Agent-Skills.md

Agent Skills: derive-system-from-goal (soul-name: *Hodos*) — the vehicle deriver

|

UncategorizedID: ekson73/multi-agent-os/derive-system-from-goal

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/derive-system-from-goal

Skill Files

Browse the full folder contents for derive-system-from-goal.

Download Skill

Loading file tree…

skills/derive-system-from-goal/SKILL.md

Skill Metadata

Name
derive-system-from-goal
Description
|

derive-system-from-goal (soul-name: Hodos) — the vehicle deriver

The law this executes: akasha ~/.claude/docs/derive-system-from-goal.md ([C22] pointer in CLAUDE.md; an on-demand reference doc, not an auto-loaded rule — KRDR) — dada a meta recorrente, construa o sistema ANTES de persegui-la; meta sem sistema é intenção sem ação. The law's only mechanized fire-point is anti-theater-grounding-protocol Layer-5 R9 (additive · fires ONLY on recurring/open-ended goals); this skill is R9's operational how. Position: ooda-loop ORIENT-b. Boyd's ORIENT is synthesis — Prisma gives the destination's coordinates (dod-as-prompt), Hodos gives the route (system-as-prompt). Both are ORIENT; neither drives.

Purpose

Derive the minimal recurring system (the vehicle) that conducts to a recurring/open-ended goal — BEFORE pursuing it — and emit it as a typed system-as-prompt envelope for ooda-loop ORIENT-b. A one-shot/bounded goal is N/A by design (its plan IS its system — law doc §2): this skill never manufactures a "recurring cadence" for bounded work.

When to use

  • A recurring/open-ended goal is given (or self-imposed) and the next move would otherwise be "start doing things toward it". R9 fired (recurring goal + one-shot-only vehicle).

When not to use

  • (a) One-shot/bounded goal — N/A: a plan IS its system (law doc §2); exit without an envelope. (b) Trivial action — its system IS the single step; deriving one is ceremony (Gordian floor). (c) The operator's personal goals — [C17] §2 HUMAN_DOMAIN: available on request, never auto-prescribed. (d) Driving — that is gap-loop/quiesce. (e) Measuring done-ness — that is Prisma.

Trigger Phrases

  • "derive a system for this goal" · "qual é o menor sistema que me leva a esta meta?" · "build the vehicle before driving" · "system-as-prompt" · an ooda-loop ORIENT-b invocation · an R9 REFINE verdict from anti-theater-grounding-protocol.

Parameters

| Flag | Default | Allowed / Notes | |---|---|---| | "<goal>" (positional) | — | the goal to build a vehicle for. Omit when --from-handoff supplies it. | | --from-handoff | (none) | path to a validated handoff-as-prompt — sources for_goal + context_lock typed (no re-inference). | | --from-dod | (none) | path to a validated dod-as-prompt — the destination's coordinates; the route is derived toward these leaves, not a vibe. | | --output | text | text | json (emit the system-as-prompt envelope). Convention mirrors the family — --output, not --format/--kind. | | --dry-run | off | derive + print, do not write the envelope file. |

The algorithm (3 rules + the guard — each a gate, not a vibe)

P0  CONTEXT-LOCK   inherit from --from-handoff (context/purpose/stakeholder/targets).
                   | absent -> SpecError. "Minimal" is meaningless without whose day it fits in.
                   v
P0.5 GOAL-SHAPE    recurring/open-ended goal? -> proceed.
                   | ONE-SHOT/BOUNDED -> N/A: exit WITHOUT an envelope (a plan IS its system;
                   |   R9 = N/A per the law doc §2). Demanding a "recurring cadence" from a
                   |   bounded goal is the missing-middle the red-teams killed -> never do it.
                   v
P1  MINIMAL        propose the single recurring ACTION so small refusing it is practically impossible.
                   | GATE: state `why_minimal` — the concrete cost + why it cannot reasonably be refused
                   |   in THIS context_lock. "it is small" unfalsifiable -> theater (anti-theater R2) -> re-derive.
                   | GATE: if the action enumerates sub-steps it is NOT minimal -> decompose further.
                   | {ts:511} "tao facil que seja praticamente impossivel voce dizer nao"  (Fogg / Lally 2010)
                   v
P2  CONSISTENCY    bind it to a concrete recurring TRIGGER (if-then) + a CADENCE relative to that trigger.
                   | GATE: "when I feel like it" is not a trigger. Deterministic cue where one exists.
                   | {ts:518} "melhor 10 min de caminhada por dia do que 3 horas"  (Gollwitzer & Sheeran d=0.65)
                   v
P3  SIGNAL         derive the Metron-ADMISSIBLE adherence signal (NOT a run-count — that is blacklisted):
                   |   "does the mechanism still FIRE — AND does firing still MOVE the goal?"
                   | decision-gated (no-fire => rebuild) + outcome-anchored (the system's purpose IS the goal)
                   v
REV REVISION       derive the falsifier: adherence-high + goal-movement-absent after N => the VEHICLE is wrong.
                   | this is the guard the source video LACKS and the answer to Eliason (see Caveats).
                   v
    EMIT           system-as-prompt envelope -> validate -> ooda-loop DECIDE gate (which routes; never authors)

