Context Construction
Context construction is the skill of deciding what the agent should know, what it should not assume, and what should survive the handoff.
More context is not automatically better. Useful context is selected, structured, provenance-backed, and small enough to govern.
Learn
A context pack should answer:
- What is the task identity?
- What outcome is this work meant to move?
- What files, systems, or users are in scope?
- What constraints are hard, soft, or assumed?
- What prior attempts or landmines matter?
- What evidence already exists?
- What should trigger stop, dissent, or salvage?
This is where The Context Stack enters the curriculum. The agent should not infer strategy from a messy transcript. The human should construct a context object the agent and reviewer can inspect.
Practice
Use templates/context-pack.md.
Build the pack before /aim or /execute:
- Name the intent.
- List relevant subsystems and files.
- Separate hard constraints from soft constraints.
- Mark assumptions that should be tested.
- Add provenance for claims.
- Add stop, dissent, and salvage triggers.
Keep the pack short enough that a reviewer can challenge it.
Artifact
Use templates/context-pack.md. A context pack must preserve:
| Field | Why it matters |
|---|---|
| Task identity | Keeps later skills from solving the wrong problem. |
| Selected sources with provenance | Lets reviewers challenge authority and freshness. |
| Hard, soft, and assumed constraints | Separates rules from preferences and assumptions. |
| Evidence available now | Grounds the run in observable reality. |
| Landmines and prior attempts | Prevents repeated dead ends. |
| Stop, dissent, and salvage triggers | Names when the agent must pause instead of improvising. |
The pack is consumed by /aim, /problem-space, /problem-statement, /solution-space, /execute, /review, and /dissent.
Review check
Reject a context pack if:
- it hides provenance;
- it dumps files without explaining why they matter;
- constraints are mixed with preferences;
- it contains unverified claims as facts;
- it does not name stop conditions;
- the agent would still need to infer the actual task.
Go deeper
- The Context Stack — context as governed memory, provenance, task identity, and promotion path.
templates/context-pack.md— starting point for the artifact.docs/prompt-and-context.md— how to turn the context pack into a checkable prompt.docs/context-to-agent-tutorial.md— guided path from context pack to prompt, skill, and subagent.docs/open-horizons.md— where the context pack sits in the full loop.
Navigation
- Previous: Model-Fit Framing
- Up: Docs Home / Curriculum
- Next: Prompt and Context Assembly