process

Process - Structured Document Build

Meta-skill for structured document production. Generates a tailored program (INSTRUCTIONS.md + BENCHMARK.md), scaffolds the WIP folder, executes the four-phase workflow. Does not process documents directly - it orchestrates.

Scope boundary. This skill builds a deliverable from sources. It is not the verification path: validating a finished document against its source for grounding + tone/style/format compliance is the validate skill; running the grounding CLI over claims (single or batch via source_map.yaml) is the grounding skill. The Verify & Ground phase below invokes the grounding skill rather than re-implementing grounding.

Pre-flight install (MANDATORY - run every session, no asking)

Always run this single line BEFORE invoking document-processing. No-op when the package is already importable; auto-installs when missing OR when a stale shim is on PATH but the package is uninstalled in the active Python:

python3 -c "import stellars_claude_code_plugins" 2>/dev/null || python3 -m pip install --user --upgrade stellars-claude-code-plugins

Never ask the user whether to install - just run the line.

Invocation

/document-processing:process <objective description>

Objective describes what to produce. Examples:

• /document-processing:process reconstruct complete timeline from all court documents
• /document-processing:process draft 2-page court statement addressing mother's claims
• /document-processing:process extract and categorize all court findings by topic

Execution Flow

Phase 0: Objective Refinement

Step 0.1: Read all files in 1-input/ - catalogue sources. Read 4-references/ for examples and universal facts. Read .claude/CLAUDE.md for project context.

Step 0.2: Present consolidated clarifying questions:

1. End state: final shape? Format, length, audience, language?
2. Source scope: which input documents? All or specific?
3. Quality criteria: good vs bad output? Dealbreakers?
4. Constraints: what must NOT appear? Tone/perspective?
5. Grounding priority: how strictly must claims trace to source?
6. Format rules: structural requirements (date format, entry format, section order)?

Step 0.3: Iterate follow-ups until objective is crystal clear.

Step 0.4: Summarize refined objective for confirmation.

Phase 1: Program Generation

Load references:

• references/WORKFLOW.md - 3-phase workflow template
• references/GROUNDING.md - assumption classification + verification methodology for the draft (DIRECT QUOTE / PARAPHRASE / INFERENCE / INTERPRETATION / UNSUPPORTED)
• references/UNIFORMIZATION.md - quality control
• references/FOLDER-STRUCTURE.md - folder conventions

For CLI-assisted grounding of HIGH-impact claims during Phase 2 (Verify & Ground), invoke the grounding skill / document-processing batch-ground - it carries the deterministic three-layer lexical grounder (+ optional semantic), the OCR fallback chain, and the verdict rules. references/GROUNDING.md is the classification methodology; the grounding skill is the operational tool.

When deriving uniformization rules, load the real rule-set examples in ${CLAUDE_PLUGIN_ROOT}/examples/ - a full in-use INSTRUCTIONS.md with measurable rules R0-R4 (no-fluff, length range with a preferred band, focus exclusions with example quotes, text format, list-section format) plus a worked uniformization checklist. Use them to make every rule measurable (word ranges, per-bullet word counts, exclusion lists with example quotes, a falsifiable "does this sentence change what the reader knows" test). Adapt the shape, not the content.

Generate INSTRUCTIONS.md at project root:

1. Context: CLAUDE.md context + refined objective
2. Directory Structure: 1-input, 2-wip, 3-output with task-specific WIP subfolder
3. Source Documents: catalogue of relevant inputs
4. Uniformization Rules: task-specific R1, R2, R3... from:
   • Stated quality criteria
   • Domain conventions (legal citation, date formats)
   • Output format requirements
   • CLAUDE.md markdown/typography standards
   • The rule-set examples in ${CLAUDE_PLUGIN_ROOT}/examples/ (shape reference)
5. Workflow Steps: 3-phase pattern (Analyze & Draft -> Verify & Ground -> Uniformize & Deliver)
6. Phase Gates: user review points
7. Execution Modes: Interactive / Semi-automated / Headless

Present for approval. Do not proceed until approved.

Phase 2: Benchmark Generation

From approved INSTRUCTIONS.md, generate BENCHMARK.md:

1. Scoring approach: MINIMIZE penalty score (target: 0)
2. Programmatic checks: measurable, verifiable
   • Word count within range
   • Required sections present
   • Date format consistency
   • Forbidden patterns absent (grep-checkable)
3. Grounding checks: each HIGH-impact claim has source reference
4. Rule compliance: one checklist item per uniformization rule
5. Subjective quality (sparingly): only for non-measurable aspects, with rubric

Present for approval. Do not proceed until approved.

Phase 3: Scaffolding

• Derive task name from objective (kebab-case, e.g., timeline-reconstruction)
• Create 2-wip/<task-name>/
• Create 2-wip/<task-name>/README.md manifest (empty table)
• Ask: execution mode A) Interactive, B) Semi-automated, C) Headless

Phase 4: Execution

Execute INSTRUCTIONS.md step by step per mode.

• Show progress after each step
• All WIP in 2-wip/<task-name>/
• At each phase gate, display deliverables, wait for approval
• After Phase 3, evaluate against BENCHMARK.md
• Deliver final to 3-output/
• Update manifests in both locations

Rules

• Never modify 1-input/
• Intermediate artifacts in 2-wip/<task-name>/
• Only final documents in 3-output/
• 4-references/examples/ = format guidance, never copy content
• 4-references/facts/ = grounding anchors (legal provisions, precedents)
• Every factual claim needs source reference
• INSTRUCTIONS.md and BENCHMARK.md need explicit approval before execution
• Phase gates MANDATORY in interactive and semi-automated modes
• INSTRUCTIONS.md must be self-contained - inline task-specific rules, never just reference files