/imperatives

Extract atomic imperatives (MUST/SHOULD/MAY) from markdown instruction files into JSONL.

Usage

/imperatives                                # default: ai-workspace/rules/*.md + AGENTS.md
/imperatives .claude/rules/*.md             # specific globs
/imperatives --output imperatives.jsonl     # write to file

Steps

1. Resolve files. Default: ai-workspace/rules/*.md and AGENTS.md. Expand globs to absolute paths. Zero matches → error, stop.

2. Pass 1 — Script extraction (fast, deterministic).

   node --import tsx "${SKILL_DIR}/scripts/extract-imperatives.ts" <files...> --output /tmp/imperatives-pass1.jsonl
   

   Copy

   ${SKILL_DIR} = this skill's base directory (shown in the "Base directory for this skill:" header above). Keyword-based: catches RFC 2119 terms + imperative verbs. ~80% coverage, <1s.

3. Pass 2 — Subagent reasoning (catches implicit imperatives). For each file, dispatch a subagent to identify imperatives the regex missed — contextual rules, implicit constraints, prose-embedded obligations.

   Subagent prompt:

   Read <file>. Here are the imperatives already extracted by the script for this file:

   <pass 1 JSONL lines where source.file matches>
   

   Copy

   Identify additional imperative statements the regex missed. Look for:

   • Implicit obligations hidden in prose (e.g., "The workflow is a menu, not a pipeline" implies MUST NOT enforce order)
   • Contextual constraints (e.g., section context that makes a statement imperative)
   • Conditional rules not triggered by keyword patterns

   For each new imperative found, output one JSON line (same schema — id, level, polarity, subject, predicate, when, source, tool_scope, tags, raw). Do NOT duplicate entries already in pass 1. If none found, output nothing.

   Use subagent_type: "explorer" (Haiku) for most files. If a file had zero pass-1 imperatives OR contains dense prose (>50 lines without a keyword match), escalate that file to Sonnet for deeper reasoning.

4. Merge passes. Concatenate pass 1 + pass 2 outputs. Deduplicate by raw text. Write final output to --output <path> or present inline.

5. Present summary. Report in a table:

   • Total count (pass 1 baseline + pass 2 additions)
   • Breakdown by level and tool_scope
   • Files with zero imperatives after both passes
   • Pass 2 additions highlighted separately

6. Downstream callers. If invoked by /policy-algebra or /distill, return the JSONL path directly — skip the summary.

Failure modes

Cross-tool notes

• Codex / Cursor: run the script directly — it's tool-agnostic.