meta-prompt

name	meta-prompt
description	Recursive meta-prompt ledger for human-guided complex tasks spanning multiple sessions. Maintains RELIANCE.md — a single-source, append-aware file carrying FINALITÉ, the RAIL session-chain, and the decision/collaboration trail across context resets. Use when: a complex task spans sessions and you want loss-free hand-off + drift detection + proportional AI challenge, WITHOUT autonomous loops or heavy regression tests. Triggers: meta-prompt, reliance, start a guided task, resume reliance, close session, next RAIL.
argument-hint	[init|resume|close] [task-name]
allowed-tools	["Bash","Read","Write","Edit","Glob","Grep","AskUserQuestion","Skill"]
model	opus
effort	high
context	main
user-invocable	true

Meta-Prompt — RELIANCE ledger

Engine for RELIANCE.md: a recursive ledger for human-guided complex tasks spanning multiple sessions. NOT for autonomous loops or dev with heavy regression tests. Read reference.md for the full model, drift protocol, challenge dials, pick-model embedding, and articulation.

Stance: single source on disk (each session RE-READS the file, no copy travels → no drift) · propose, don't emit (end by proposing the next step, wait for the human gate) · three altitudes of "good": FINALITÉ (whole run) → RAIL.landed (session gate) → HEADING.next (pebble) · RAIL = one full session, future RAILs are stubs · the live file is token-lean (procedure lives in this skill).

Concepts (full legend → reference/morin-glossary.md): RELIANCE re-link what division separates · FINALITÉ orienting why · HEADING revisable now · LOOP output-becomes-input (append-only trail + 🤖 AI behaviour) · DIALOGIC hold the tension · ÉCO coupled to files, re-organise on drift · MAIN human inbox (not Morin) · RAIL = instruction payload (acronym).

Modes

$ARGUMENTS = {init|resume|close} [task-name]. No mode → infer: no RELIANCE.md → init; else resume.

Mode	Does
init	Elicit FINALITÉ + RAIL-0 (role·bounds·ready·landed) + coarse RAIL-chain (ask human, don't invent scope) → /pick-model for RAIL-0's model: field (recommendation, human confirms) → list hot files by tier + fingerprint → emit lean RELIANCE.md from reference/RELIANCE-template.md §LLM SKELETON (zones+data only, strip all prose/diagrams) → confirm before first work.
resume	Run the contract IN ORDER: 0 MAIN (drain inbox → ack to LOOP → clear; overrides drift) · 1 ÉCO (drift-check ref/wip → back-sync+flag) · 2 LOOP (orient from trail, esp. logged failures) · 3 HEADING (propose next pebble, cite LOOP) · 4 DIALOGIC (Dial 1 mandate + Dial 2 one lens) · 5 surface + wait (no proceed without human).
close	1 append LOOP two-level entry (append-only, never prune) · 2 mark NOW done · 3 gate RAIL-0 LANDED; if met → detail next RAIL (READY = this LANDED) + re-run /pick-model; else keep open · 4 if /save-context active, thin its Next/Decisions to → see RELIANCE.md.

Gate granularity: gate at the boulder boundary, not every pebble — pebbles flow light-touch once a boulder is confirmed.

Anti-patterns

• ❌ Copying the plan into the prompt (travelling state → drift). Single disk source only.
• ❌ Pruning/rewriting LOOP — append-only and sacred.
• ❌ Emitting the next step without the human gate (that's /goal, not this).
• ❌ Pre-scripting within-session steps — the loop generates the next pebble; only the RAIL-chain is mapped ahead.
• ❌ Loading the rich template into context — it lives in reference/, Read on demand at init only.
• ❌ Drift-checking park-tier or non-hot files — mutability tier = drift budget.

// future: per-RAIL autonomy flag — a RAIL whose LANDED is machine-checkable could run /goal-like. // RELIANCE = the human-in-the-loop superset of /goal.