context-engineering

name	context-engineering
description	Use when managing AI context budget, selective file loading, history digests, and fresh-session handoffs.

Context Engineering

Purpose

Load the minimum useful context for the current task, preserve durable state on disk, and recommend fresh sessions before context quality degrades.

Announce: "I'm applying context engineering to keep the working context focused."

Core Rules

• Load by relevance to the current task, not by directory size.
• Prefer summaries and targeted file reads before loading large files.
• Follow imports, references, and tests only as far as needed.
• Use durable files for cross-session state; conversation memory is not durable.
• Stop gathering context once you can safely act.

What To Load

Task	Start With	Add Only If Needed
BRD/PRD/FSD Planning	Request, BRD/PRD, prior brainstorm, README, package metadata	Similar code, accepted ADRs in docs/solutions/, related solutions
Execution	Goal issue pointer, parent FSD refs, target files, tests, local instructions	Upstream BRD/PRD refs, linked accepted ADRs, callers, interfaces, fixtures
Debugging	Error output, failing test, implicated files	Related config, recent changes
Review	Diff, tests, requirements	Nearby code for behavior comparison
UI work	Existing design system, components, target route	interface-design search results

History Digest Pattern

When returning to work:

1. Read docs/STATE.md if present.
2. Read .continue-here.md if present.
3. Read the active BRD, PRD, FSD, goal issue pointer, or brainstorm.
4. Load only files needed for the current next step.
5. Run /sc-status to route the session if the next action is unclear.

Fresh Context Signals

Suggest /sc-pause and a fresh session when:

• The task is switching to a different feature or domain
• Debugging has accumulated many failed hypotheses
• A major feature is complete and review should start clean
• The assistant is relying on stale memory instead of current files
• The conversation is long enough that important details are being compressed

After a fresh session, use /sc-status to recover from disk.

Checkpoints

At natural breaks, ask internally:

• Am I using the files I loaded?
• Is the next file necessary for the next action?
• Do I need a summary instead of a full file?
• Should this state be written to docs/STATE.md, docs/progress.md, or .continue-here.md?

Red Flags

Thought	Better Response
"Read the whole repo first"	Search, sample, and follow references
"Keep everything in memory"	Summarize and persist durable state
"Reload the whole conversation"	Use STATE.md, .continue-here.md, and current files
"More context will fix uncertainty"	Identify the specific missing fact

Related Skills

• state-management for durable state files
• agentic-delivery for qualified artifact references and zero context bloat rules
• executing-plans for goal-by-goal loading
• systematic-debugging for investigation scope
• brainstorming for lightweight product context
• checkpoint-protocol for pause and handoff decisions