# Context Not Constraints Is What Unhobbles an Agent ## Claim A capable model's performance is limited more by missing context than by insufficient instruction, so removing constraints and supplying situation raises capability where adding rules does not. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/Field Guide to Fable — Thariq Shihipar, Anthropic]] — AI Engineer talk, launch day The concrete result is that roughly **80% of the Claude Code system prompt was removed**, and the product got better. That is only explicable if most of what was removed was not adding capability. It was constraining a model that already knew how to do the thing, and the constraints were interfering. The framing behind it is *capability overhang* — the model is grown rather than designed, so its abilities exceed what any harness has been built to elicit. Under that view a system prompt is not primarily how you make a model competent; it is a mask, and the engineering problem is which parts of the mask are load-bearing. The prescription inverts the usual instinct. When output is wrong, the reflex is to add a rule. Shihipar's version is to ask what the model does not know about the situation — the codebase conventions, the definition of done, what already failed — because a rule addresses a symptom while context addresses the cause. This lands directly against the failure-diagnosis order that already governs this vault's agent work: context problem, then effort problem, then model problem. The talk supplies a mechanism for why context comes first, and evidence that the reflex to over-specify is not a small inefficiency but the dominant one. The second half — mapping your unknowns through blind-spot passes, interviews, references — is the same claim pointed at the human. The person's job is to specify intent and adjudicate what neither party knows, which is a context problem too. ## Scope & conditions - **Holds** for frontier models on tasks within their competence, where failure looks like the wrong thing done well rather than an inability to do it. - **Breaks** for genuine capability gaps. No amount of context makes a model do what it cannot do, and reading every failure as a context problem wastes real effort. - **Breaks** for safety and policy constraints, which exist precisely to bind behaviour the model is otherwise capable of. "Remove constraints" is a capability argument, not a governance one. - **Weakens** where the constraint encodes something genuinely unguessable — a house convention, a regulatory requirement, an arbitrary interface. That is not a mask, it is context in imperative form. - **Caveat on the source**: this is a launch-day talk by the vendor, and "tradeoffs aren't real" is stated with more confidence than the 80% figure alone supports. The measurement is the durable part. ## Links - [[3. Resources/Permanent Notes/Vague Prompts Are Expensive Specificity Cuts AI Error and Cost Together]] — because contrasts, and the tension is the useful part: that note says be specific, this says stop adding rules. They reconcile as *be specific about the situation, not about the procedure*. - [[3. Resources/Permanent Notes/The Harness Now Outpaces the Model as the Source of Agent Capability Gains]] — because extends: capability overhang is why the harness has the leverage. There is unclaimed ability sitting there, and the harness decides how much is reachable. - [[3. Resources/Permanent Notes/Effort Is a Cost-Quality Dial That Subsumes Prompt Complexity]] — because extends: both say elaborate prompt machinery is the wrong lever, and name a different one. - [[3. Resources/Permanent Notes/Bounded Action Spaces Not Better Models Make Agents Safe to Run]] — because contrasts: the necessary counterweight. Removing constraints raises capability and raises blast radius, and these two notes should be read against each other rather than separately. - [[3. Resources/Permanent Notes/Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]] — because extends: written intent is exactly the context that makes constraints unnecessary, and it is what a `SKILL.md` should hold. - [[3. Resources/Permanent Notes/AI's Leverage Comes From the Data Foundation Not the Model]] — because extends: the same claim at organizational scale — what the system can see determines what it can do. - [[3. Resources/Permanent Notes/Subagent Configuration Inherits From the Parent With No Override]] — because applies-in: inheritance is a context-propagation mechanism, so what the parent knows sets the subagent's ceiling. - [[3. Resources/Permanent Notes/Encoding What Good Looks Like Is What Makes Agent Verification Trustworthy]] — because extends: the definition of done is the highest-value piece of context, and it is the one most often left implicit. - [[AI Agent Safety MOC]] — because applies-in. - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the MOC lists this note under Synthesis additions (2026-08-01, run 3–4). ## Implications - The practical edit to any skill or instruction file is subtractive first: delete a rule and check whether the behaviour degrades. Most rules were written against an older model and never re-tested. - It sharpens the existing diagnosis order into a test rather than a heuristic. Before escalating model or effort, ask what the agent could not have known — and if there is an answer, the escalation was going to be wasted. - `SKILL.md` files should read as briefings rather than procedures — conventions, constraints that genuinely exist in the world, what has already been tried — because that is context, and procedure is mask. - The safety caveat has to travel with it. Unhobbling and bounding are independent axes; this argument is about the first and says nothing about the second, and applying it to safety constraints misreads it. ## Open Questions - Is the 80% figure specific to a harness that had accreted constraints over time, or does it generalize to prompts written fresh against current models? - How do you distinguish, before testing, a constraint that masks capability from one that encodes unguessable context? - Does the argument hold for smaller models, where more of the system prompt may be doing real scaffolding work rather than masking? ## Source - [[3. Resources/Readwise/Articles/Field Guide to Fable — Thariq Shihipar, Anthropic]] — capability overhang, the 80% system-prompt reduction, and the map-your-unknowns method