure.
Cloudflare postmortem, 18 Nov 2025
OpenAI announced the retirement of the Codex models on 20 March 2023 and shut them off on 23 March 2023. Longer notice periods were published afterwards.
OpenAI deprecations, entry 2023-03-20
On 3 June 2025 Windsurf lost nearly all direct access to Claude models with, by its own account, under a week's notice. The stated reasons were compute capacity and a reported acquisition by a competitor. The access simply ended, on a timescale Windsurf did not set.
TechCrunch, 3 Jun 2025
Asked at the French Senate on 10 June 2025 whether he could guarantee that French citizens' data would never be passed to US authorities, Microsoft France's Director of Public and Legal Affairs answered: Non, je ne peux pas le garantir.
Sénat, commission d'enquête sur la commande publique, 10 Jun 2025 · Sénat hearing · Proceedings
Three US-based companies account for 65% of the EU cloud services market. The UK competition authority closed its cloud market investigation in July 2025 and recommended strategic market status investigations into two of them. Under DORA, the largest ICT providers to European finance already sit under direct European oversight.
People change roles and skills fade without practice. Keep the manual path in use, train replacements and verify their performance in drills.
06 / Already required in some sectors
Two EU laws already require something close, each within its own sector. Neither names AI models. KLOD is one way to document and test that continuity, one capability at a time.
| Rule | Requirement | Who is covered |
|---|---|---|
| DORA Art 28(8) | Documented, tested exit plans for ICT services supporting critical or important functions, with alternatives and transition arrangements | Financial entities within DORA's scope |
| Regulation (EU) 2022/2554 Art 29(1)(a) | Assessment of concentration risk where a provider is "not easily substitutable" | Financial entities |
| Regulation (EU) 2022/2554 Art 11, Art 12 | ICT continuity policy, business impact analysis, yearly testing, backup and restoration, RTO and RPO | Financial entities |
| Directive (EU) 2022/2555 (NIS2) Art 21(2)(c) | Business continuity, backup management, disaster recovery, crisis management | Essential and important entities |
| Directive (EU) 2022/2555 Art 21(2)(d) | Supply chain security | Essential and important entities |
NIS2 covers qualifying entities in sectors including energy, transport, healthcare, digital infrastructure and public administration. Check its scope and the applicable national law; sector membership alone does not determine coverage. These laws address continuity and supplier risk. They do not prescribe KLOD or require every AI task to have a manual equivalent.
Where KLOD fits: the register (section 4.1) records each AI supplier and the work that depends on it. The manual path (section 4.2) is an alternative people can operate. The measurements and drills (sections 4.3 and 4.4) show, with a date, that the plan was tested on real work. Whether your organisation has a legal duty to keep such a plan needs a separate assessment.
EU systemic-risk warning, 25 June 2026
The emergent frontier AI models increase inherent ICT risk, leading to weakened operational resilience … particularly across four areas … concentration, including dependencies on a limited number of AI providers, dependent in turn on cloud providers, open source components and other widely used software.
European Systemic Risk Board, Warning of 25 June 2026 on systemic cyber risks stemming from frontier artificial intelligence models (ESRB/2026/3), OJ C/2026/3795, 16 July 2026. ESRB Warning · Official Journal
The same Warning records that the geographical concentration of leading AI providers outside the Union leaves it exposed to strategic dependency and geopolitical risk.
Informative only, reviewed on 11 September 2026. Applying KLOD does not by itself establish legal compliance. Read the full page: what EU law already requires, and what the AI Act does and does not require
Where to start
Answering customer requests, approving invoices or completing reports: start with a job your team knows. KLOD calls it a capability. Then, in this order: