Skip to content

Fast LLM instruction plan

This document is a compact instruction set for generating planning artifacts quickly and safely.

It is intentionally strict. It exists to reduce ambiguity and avoid repeated planning rewrites.

  • In-scope: planning research, critique, document drafting, consistency audits, and governance updates.
  • Out-of-scope: code implementation tasks, runtime/build changes, or direct rollout execution.
  • Use this document as the default fast path for planning cycles.
  • Escalate to 03-weighted-deep-planning-manual.md when any section is W3 or W4, or when blocker-class ambiguity appears.
  • Keep both docs aligned on taxonomy, gate language, and authority references.
  1. Use canonical terminology from 06-planning-taxonomy-glossary.md.
  2. Follow authority hierarchy in 01-master-planning-index.md.
  3. Never mix implementation execution tasks into plan-authoring documents.
  4. Every plan section must define acceptance evidence.
  5. Complex sections must include explicit anti-foot-gun controls from 05-anti-foot-gun-planning-standard.md.
  • Gather source docs:
    • blueprint,
    • ADR 012,
    • architecture/lowering explainers,
    • governance and doc acceptance checklist.
  • Build a one-page “source-of-truth map” before drafting.
  • Produce severity-ranked findings.
  • For each finding: define root cause, risk mechanism, and correction strategy.
  • Map each correction to a target planning document.

Step 3: define plan information architecture

Section titled “Step 3: define plan information architecture”
  • Decide document set, authority tiers, and non-overlap boundaries.
  • Declare owner role per document.
  • Declare update cadence and review path.

Step 4: write specifications/templates first

Section titled “Step 4: write specifications/templates first”
  • Write task schema spec.
  • Write milestone/gate evidence spec.
  • Write deferral/exception policy.
  • Write anti-foot-gun planning standard.
  • Draft fast plan for short-cycle work.
  • Draft deep weighted manual for complex/high-risk work.
  • Ensure both plans reference the same taxonomy and gate model.
  • Check for contradictory gate names/threshold references.
  • Check for duplicate ownership claims.
  • Check for terminology drift.
  • Check for implementation leakage into doc-only artifacts.
  • Record version/update metadata.
  • Record unresolved issues and owner.
  • Publish corpus and read-order guidance.

Each planning document must include:

  • purpose statement,
  • scope boundaries,
  • authority tier,
  • acceptance criteria,
  • dependencies/cross-links,
  • owner role.

For high-risk documents (deep manual, gates spec, anti-foot-gun standard), also include:

  • failure modes,
  • stop conditions,
  • escalation path.

Stop drafting and request clarification when:

  1. authority conflict cannot be resolved via hierarchy rule,
  2. gate definitions differ across Tier 1 docs,
  3. requested scope includes implementation execution despite doc-only mode,
  4. non-goals are missing and scope is unbounded,
  5. acceptance evidence is absent for milestone or gate definitions.

Before finalizing any plan doc:

  • Does this section create a backdoor for legacy semantic ownership?
  • Does this section depend on silent fallback behavior?
  • Does this section defer work without owner/expiry/closure criteria?
  • Does this section use ambiguous terms that conflict with glossary?
  • Does this section imply rollout behavior without rollback evidence requirements?

If any answer is yes, revise before acceptance.

When writing concise planning outputs:

  • Keep section hierarchy shallow.
  • Use one line per mandatory constraint.
  • Use explicit “do/don’t” formulations.
  • Prefer deterministic checklists over narrative prose.

Every fast-plan output must link to:

  • 01-master-planning-index.md
  • 05-anti-foot-gun-planning-standard.md
  • 07-task-catalog-authoring-spec.md
  • 08-milestone-gate-definition-spec.md

This fast plan is complete when:

  • a planner can produce or revise the 10-document core corpus in one pass,
  • no implementation execution tasks are included,
  • consistency checks can be run using only this doc plus the Tier 1 docs.