Design reliable context with ak:context-engineering

Use ak:context-engineering to reason about context budgets, degradation, compression, memory, tools, and agent coordination. It starts as guidance: the Skill selects the relevant source material and proposes a measurable change. It produces local artifacts only when you ask it to run its bundled analyzers or to write a design, summary, or memory file.

Choose ak:context-engineering for context problems

Use ak:context-engineering when

• A long session is losing important facts, mixing tasks, or approaching its context limit.
• You need to reduce token use without losing decisions, file history, or next steps.
• You are designing retrieval, memory, tool descriptions, or context isolation for an agent system.
• You want a probe-based way to compare an original context with a compressed summary.

Choose another workflow when

• You need repository facts or file locations. Use ak:scout first.
• You need current external documentation. Use ak:docs-seeker.
• You need implementation rather than a context design. Move to planning or ak:cook after the proposal has measurable acceptance criteria.
• You only need the runtime's authoritative quota display. Check that runtime's own status and account controls; the Skill cannot invent unavailable usage data.

Prepare the evidence and runtime

Before you start:

• Complete Onboarding, and confirm Engineer Kit is installed for the runtime and scope you use.
• State the failure you observe, the model or runtime limit when known, and the evidence that must survive any compression.
• Identify sensitive messages, tool output, credentials, or customer data that must not be copied into summaries, files, or external services.
• For script-based analysis, prepare UTF-8 JSON containing either a message array or an object with a messages array.

See Runtime adapters before treating a runtime-specific usage signal as portable evidence.

Run the Skill

/ak:context-engineering "Our tool outputs dominate a long debugging session. Design a compaction policy that preserves errors, decisions, changed files, and next steps; do not write files."

/ak:context-engineering "Our tool outputs dominate a long debugging session. Design a compaction policy that preserves errors, decisions, changed files, and next steps; do not write files."

$ak:context-engineering "Our tool outputs dominate a long debugging session. Design a compaction policy that preserves errors, decisions, changed files, and next steps; do not write files."

The Skill accepts a topic or question. It has no published mode flag. Ask for a specific output such as a diagnosis, budget, compression policy, memory design, tool review, or evaluation plan.

Optional local analyzers

These Python utilities read files up to 100 MiB and do not create output files on their own. Shell redirection or a separate write request changes that boundary. The evaluator is heuristic; comments in the source reserve real model-based judging for a production implementation.

Understand what happens during a run

1. The Skill identifies the context failure. It separates capacity pressure from distraction, contradiction, poisoning, lost-in-the-middle behavior, or missing persistence.
2. It measures before optimizing. It uses available runtime evidence or a supplied context file and labels estimates as estimates.
3. It selects a bounded strategy. The source organizes choices as write, select, compress, and isolate, then loads only the relevant reference.
4. It preserves the task contract. Goals, decisions, constraints, artifact history, current state, and next steps remain explicit.
5. It proposes or performs the requested operation. Advice stays in the conversation; analyzers read local files; requested designs or summaries may write to an approved path.
6. It checks quality. Probe-based checks test factual recall, artifacts, continuation, and decision rationale instead of trusting token reduction alone.

Keep context and credentials safe

• Reading local context, summaries, or memory files requires access to those paths; writing or replacing them requires an explicit artifact boundary.
• Claude-specific usage awareness can read local status data and use an Anthropic OAuth credential to call the provider's usage endpoint. Do not copy credential values into prompts or reports.
• Context isolation through another agent can multiply model use and cost. Ask before adding paid providers, broad parallelism, or external storage.
• A proposed memory system does not authorize collection, retention, or cross-session storage of personal data.
• Do not discard the original context or overwrite a trusted summary until the compressed result passes the required probes.

Verify the result

A complete guidance run should provide the observed failure, evidence used, selected strategy, preservation contract, expected token or latency effect, and a test that can falsify the proposal. A script run should also report input files, exact operation, exit status, JSON output, and any estimate or heuristic limit.

Treat the result as evidence only when critical facts remain recoverable, artifact history and next steps survive, contradictory or stale material is identified, and the proposed change does not silently expand persistence, provider access, or cost.

Troubleshoot or continue

Know the current limits

• Token estimates use simple character-based approximations in the bundled scripts; they are not provider billing records or tokenizer-exact counts.
• Attention, degradation, cost, and performance figures in the source material are design guidance, not measurements of your system.
• The Skill does not automatically create a context bundle, memory store, graph, or evaluation dataset.
• External model judging, retrieval services, and persistent stores require separate tools, credentials, privacy review, and cost approval.
• Stable and beta package identical ak:context-engineering Skill content and bundled analyzers.