Skip to content
Agent Orchestration
Skill

/spectre-plan-route

Internal semantic classifier for Plan-or-Execute preparation. Use only from Plan or Execute with bounded routing input; never for planning, artifacts, telemetry, or user gates.

BOOST
From plugin
spectre
16521 skills8 agents
Install
$ npx -y skills add joenandez/spectre --skill spectre-plan-route --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/spectre-plan-route

Context preview

The summary Claude sees to decide when to auto-load this skill.

Internal semantic classifier for Plan-or-Execute preparation. Use only from Plan or Execute with bounded routing input; never for planning, artifacts, telemetry, or user gates.

SKILL.md

spectre-plan-route.SKILL.md
name: "spectre-plan-route"
description: "Internal semantic classifier for Plan-or-Execute preparation. Use only from Plan or Execute with bounded routing input; never for planning, artifacts, telemetry, or user gates."
user-invocable: false

plan-route

Purpose

Classify into plan-routing/v1 for Plan/Execute; caller owns authority.

Inputs

  • Canonical Scope/authority + bounded work, not size.
  • Initial; observed uses prior decision + selection before drafting; legacy may include Routing Observations.
  • Execute when transported assessment is missing/materially stale.
  • Resume-only legacy size is normalized once.

Working Set

  • Bounded scan: topology, uncertainty, evidence, boundaries, graph risk, shipped precedent, user decisions.
  • If missing evidence alone would cross a costlier route, use exactly one bounded probe (`@spectre_finder`/`@spectre_patterns`) for one repository question.

Outputs + DONE

Produce `plan-routing/v1`: shape ATOMIC|DIRECT|STRUCTURED; uncertainty LOW|MODERATE|HIGH; evidence SUFFICIENT|PROBE_REQUIRED|PROBED; protected_boundaries[] {type, threatened_invariant, failure_mode}; task_graph_risk LOW|HIGH; design flag, size, route, rationale. Observed adds regret, reason codes, KEEP|RERUN_SMALLER|RERUN_LARGER.

DONE when fields validate; probe is consumed; STRUCTURED names workstreams; HIGH graph risk names an implementation-graph failure; only table derives size/route.

Method / guardrails

| Semantic result | Size · route | |---|---| | ATOMIC + LOW + no changed protected boundary | XS · XS_DIRECT | | ATOMIC + LOW + changed protected boundary OR DIRECT + LOW | S · S_DIRECT | | ATOMIC/DIRECT + MODERATE/HIGH | M · M_REVIEWED_DIRECT | | STRUCTURED + LOW/MODERATE + ordinary graph risk | L · L_STRUCTURED | | STRUCTURED + HIGH uncertainty or HIGH task-graph risk | XL · XL_REVIEWED_STRUCTURED |

  • STRUCTURED requires multiple independently implementable workstreams; render variants of one surface (layout, density, breakpoint, theme), dependencies, supporting artifacts, workflow/acceptance steps, and pilots are not workstreams.
  • Two shipped instances of a change-shape make the next DIRECT; its layers are not workstreams; STRUCTURED requires a named delta beyond repetition.
  • HIGH graph risk requires a credible implementation ordering/coordination/rollback failure; workflow gates/state transitions do not qualify.
  • A protected boundary needs a concrete invariant + credible failure mode; it floors size at S but never creates structure or HIGH graph risk.
  • Honor confirmed Scope assumptions and explicit selected-source boundaries. Missing paths/evidence permit at most the one probe; when Scope mandates an abstraction, spend it on whether it ships; classify the real delta. Only unresolved, approach-changing uncertainty affects size.
  • Design authority needs unresolved product, compatibility, destructive, migration/rollback, or architecture choice; size and routine placement never create it.
  • Observed consumes the completed minimum-solution selection before drafting, reports regret and KEEP|RERUN_SMALLER|RERUN_LARGER, and never repeats work or removes artifacts. The same routing table alone maps selected structural facts, uncertainty, protected boundaries, graph risk, and assurance floor; it does not select solution shape.
  • Normalize legacy once: MICRO→XS, LIGHT→S, STANDARD-DIRECT→M, STANDARD→L, COMPREHENSIVE→XL; labels never decide.
  • Never use file volume, dependency count, surface counts, or sensitive-domain keywords as size authority, and never let them create workstreams. Never plan, write artifacts, emit telemetry, or present gates while classifying; these limits end with the record and never bind the caller's next step.

Handoff

Inline (Plan): keep record/probe as internal data; never print it as a reply. DONE does not end caller's turn: in the same response, run the caller's next step. Child (Execute): final message is the record. Caller owns persistence, explanation, orchestration, gates, telemetry.

Escalate-If

  • Canonical Scope absent/conflicting, invalid enum, or boundary lacks predicate.
  • Material user-authority decision: report `design_authority_required`, never decide.
Read more
Ships withspectre

SPECTRE is an Agentic Engineering workflow for Claude Code and Codex designed to help you do ONE THING more, faster, and with higher quality.

Get the whole plugin

Other skills on spectre.