Context engineering

mk:context-engineering — Runtime Context Router

Decide the next context move for the task in front of you: read more, stop, ask, assume, or delegate — then load at most one pattern reference to execute it. This skill routes decisions; it does not preload a knowledge library.

Boundary vs mk:context-audit: that skill audits structural .claude/ overhead (what is statically loaded every session). This skill decides runtime context moves for one task. Structural question → context-audit. Runtime question → here. If a request is ambiguous between the two (e.g. "this project's context is messy"), ask ONE question — structural .claude/ setup or this session's runtime context? — then route.

When to Use

• Context is near its working ceiling and phases remain — decide what to keep.
• A long session is drifting or output quality is slipping.
• You must choose the minimal read-set before a task (Plan/Build/Review).
• You are choosing between /compact, /clear, and spawning a sub-agent.
• A known pattern applies (lost-in-middle, RAG dump, few-shot drift) and you want the routing.

When NOT to Use

• Auditing .claude/ token overhead → mk:context-audit.
• Persisting a learning or reading memory topic files → mk:memory.
• Finding where code lives → mk:scout (or Read for one known file).
• USD run cost → mk:budget.
• A trivial 1–2 file task — no routing needed; just do it.

Four Operations

The whole surface reduces to four moves. Name the one you need, then triage.

• Write — push durable state out of the window (memory file, scratchpad, handoff brief) so it survives compaction. Route via the persistence patterns.
• Select — load the smallest high-signal set; extract the function, not the file. Default stance for every read.
• Compress — summarize essential state and restart when the window degrades; re-ground the summary against sources.
• Isolate — give a sub-agent its own focused window for an independent slice; return a distilled result, not raw context.

Decision Framework

Pointer-style — the rules below own the detail; do not restate them here.

• Read more — only when a Retrieval Trigger in context-budget-rules.md matches (schema/contract change, security surface, cross-cutting refactor, failed verification, resume). No trigger → no read. (Rule file is present in a toolkit install; if absent downstream, fall back to: read only what a named acceptance criterion requires.)
• Stop — you have enough to act, OR two consecutive reads changed no decision, OR the phase/tier read ceiling is reached. "Seems right" without evidence is not stop — verify.
• Ask — an ungreppable fact, a spec-vs-code conflict, or a business/irreversible decision (core-behaviors.md #2, if present). Scout first: if grep/read answers it, don't ask.
• Assume — cheap-to-fix uncertainty. Emit an ASSUMPTIONS I'M MAKING block (format in Output Format below; matches core-behaviors.md #1 where present) and proceed; never silently guess.
• Delegate — an independent sub-scope that would bloat this window → Isolate into a sub-agent.

Anti-overengineering (keep it small): pick the smallest read-set that satisfies the acceptance criteria; do not introduce a new pattern or abstraction without an observed failure; prefer the boring move.

Anti-hallucination (don't invent): never quote an API, number, or config from memory when it is greppable — read it. An unverifiable fact is an Ask or an explicit Assume, never a silent guess.

Triage (copy this checklist)

Context triage:
- [ ] 1. Phase + tier — which of Orient/Plan/Test/Build/Review/Ship/Reflect, and low/standard/complex?
- [ ] 2. Context risk — near ceiling? drifting? poisoned by a prior claim? key fact buried? (labeled heuristics, not measured)
- [ ] 3. Minimal read-set — the smallest set for THIS step (see context-budget-rules.md phase×tier, if present)
- [ ] 4. Route — load exactly ONE section of references/pattern-index.md matching the intent
- [ ] 5. Apply + emit the output contract (Context decision: …)

Step 2 thresholds are labeled heuristics (e.g. "compact around ~60–70% of the effective window") — signals to act on, not measured facts. Do not present them as measured.

Pattern Routing Table

Match the runtime intent, load that ONE section of references/pattern-index.md. Never load several.

Intent (runtime signal)	Load section
Too much loaded; cut before adding	less-noise
Long session, output drifting	degraded-session
State must survive compaction / new session	persistence
Decompose across sub-agents	delegation
RAG / index-first / grounding	retrieval
Prompt ordering, trust levels, constraints	prompt-structure
Few-shot, structured output, tool selection	examples-tools
Images / video in context	multimodal
Key fact ignored mid-context	critical-position
A model-generated claim treated as truth	safety

If no intent matches, note "no matching pattern" and proceed with the Four Operations

• Decision Framework — do not stall or invent a reference.

Host-Runtime Operations (host-specific)

The only host-coupled choices. Everything else above is host-agnostic.

Situation	Move	Why
Window long, task continues, keep thread	/compact	Summarize in place; costs latency — skip on small tasks.
Switching to an unrelated task	/clear	Fresh window; only after durable state is Written out.
Independent sub-scope, would bloat this window	Sub-agent	Isolated window; returns a distilled result.

Output Format

When this skill drives a decision, emit one line:

Context decision: <read-more | stop | ask | assume | delegate> — <one-line reason>

For assume, follow it with the ASSUMPTIONS I'M MAKING block and → Correct me now or I'll proceed with these.

Gotchas

• Compacting at ~50% is safe but wastes latency on small tasks — the threshold is a heuristic, not a rule.
• Bad routing turns this into "a dump with extra steps" — load ONE reference section per decision, never several.
• The risk thresholds here are labeled heuristics, not measured baselines — don't quote them as data.

Related Skills

• mk:context-audit — use instead for structural .claude/ overhead (tokens/% statically loaded).
• mk:memory — use instead to actually read/write canonical .meowkit/memory/ JSON stores; this skill only decides a write is warranted.
• mk:scout — use instead to discover where code lives; this skill decides how much to read.
• mk:agent-detector — use the canonical agent inventory for routing and on-demand instruction loading.