Ground

Ground Claims with Research

Prevent ungrounded confabulation when producing methodologies, frameworks, scoring systems, taxonomies, or best-practice documents. Encode the diverge-converge research pattern proven during prioritization skill creation.

Core principle: Internal reasoning alone confabulates plausible structures. External research alone produces generic frameworks. Parallel diverge + synthesis produces grounded, project-adapted output.

Procedure

Phase 1: Scope

Frame the research question before searching.

• Define the topic: "What established frameworks exist for [X]?"
• Set inclusion criteria: relevance to project context, actionable methodology, procedural (not pure theory)
• Set exclusion criteria: wrong domain, no procedure, outdated
• Select parameters (consult references/grounding-criteria.md for guidance):
  • Internal scope: codebase (descriptive — "what patterns exist for X") or conceptual (generative — "what dimensions/constraints apply to X")
  • Model tier for internal branch: codebase → sonnet; conceptual → opus (generative divergence)
  • Research breadth: narrow (1-2 searches) or broad (3-5 searches)
  • Output format: reference document, skill body, or decision entry

Phase 2: Diverge

Execute two branches as parallel Task agents. Launch both in a single message. Neither is optional — both are required.

Branch A — Internal explore: Scope determines agent type and model:

• Codebase scope: Delegate to a Task agent (subagent_type: general-purpose, model: sonnet). Surface existing codebase patterns, prior decisions, current conventions. Output: inventory of existing patterns with file references. Write to plans/reports/<topic>-internal-codebase.md.
• Conceptual scope: Delegate to Task agent (subagent_type: general-purpose, model: opus). Generate project-specific dimensions, constraints, desiderata, evaluation axes. Focus on what no external source would surface. Write to plans/reports/<topic>-internal-conceptual.md.

Branch B — External research: Delegate to Task agent (subagent_type: general-purpose, model: sonnet). Model is always sonnet. Agent prompt includes:

• Research topic and framing question from Phase 1
• Search query templates from references/grounding-criteria.md
• Evaluation criteria: relevant to project context? Actionable methodology? Named framework with structure?
• Output contract: framework names, component structures, scoring mechanics, known limitations

Write to plans/reports/<topic>-external-research.md.

Both agents write reports and return filepath only (quiet execution pattern).

Phase 3: Converge

Synthesize both branches into grounded output.

• Map internal dimensions onto external framework skeleton
• Adapt scoring criteria or structural elements using project-specific evidence
• Apply convergence template from references/grounding-criteria.md (Framework Mapping, Adaptations, Grounding Assessment, Sources)
• Attach grounding quality label (mandatory): Strong / Moderate / Thin / None. See references/grounding-criteria.md for label definitions and evidence requirements.

Framing rule — general first: State each principle as the general insight derived from external research. Project-specific implementation is an instance that validates the principle, not the principle itself. The internal branch confirms the principle applies locally; it does not define the principle.

• ❌ "Git state queries return booleans — _git_ok(*args) -> bool"
• ✅ "Expected states should return booleans, not raise exceptions — implemented as _git_ok() in this project"

Phase 4: Output

Write the grounded reference document to plans/reports/<topic>.md (persistent, tracked).

Required sections:

• Research foundation (named sources with links)
• Adapted methodology — general principle per item, then project implementation as instance
• Grounding quality label with evidence basis
• Sources section with retrieval context

Branch artifacts: Both branch reports in plans/reports/ are retained as audit evidence supporting the synthesis. The grounding report references both.

References

• references/grounding-criteria.md — Trigger criteria, quality labels, parameter selection, search query templates, convergence template