Write Plans

Write for a capable executor who will have less conversational context and more evidence than you do now. Carry forward intent, a useful problem representation, a repertoire of methods, and the grounds for choosing the next action. Use high-salience domain concepts to compress shared knowledge and make room for breadth. Supply the local distinctions that change their application.

Select and combine the lenses below as the task warrants. Scale detail to uncertainty, dependencies, and consequences. Let the requested artifact determine the structure; a short plan can be sufficient. Match the surrounding assignment: deliver a plan when planning is requested, or use it to continue work already authorized.

Intent; constraint satisfaction; acceptance criteria. Recover the desired outcome, its purpose, scope, and consequential user preferences. Distinguish requirements from preferences and proposed solutions. Define success through observable results or artifacts. For exploratory work, specify the question and what would constitute an informative result without assuming the conclusion. Preserve why a constraint matters so later choices can honor its purpose. Keep uncertainties about the goal separate from uncertainties about implementation.

Problem representation; affordances; analogical transfer. Expose the relationships that make the work tractable: state transitions, dependencies, invariants, audiences, incentives, representations, or asymmetries between construction and verification. Name applicable methods and useful adjacencies, including approaches whose value may become apparent later. Connect them through shared structure, prerequisites, alternatives, and tradeoffs. Incremental derived state suggests recomputation oracles; compatibility constraints suggest staged migration; competing explanations suggest discriminating experiments. Teach recognition beyond the initial examples. Distinguish committed actions from the available methodological repertoire.

Epistemic state; provenance; decision records. Separate observations, inferences, assumptions, proposals, and decisions already made. Ground consequential claims in inspectable artifacts; record revisions or dates where freshness affects the choice. Preserve the rationale for significant decisions, including rejected alternatives when the reason for rejection could change. Carry forward corrections and constraints established in the conversation. An unresolved premise should remain visible as a question, assumption, or proposed check; confident prose must not convert it into an established fact.

Backward chaining; means–ends analysis; proof obligations. Work backward from the intended result to the evidence, intermediate capabilities, and actions needed to obtain it. Explain how consequential steps contribute to the outcome. Make the first useful move executable with the available context: an entry point, a question to resolve, a candidate method, or a bounded change with a meaningful check. Specify interfaces and failure-sensitive details where they constrain correctness; leave ordinary implementation choices to the executor. Carry useful partial results forward when a later question remains open.

Partial-order planning; causal links; critical path. Order work by real dependencies. Identify what each substantial work item needs and produces, and which actions could invalidate another item's prerequisites. Distinguish necessary ordering from convenient sequencing. Where parallel work helps, define independent responsibilities, shared interfaces, resource constraints, and integration evidence. Account for shared mutable artifacts and duplicated effort. Make the current bottleneck visible so new information can change the allocation of work.

Contingent planning; hypothesis discrimination; value of information. Turn important uncertainties into observations that can change a decision. Pair a consequential question with a useful probe, interpretation of its possible results, and the resulting branch. Prefer investigations that discriminate between plausible approaches or unlock several dependent tasks. Keep independent work moving while a question is unresolved. Spend planning and research effort in proportion to their expected effect on the next decision; stop elaborating when the executor has a sound next move and useful grounds for adaptation.

Assumption-based planning; signposts; sensitivity analysis. Identify assumptions whose failure would materially change the approach. For uncertain, consequential assumptions, describe observable signs that warrant reassessment and the affected decisions. Consider how the plan changes if a dependency behaves differently, a resource is unavailable, or an early hypothesis is wrong. Preserve useful fallback options and actions that reduce dependence on uncertain premises. Investigate plausible, consequential failure modes in proportion to the task; ordinary work need not become a catalog of hypothetical disasters.

Receding horizon; least commitment; option value. Make near-term work concrete and later work conditional on what it will reveal. Defer expensive commitments whose justification depends on missing evidence; retain choices that make subsequent revision cheap. Refine the plan as uncertainty resolves. Treat estimates as estimates and tie checkpoints to useful evidence or deliverables. Avoid inventing deadlines, budgets, or certainty about implementation details the task has not established.

Feedback control; observability; independent verification. Connect actions to expected observations and criteria for continuing, changing approach, or finishing. Separate activity measures from evidence that the intended result holds. Choose checks appropriate to the claim: executable regressions, independent calculations, source checks, representative examples, or editorial review. If a check or search is quiet, consider whether it could detect the relevant failure before treating silence as assurance. Keep proposed validation separate from completed validation and retain the limitations of the evidence.

State externalization; prospective memory; dependency invalidation. Make the plan resumable: retain the current objective, consequential decisions, completed evidence, active work, next useful actions, and unresolved dependencies. Link durable artifacts instead of copying transcripts or tool output. When evidence invalidates a premise, revise the downstream approach and mark superseded decisions where their history still matters. Keep current guidance coherent. Preserve enough context to resume without repeating completed work or mistaking a stale instruction for the current plan.

Counterfactual review; executable walkthrough; proportionality. Read the plan as a fresh executor. Can you begin, choose among methods, inspect the evidence, and recognize completion? Perturb an important assumption and check whether the plan supports an intelligent response. Look for missing causal links, circular dependencies, unsupported assertions, and steps whose outcome would not affect subsequent work. For simple tasks, check that the plan remains simple. Trim repeated exposition while retaining distinct useful concepts, consequential constraints, and reasons that would otherwise be lost at handoff.

Use references/methods.md when the underlying planning concepts or their limits need grounding. The repertoire remains open: recruit other methods when they make the work easier to recognize, investigate, or complete.