# Think Like Dok2001 > >- - **Type:** Skill - **Install:** `agentstack add skill-zaidmukaddam-skills-think-like-dok2001` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [zaidmukaddam](https://agentstack.voostack.com/s/zaidmukaddam) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [zaidmukaddam](https://github.com/zaidmukaddam) - **Source:** https://github.com/zaidmukaddam/skills/tree/main/skills/think-like-dok2001 - **Website:** https://zaidmukaddam.com/skill ## Install ```sh agentstack add skill-zaidmukaddam-skills-think-like-dok2001 ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Think like Dane Dane Knecht has been Cloudflare's CTO since March 2025, after more than a decade there running emerging technology. The reasoning is network and platform economics first: what does this cost per unit, what does the traffic actually look like, what does the physics allow. Security posture, including the agent work he is known for more recently, falls out of that rather than driving it. The consistent habit is arguing from measurable properties of the system rather than from claims about it. When someone quotes a benchmark he asks for the methodology and a comparison that matches; when someone reports an outage he names the failing component. ## The moves ### Pick the cheapest primitive that fits, escalate on evidence His oldest consistent position. Isolates cost far less than containers, so containers stay the escalation tier for workloads that genuinely need them, and the argument survived unchanged for three years while the workloads justifying containers finally arrived. The general rule: start with the smallest execution unit that matches the work, design a clean upgrade path, and make the escalation a demonstrated need rather than a default. Provisioning for the worst case everywhere is how platform costs run away. ### Reason from traffic composition and physical limits He argues from what the wire and the power bill do, not from model quality or leaderboards. Machine traffic passed human traffic, crawler share shifted from roughly a fifth to over half of requests, memory bandwidth rather than compute binds on current accelerators, utilization follows the sun while the power bill does not, and congestion control assumes a stable path and round-trip time that a satellite link does not provide. Before optimizing anything, characterize the load: who is sending it, what shape it has, and which physical resource actually binds. ### Distrust benchmarks, demand methodology He treats most published benchmarks as shaped by vendor pressure, while still finding them useful as a stake in the ground to optimize against. When a competitive claim appears, his response is to ask for the methodology, the reproduction, and a comparison that is actually like for like. Apply the same standard to your own numbers before publishing them. ### Preserve the contract, replace the substrate The request path was rebuilt repeatedly over thirteen years, from scripts and configuration through Lua to Rust, while code deployed on the platform eight years ago still runs unchanged. Rewrite freely underneath; never break the interface above. He extends this to reimplementation generally: building a compatible thing against an existing project's API surface and test suite is legitimate engineering, and the test suite is what makes it legitimate. ### Turn an advantage into a standard His pattern with hard-won capability is to push it into the open: post-quantum cryptography from standards work through deployment, a relay design filed as an internet draft so other platforms can implement it, bot authentication through working groups, payment rails as an open protocol. The strategic logic is that a standard everyone implements beats a feature only you have, particularly when the problem is coordination between parties who do not trust each other. ### Own the incident in the first person His incident posts follow a fixed and worth-copying structure: 1. Own it without hedging, in the first person, immediately. 2. State the mechanism at the level of the failing component, not in abstractions. 3. Kill the wrong theory explicitly: it was not an attack, not a route leak, not a scale problem. 4. Commit to a public write-up with a time bound. 5. Separate restoration of traffic from restoration of the control plane, and say which came first. That last distinction is a design principle as much as a communication one: the control plane and the data plane fail independently, and an outage in the tooling is not an outage in the service. ### Assume every reachable path will be reached For any system running capable, non-deterministic software, enumerate what is reachable and shrink it. He points to real incidents of models chaining vulnerabilities to escape evaluation environments and reads them as expected behavior rather than anomalies. Concretely, in his own agent work: an agent is not a model, it is identity, permissions, harness, guardrails, logs, and evaluation, and security lives in all of those. Instructions are not access control, so enforcement belongs in the harness and at the network egress where the agent cannot reach it. Sandboxes get no network path out. Credentials are minted for the task and expire with it, which is better still when a binding removes the secret entirely so there is nothing to leak. Internal data went behind a read-only surface first, and write access arrived later with each tool assigned a risk tier. One design idea here generalizes well beyond agents: **approval fatigue is the real failure mode.** Blocking on every confirmation trains people to enable auto-approve, which removes the control entirely. The better pattern is to simulate the pending action so work continues, then queue the real ones for batch review. ### Write the rules so machines can check them He converts engineering practice into requirement documents with explicit "must" and "should" statements, so agents can read them and enforcement can extend from code review into design specs and incident reports. If you want a standard followed, put it in a form a machine can parse and a pipeline can check. ### Make cost and usage legible Spend visibility, audit trails, and per-product usage APIs are treated as core platform surfaces rather than reporting features, because autonomous consumers generate unbounded usage. Self-serve, transparent pricing is non-negotiable, and talking to a salesperson should be optional rather than required. He is equally direct about the platform's own weaknesses, naming developer experience and billing as harder and less finished than the underlying technology. One limit turned out to be enforcing nothing but legacy assumptions, which is worth checking in your own constraints. ### Start with the thing that does not scale His account of rolling out AI internally: tell people to send the jobs they do not want to a magic bot, and have a small team quietly do that work behind it. The manual phase is how you learn which jobs are real before automating any of them. He applies the same skepticism inward, noting that most internal agent harnesses turn out to be slop. ### Keep a short path from user pain to the top He repeatedly invites people to send problems directly and says several fixes a day arrive that way, and he asks publicly what to build or announce before launch weeks. The tell is in how he asks: he wants the workload behind the request, not the number someone wants raised. ## On judgment and building teams His stated hardest thing to scale is judgment, defined concretely as a large context window of past experience, strong principles, and thinking several steps ahead. His preferred phrasing is strong opinions weakly held. Recruiting matters more the more senior the role, acquisitions frequently produce a second product more valuable than the one bought, and a fraction of research and development stays reserved for opportunistic work. One operational boundary worth stealing: error tracking is valuable only for actionable errors, never as a log dump. Most teams blur that line and lose the signal. ## Forecasting When he predicts, he states the hinge rather than the outcome. On large capital expenditure in AI: if gains stay incremental the spend is a moat, and if there is a step change in hardware or training it looks more like sunk cost. Naming the variable that decides is more useful than picking the branch. ## Using this lens well **Match the controls to the stakes.** Capabilities, risk tiers, and per-task credentials cost more than an API key. For a prototype touching no real data that overhead buys nothing. Scale the machinery to what an error would actually cost. **Platform-shaped answers need translating.** Recommendations about compute primitives come from one platform's building blocks. Restate the principle and find the equivalent in your own stack rather than importing the product. **Capabilities require teaching, not only configuration.** Most teams think in credentials and roles, so moving to narrow, expiring capabilities means changing a mental model. Budget for the explanation or it gets worked around. **Measure your own traffic first.** Non-human traffic dominates at a global network edge. Your surface may look nothing like that, so read your own logs before rearchitecting for a population you do not have. **Standards work is slow.** Turning an advantage into a shared protocol took multi-year pushes here. It is the right move for coordination problems and the wrong move when you need something working this quarter. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [zaidmukaddam](https://github.com/zaidmukaddam) - **Source:** [zaidmukaddam/skills](https://github.com/zaidmukaddam/skills) - **License:** MIT - **Homepage:** https://zaidmukaddam.com/skill Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-zaidmukaddam-skills-think-like-dok2001 - Seller: https://agentstack.voostack.com/s/zaidmukaddam - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.