Context Engineering¶

The difference between a good AI agent output and a mediocre one is almost always context quality. learnship doesn't let agents guess: every invocation is loaded with the right structured context for what it needs to do.

The problem with raw AI agents¶

Without structured context, every AI session starts cold:

• The agent doesn't know your tech stack
• It doesn't know what decisions you've already made
• It doesn't know which phase you're in or what was built before
• It makes assumptions: and those assumptions become bugs

The naive fix (pasting context manually at the start of every session) is fragile, inconsistent, and exhausting.

How learnship solves it¶

flowchart LR
    subgraph CONTEXT["Loaded into every agent call"]
        A["AGENTS.md<br/>Project soul + current phase"]
        B["REQUIREMENTS.md<br/>What we're building"]
        C["DECISIONS.md<br/>Every architectural choice"]
        D["Phase CONTEXT.md<br/>Implementation preferences"]
    end
    CONTEXT --> AGENT["AI Agent"]
    AGENT --> P["Executable PLAN.md"]
    AGENT --> S["Commits + SUMMARY.md"]

AGENTS.md: persistent project memory¶

Generated by /new-project and placed at your project root. Your AI platform reads it as a system rule at the start of every conversation: automatically, without any action from you.

AGENTS.md
├── Soul & Principles        # Pair-programmer framing, 10 working principles
├── Platform Context         # Points to .planning/, explains the phase loop
├── Current Phase            # Updated automatically by workflows
├── Project Structure        # Filled from your new-project answers
├── Tech Stack               # Filled from domain research
└── Regressions              # Updated by /debug when bugs are fixed

Every workflow updates AGENTS.md when the phase advances: so the agent always knows where the project stands.

DECISIONS.md: the decision register¶

## DEC-001: Use Zustand over Redux
Date: 2026-03-01 | Phase: 2 | Type: library
Context: Needed client-side state for dashboard filters
Options: Zustand (simple), Redux (complex, overkill for scope)
Choice: Zustand
Rationale: 3x less boilerplate, sufficient for current data flow
Consequences: Locks React as UI framework
Status: active

The planner reads DECISIONS.md before creating any plans and never contradicts active decisions. The debugger adds lessons from bugs. /audit-milestone checks that decisions were honored.

Phase CONTEXT.md: your preferences per phase¶

Before planning, /discuss-phase writes your implementation preferences to .planning/phases/N-*/N-CONTEXT.md. The planner reads this and creates plans that match your choices: not generic best practices.

Parallel execution with full context budgets¶

On Claude Code, OpenCode, and Codex CLI, parallel subagents are on by default: execute-phase dispatches each independent plan to its own dedicated executor agent with a full 200k token context budget. Plans in the same wave run in parallel.

To disable parallel execution: set "parallelization": { "enabled": false } in .planning/config.json.

What "nothing is guessed" means in practice¶