Step 4 · Playbook·Operator workflow ENPT
Alembic · Blind Planning

Blind prompt playbook

How to paste prompts, compare N independent plans, score with rubric R1–R10, and require all 16 delivery sections in PLAN.md.

1

Operator workflow

Goal: N comparable plans from N clean agent sessions. Same prompt pair, same rubric, same 16 sections — then aggregate.

  1. New session — no prior Alembic RE context, no gist, no docs from other agents as truth.
  2. Paste both files — full PROMPT-BLIND-PLANNING.md (or <prompt> block) + entire PROMPT-CORPUS-EMBEDS.md.
  3. Agent starts Phase 1 — RE-A (Droid) first; do not skip to implementation.
  4. Output path — outputs/alembic-blind-plan-<agent-id>/PLAN.md + reverse/ artifacts.
  5. Compare — score each plan R1–R10 (0–2); check all 16 sections present.
2

Rubric R1–R10

Self-score 0–2 per criterion at end of each plan. Source: PROMPT-BLIND-PLANNING.md § Fase 4.

IDCriterion02
R1Three REs with binary + disk + web evidenceMissing REAll three forensically proven
R2Unified synthesisContradictionsClean fusion table
R3Loop Engineering operationalizedCited onlyGates wired in runtime
R4@alembic/harness spine reuseGreenfieldExplicit package map
R5Operator interfaceBackend onlyCLI + TUI + web mockups
R6alembic.plan.ts specSketchFull hooks API
R7~/.alembic/ layoutVagueComplete tree
R8API protocolPartialREST/SSE/MCP/CLI
R9Slices A0–A9No proof gatesMeasurable per slice
R10Comparable structureAd hocAll 16 sections below
3

16 delivery sections

Every PLAN.md must contain exactly these sections (markdown headings):

  1. Executive Summary (1 page)
  2. RE-A: Droid Missions
  3. RE-B: Claude Workflows
  4. RE-C: Kimi AgentSwarm
  5. Synthesis: unified table + fusion rules
  6. Architecture Alembic (mermaid diagrams)
  7. Loop Engineering: gates in runtime
  8. State on disk + event schema
  9. alembic.plan.ts + hooks API
  10. Operator interface (CLI + TUI ASCII + web wireframe)
  11. API protocol (REST, SSE, MCP, CLI)
  12. Existing code reuse map
  13. Slices A0–A9 with dependency DAG
  14. Risks, forks, user-gated decisions
  15. Rubric auto-score + honest gaps
  16. Appendix: commands executed + key outputs
Save location 
outputs/alembic-blind-plan-codex-gpt55/PLAN.md
outputs/alembic-blind-plan-codex-gpt55/reverse/
# Replace agent-id: codex-gpt55, kimi-k2, grok-4, ...
bash docs/alembic/build-corpus-embeds.sh
# After new brightdata scrapes → corpus-cache/*.md

Working language: reasoning + deliverables in PT-BR; code, commands, identifiers in English (per blind prompt).

4

Handoff checklist

You can now:
  • Paste prompt + embeds to a fresh agent session
  • Enforce evidence discipline (Lesson 2) during review
  • Judge three-chord fusion quality (Lesson 3)
  • Score plans with R1–R10 and verify 16 sections

Primary sources: docs/alembic/PROMPT-BLIND-PLANNING.md · docs/alembic/PROMPT-CORPUS-EMBEDS.md · loop-engineering/SKILL.md