Skill Narrative

Linguistic style modes (required)

Use explicit mode labels so tone is intentional and stable:

• operator-brief (default internal mode)

  • Tone: sober, analytic, compact.
  • Use for: synthesis diagnostics, decision support, ledger rationale.
  • Constraint: minimal rhetoric; emphasize clarity and provenance.

• narrative-stance (distinct narrative mode)

  • Tone: slightly warmer, human-scale, forward-leaning.
  • Use for: principal memo paragraph and narrative posture lines.
  • Constraint: still evidence-bound; no theatrical language; no persona roleplay.

• public-hook

  • Tone: concise, persuasive, high-signal.
  • Use for: short hooks/candidate posts.
  • Constraint: draft-only; requires human review before publication.

Mode policy:

• Always produce outputs with mode labels.
• Keep operator-brief for governance/decision blocks.
• Require at least one narrative-stance paragraph each run so this skill remains distinct from pure operator analytics.

Method

1. Ingest last N briefs (N=3 default, N=7 in high volatility).
2. Extract:
   • stable themes,
   • new deltas,
   • stale/repeating frames.
3. Draft synthesis + content outputs.
   • Synthesis can remain operator-brief.
   • Principal narrative paragraph must be narrative-stance.
   • Short hooks should be public-hook.
4. Log only high-value decisions:
   • high-impact + high-uncertainty,
   • companion/agent disagreement,
   • repeated stale frame.
5. Record each decision with rationale and next test.
6. For deferred items, set a required review date.

Convergence Ledger schema

Each entry should include:

• id
• timestamp
• domain
• companion_signal
• agent_interpretation
• decision (accept|defer|reject)
• rationale
• confidence
• impact
• next_test
• review_by (required if decision=defer)
• supersedes (optional)
• source_briefs

Cadence

• Daily: run synthesis + content pack + 1-3 ledger entries.
• Weekly: sweep all deferred entries (close|extend|promote) and archive resolved ones.

Guardrails

• WORK-only process layer; do not treat outputs as Record truth.
• No automatic promotion to SELF/EVIDENCE.
• Public-facing copy still requires human review and source checks.
• If an item has no clear next_test, usually do not log it.

Cursor / grace-mar instance

Grace-mar paths and runbook links for this repository (from .cursor/skills/skill-narrative/).

Run reminder: keep ledger entries sparse (1-3 high-value items per run) and enforce review_by on all defer decisions.