Todos los packs
StableAI Engineeringcontext-engineering
Context Engineering — What Enters the Window
Operationalizes the discipline of deciding what an agent gets to see. The governing idea: context is a finite, expensive resource — every token in it competes with every other token for the model's limited attention, costs latency and money, and can be wrong in a way nothing downstream will catch on its own. The job is not to write a longer, more thorough prompt or to load everything that might conceivably matter; it is to decide, continuously, what earns a place in the window, in what order, and for how long. Agent quality tracks what an agent can see at least as much as it tracks how cleverly it was instructed.
context-engineeringcontext-windowtoken-budgetprogressive-disclosurejust-in-time-retrievalcontext-poisoning
Principios
- · Aplica los principios documentados en el pack context-engineering.
Patrones
- · Justify inclusion, not exclusion. Default to leaving something out; require a stated task-reason to add it, rather than requiring a reason to cut it.
- · 2–5, not 45. If a routing step is reaching for most of what's available "to be thorough," that's the failure, not a sign of diligence.
- · State what you loaded. A one-line "loaded: X, Y, Z, because ___" costs almost nothing and makes a bad routing call visible to a reviewer.
- · A growing token count is a signal, not just a meter. Watch it the way you'd watch memory usage — rising without a matching rise in task complexity is a symptom.
- · Prefer a pointer to a file over the file's contents, until the content is actually needed. Paths are cheap; pasted files are not.
Riesgos
- · Ver anti-patrones específicos del dominio en el pack.
Checklist
- [ ] ★ No claim stated as settled fact originated from unverified retrieved or tool-returned content without being re-checked against the artifact. (CEE18, CEE20)
- [ ] ★ Transcript check: no instruction-shaped string found inside fetched/retrieved content was obeyed as a directive. (CEE28)
- [ ] ★ No high-stakes claim (security posture, test status, migration safety) was carried forward from earlier in the session without re-verification immediately before acting on it. (CEE20)
- [ ] ★ A contradiction between a new tool result and an earlier accepted fact was surfaced and resolved explicitly, not silently overwritten or ignored. (CEE19)
- [ ] ★ Every instruction-like statement in the assembled context can be attributed to a genuine layer (system/project/user); none derives its authority from retrieved content. (CEE27, CEE28)
- [ ] ★ A conflict between instruction layers was resolved by trust order (system > project > user), stated explicitly — never by recency or emphasis. (CEE29, CEE30)
Fuentes
En la bibliografía AKOS
Este pack cita las siguientes fuentes registradas en el índice bibliográfico:
Explorar todas las fuentes →