| name | structure-planning |
| description | Use when a report, experiment plan, Eval output, presentation storyboard, PPT/deck plan, document, paper, HTML view, or refactor needs a structure contract before prose, rendering, interpretation, follow-up runs, or edits. |
Structure Planning
Tool Commands
Use the command packet before applying this skill's workflow:
python3 tools/agent_tools/skill_tool_commands.py show --skill structure-planning --format text
Execute the required and task-matching conditional commands that the packet prints.
- Read
agents/skills/structure-planning.md.
- Use this before drafting prose, writing a renderer, interpreting experiment output, planning a presentation/deck, planning follow-up runs, or editing refactor surfaces when the work has a nontrivial structure.
- For document additions or revisions, use this before prose edits when section order, responsibility, claim/support, reader path, source map, or canonical route changes; for typo/link/format-only edits, record
structure_contract=skipped with the reason and use $md-style-check.
- Create a structure contract with
structure_kind, audience, decision context, first artifact, first artifact question, visual_plan, source-to-structure map, document_unit, document_split_decision, OOP structure contract when planning experiments, metric or delta contract, ordered structure, invalid interpretations, and validation gate; for deck/PPT work, include slide/storyboard order and source-to-slide mapping.
- For document structure changes, fill
document_unit with owner, reader, source map, validation route, update cadence, canonical parent, and downstream consumers. Set document_split_decision to keep, split, merge, inline, rename, or not_applicable:format-only before prose edits.
- Choose
split only for a new owner, reader, validation route, source map, update cadence, or downstream consumer. Choose merge, inline, or keep when the same owner, reader, source map, validation route, and update cadence remain shared. Treat length, token budget, chunking convenience, section count, nearby path, temporary work queue, and shared validation oracle as invalid split boundaries.
- For
experiment-plan and experiment-report, structure from the OOP responsibility view before prose order: list reused modules/classes/functions/protocols, object creation/mutation/pass-through/artifact writes, the factory/function boundary where variants differ, and dependency direction across orchestration, domain logic, metrics, visualization, and artifact I/O.
- When paragraph/block order, connective choice, or logic-gap evidence is nontrivial, run or request
agent-canon semantic-index discourse-relations --profile <general|experiment-report|methods-protocol|academic-argument|refactor-design> --format jsonl after the semantic index is built; use it as advisory structure evidence, not as prose or policy authority.
- If a prose graph DB or projection is present, use
prose_reasoning_graph.py explain and integrate as advisory evidence for paragraph bridges, split/merge/reorder operations, and invalid interpretations.
- Choose the first artifact before implementation: figure, table, ponchi-e/concept diagram, slide, summary card, first section, experiment slice, or refactor slice.
- For reader-facing documents, reports, plans, workflow guides, and refactor maps, choose Mermaid as the default first visual when the structure includes nontrivial process flow, dependencies, ownership, routing, state transitions, review gates, or multi-step handoffs; choose
text-only only when a diagram would duplicate a simple list and record that reason in visual_plan.
- Map every source artifact to the section, slide, visual, claim, experiment slice, or refactor slice it supports; do not let unsupported claims or edits appear later.
- Define metric denominator, directionality, baseline, and caveat for reports or experiments; define allowed structural delta and forbidden semantic delta for refactors.
- Put sections, slides/storyboards, visuals, experiment slices, or refactor slices in reader or execution order rather than raw tool-output order.
- Record invalid interpretations so the structure cannot be mistaken for policy, classification, merge, deletion, ownership, or behavior-change authority.
- Hand the completed structure contract to
$report-writing, $html-output, $html-experiment-report, $experiment-lifecycle, slide-production workflow, $long-form-writing, $academic-writing, $paper-writing, or $refactor-loop as appropriate; this skill owns structure, not raw storage, experiment execution, report prose, document drafting, implementation, or domain authority.
- Record closeout tokens:
structure_planning=complete, structure_contract=<path-or-inline>, document_split_decision=<keep|split|merge|inline|rename|not_applicable:format-only>, structure_first_artifact=<name>, structure_visual_plan=<mermaid|table|text-only|html|image|slide|not-applicable>, structure_source_map=<path-or-inline>, structure_oop_contract=<path-or-inline|not_required>, discourse_relations=<path|not_required>, and structure_invalid_interpretations_recorded=yes.