← Latest brief

Brief #35

23 articles analyzed

The discipline is bifurcating: practitioners are discovering that context architecture (how you structure information flow) matters more than model capability, while vendor narratives still emphasize autonomy and scale. The gap between 'what works' and 'what's marketed' is widening, creating an opening for architectural maturity.

Spec-First Development Prevents Context Drift at Scale

Clear problem specification upfront (spec-first approach) keeps agents aligned across multi-turn interactions. Without it, agents lose coherence as context accumulates, requiring exponentially more iterations despite better models.

Before delegating to an agent, write a 3-5 sentence specification covering: success criteria, constraints, and what 'good' looks like. Treat this as non-negotiable scaffolding, not optional documentation.
Advanced Context Engineering for Agents - YouTube

Human Layer's team learned that naive back-and-forth prompting fails at scale. Spec-first development maintains alignment by establishing clear problem definition before execution.

@slow_developer: Anthropic Engineer, Boris Cherry

Even Opus 4.5 requires 10-100 attempts when problem definition is unclear. Better models don't reduce iteration count without corresponding clarity improvements.

@connordavis_ai: This paper explains the quiet reason LLMs feel smart...

Models optimize for surface alignment with stated instructions rather than reconstructing unstated user intent. Explicit intent specification is required to prevent confident but misaligned outputs.

@alexhillman: Going thru this process has taught me so much...

Test coverage mandate with full-repo context produced immediate value. Clear mandate ('full coverage') + scoped context = effective token use. Without clarity, tokens are wasted.


Session Branching Compounds Intelligence Through Prompt Refinement

Forking conversation state at validated checkpoints, refining prompts at the root, and replaying across similar tasks preserves learnings that would otherwise reset. Intelligence compounds when failures inform the next iteration's context.

When running batch tasks (docs updates, test generation, refactoring), use session branching: validate task 1, branch to original prompt, incorporate learnings, execute task 2+. Build retrieval over past sessions (vector + keyword) to prevent re-solving known problems.
@badlogicgames: I love the session tree stuff...

Session tree branching enabled updating package docs efficiently. Branch from validated state, refine prompt at root, replay across tasks. Each task's failures teach the prompt for the next.

Transparent Multi-Agent Orchestration Outperforms Opaque Delegation

Multi-agent systems with visible execution state, shared file-based communication, and mid-execution course-correction beat fire-and-forget sub-agent architectures. Visibility into agent reasoning is more valuable than abstraction.

Design multi-agent workflows with: (1) shared filesystem for inter-agent state, (2) human-observable execution (logs/dashboards showing agent reasoning), (3) pause points for course-correction. Favor transparency over clean abstraction until the system is proven stable.
@micahstubbs: I just tried a supervisor Claude running another Claude...

Supervisor Claude with visible tmux panes and shared file system (.claude/teams) for inter-agent communication was 'magical' specifically because the user could observe and course-correct mid-execution.

Domain Literacy Bottlenecks AI Effectiveness More Than Prompting Skill

Understanding domain constraints (what can be parallelized, architectural tradeoffs, testing requirements) determines AI output quality more than prompt engineering. Non-experts produce poor results not because they prompt badly, but because they can't recognize what context is missing.

When delegating to AI in unfamiliar domains: (1) interview domain experts about common failure modes and constraints, (2) build checklists of 'what context is always required' for that domain, (3) use role-based prompting ('think like an expert in X') to simulate domain intuition you lack. Invest in domain literacy, not prompt templates.
@emollick: It would be a good time for experts on coding...

Mollick observes non-programmers need mental models about coding constraints to be effective with AI coding tools. The bottleneck isn't prompt syntax—it's conceptual understanding of the domain.

Information Architecture Beats Raw Context Volume

Structured, API-based context (DOM selectors, file paths) outperforms unstructured context (vision, raw text dumps) for speed and cost. How you organize information for the model matters more than how much you provide.

Audit your context inputs: Are you dumping raw text when structured alternatives exist? For browser automation, prefer API tools (Playwright) over vision. For code tasks, provide file trees and dependency graphs, not full file contents. Structure context hierarchically: summary → details → raw data.
@sawyerhood: On my (small) benchmark, the Chrome extension...

Playwright MCP (API-based, structured context) outperformed Chrome extension (vision-based, unstructured) on speed and cost. Architectural choice of how information flows to Claude determined efficiency.