shinpr/llm-friendly-context

LLM-Friendly Context

The goal is stable downstream execution: the next agent should know what to read, what to do, what counts as success, and when to stop or escalate.

Core Rules

1. Use positive, executable instructions

   • State what the next agent should do.
   • Convert quality policies into positive 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. Make vague instructions concrete

   • Replace subjective terms with observable conditions, paths, commands, schemas, examples, or decision rules.
   • Terms that often need clarification when they leave a decision to the next agent: appropriate, proper, related, existing behavior, optional, as needed, if needed, per convention, unresolved alternatives, TBD, placeholder.

3. Specify output shape

   • Define required sections, fields, table columns, JSON keys, or checklist items.
   • For handoffs, include paths to produced artifacts and the exact status fields the caller must inspect.

4. Provide necessary context

   • Include the purpose, source artifacts, hard constraints, accepted decisions, and unresolved conditions.
   • Prefer concrete file paths and section hints over broad module names.
   • Follow references while they can change an in-scope decision, action, or verification result.

5. Decompose complex work into verifiable steps

   • Split work with 3+ objectives or sequential dependencies into ordered steps.
   • Each step needs a checkpoint: what evidence proves it is complete.

6. Permit uncertainty explicitly

   • If the source material is missing, contradictory, or not verifiable, state the uncertainty and the required escalation.
   • Record unknown business, product, security, or compatibility decisions as blocking unresolved items with the input needed to resolve them.

7. Keep constraints proportionate

   • Add only constraints that reduce ambiguity or preserve a real requirement.
   • Keep simple downstream tasks lightweight when the target action, context, and success criteria are already clear.
   • Apply minimal, a few lines, and explicit line estimates to the completed diff as one total budget.

Rewrite Patterns

Use these rewrites before treating a prompt, handoff, or artifact as complete.

| Ambiguous form | Rewrite as | |---|---| | optional used as an unresolved choice | Required, omitted, or required only under a named condition | | Multiple alternatives that the next agent must choose between | The selected option, or a deterministic decision rule | | as needed / if needed | The triggering condition and required action | | per convention | The file, function, test, or documented convention to follow | | related files | Specific paths, globs, or search hints | | existing behavior | The observable behavior, source file, test, API response, or UI state to preserve | | placeholder | Exact temporary value/behavior, allowed dependencies, and verification expectation | | TBD used as a placeholder for required information | A blocking unresolved item with owner, required input, or escalation condition | | appropriate / proper | A measurable criterion or checklist |

Handoff Checklist

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

• [ ] The target action is explicit.
• [ ] Required input paths and source artifacts are named.
• [ ] Accepted decisions and constraints are stated once, without alternate wording.
• [ ] Output format or expected status fields are specified.
• [ ] Success criteria are observable.
• [ ] Ambiguous expressions have been rewritten or marked as unresolved.
• [ ] Each instruction states the allowed action; each retained prohibition names the protected condition and allowed alternative.
• [ ] The next agent can complete its scope with explicit choices, decision rules, or blocking unresolved items.

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.
• [ ] Each instruction states the allowed action; each retained prohibition names the protected condition and allowed alternative.
• [ ] If an artifact is derived from another artifact, copied decisions stay consistent in wording and meaning.
• [ ] If downstream work is blocked by missing information, the artifact records the missing input and escalation condition.