LLM-Friendly Context

Purpose

Use this skill when writing or revising any content another agent will execute or judge: prompts, handoffs, planning artifacts, review findings, completion reports, generated instructions, test skeleton comments, work plans, and task files.

The goal is stable downstream execution. The next agent should know the target action, required inputs, accepted decisions, observable success criteria, and the condition that requires escalation.

Core Rules

1. Use positive, executable instructions
   • State the action the next agent should perform.
   • Convert quality policies into observable acceptance criteria.
   • Keep a prohibition only when it protects an irreversible boundary or shipped contract. Name the protected condition and the allowed action.
   • Example: Preserve existing public API behavior across the documented compatibility cases.
2. Resolve outcome-relevant ambiguity
   • Clarify an instruction when plausible interpretations would change correctness, requested scope, downstream usability, or verification. Use the least-restrictive sufficient condition, source, or example.
   • Leave local, reversible choices to the next agent when the stated outcome, constraints, and repository evidence are sufficient to choose. A subjective word alone does not require another rule or decision.
3. Specify output shape
   • Define only the sections or fields the next consumer uses.
   • For agent handoffs, name produced artifact paths and the result needed by the next action. Require exact serialization only when a program parses it.
4. Provide the smallest sufficient context
   • Self-contained means sufficient for the assigned action, not complete background.
   • Include only context directly consumed by the target action or required result.
   • Pass a governing source by its path or an unchanged excerpt. A governing artifact already supplies its own purpose, criteria, and decisions; a restatement adds an unreviewed second version of them.
   • Extra context can anchor judgment, contaminate independent work, and create work.
5. Decompose complex work into verifiable steps
   • Expose dependency order when a later action relies on an earlier result.
   • Reuse one execution plan to retain all required steps and final verification during multi-step work. Simple, single-action work proceeds directly.
6. Permit uncertainty explicitly
   • State missing, contradictory, or unverifiable source material and its effect on the current action.
   • Return unresolved decisions to the orchestrator with the evidence needed to resolve them; the orchestrator decides whether user input is necessary.
7. Keep constraints proportionate
   • Add constraints that reduce ambiguity or preserve a real requirement.
   • Keep simple downstream tasks lightweight when target action, context, and success criteria are already clear.

Rewrite Patterns

Use these rewrites when an ambiguity materially changes the next action or its result. Retain valid local choices within the stated boundaries.

Handoff Checklist

Before sending a prompt or artifact to another agent, verify:

• The target action is explicit.
• Every included context item is directly consumed by the target action or required result.
• User constraints are distinguished from revisable technical choices; include the necessity of selected means where the next decision depends on it.
• The next consumer can identify the artifact or result it needs.
• Success criteria are observable.
• Outcome-relevant ambiguities are resolved or identified with their effect; valid local choices remain available.
• Every retained prohibition names the protected condition and allowed alternative.
• Dependencies needed by the next action are visible.
• The next agent can complete its scope or return the unresolved decision with evidence.

Generated Artifact Checklist

Before writing or finalizing a generated document:

• Each requirement, claim, task, test skeleton, or review finding has enough source context to trace why it exists.
• Every executable instruction names the target, action, and expected result.
• Verification steps say what to run or observe and what result proves success.
• Every retained prohibition names the protected condition and allowed alternative.
• Derived artifacts preserve user-required outcomes and explicit constraints. When superseding a technical choice, identify and reconcile affected sources rather than copying an obsolete obligation.
• Blocking missing information records the missing input and escalation condition.

References

• Task File Contract — execution-carrier fields and filename routing for task decomposition, build recipes, and Small flows