The homework question, verbatim ({ts:604-617}) — it is the algorithm's intent: "Pensa na meta que você mais queria alcançar… Agora pega e esquece essa meta e pergunte para você mesmo: qual é o menor sistema que vai te levar até essa meta?"

The typed triple (the wiring — no invented interface)

| Envelope | Produced by | Consumed as | |---|---|---| | handoff-as-prompt | goal-recovery (OBSERVE) | driver --state-source=handoff:<file> | | dod-as-prompt | Prisma (ORIENT-a) | driver --condition=<termination_predicate> | | system-as-prompt | THIS (ORIENT-b) | minimal_system.action → driver positional "<instructions>" — "extra free-text appended to the goal" (gap-loop SKILL.md:194). Only the .action string is passed; minimal_system itself is an object (trigger/action/why_minimal/cadence) whose other fields govern the derivation, not the driver call. ⚠️ No flag for "the recurring action" exists and inventing one would be interface fabrication (anti-theater R4). The recurring step rides the existing surface. |

Validator-gated like its siblings (skills/goal-recovery/bin/validate_envelope.py). The triple is revisable, not frozen — REV exists precisely to revise it.

Composition (DRY — reimplements nothing)

| Concern | Owned by (cited, not rebuilt) | |---|---| | recover the goal | skills/goal-recovery | | measure the destination | skills/decompose-abstract-to-measurable (Prisma) | | derive the route | THIS | | gate / route | skills/ooda-loop DECIDE (COWORK-AUTONOMY-POLICY bands + [C17] §2 + anti-theater 9Q) | | drive | skills/gap-loop / skills/quiesce | | verify (verifier != generator) | the driver's own audit — maos:persona-pipeline / perspective-trio. NEVER re-implemented here (Huang et al. 2310.01798). | | measure over time | agentic-observability-protocol (Metron) |

The anti-duplication test this skill must keep passing: it contains no loop, no convergence gate, no verifier. If any appears here, it has become gap-loop/ooda-loop re-implemented → cut it (Gordian).

Caveats (honest — deleting these makes this marketing)

| ⚠️ | The honest statement | |---|---| | The source video is n=1 | An anecdote (a youtuber citing a youtuber) — the weakest evidence behind this skill. Adoption is selective by the operator's own instruction ("leitura é obrigatória, a adoção é seletiva"): the three rules map onto real research; the raw thesis "systems > goals" is NOT adopted. | | What IS evidence-backed | Gollwitzer & Sheeran 2006 implementation intentions (d=0.65, 94 tests, 8,000+ participants; 2024: 642 tests) → P2 · Fogg Tiny Habits / Lally 2010 (~66d median) → P1 · Michie BCT self-monitoring → P3. | | The Eliason counter | "Systems without goals is a path to mediocrity." Answered by the operator's own formulation — goals stay valid, as direction. We adopt the complement, never the substitution. REV is the guard the video lacks. | | The Metron tension — narrow, real | "Track the system, not the goal's number" does not contradict Metron — it is its §1 verbatim ("Measure-to-Inform, Manage-to-Outcome"). BUT §5 gate-2 demands outcome-anchored and the vanity blacklist bans activity metrics ⇒ "count how many times we ran it" is blacklisted. Surviving formulation: system = managed object; goal's number = informing signal; admissible signal = fires + still moves. | | No cadence substrate exists | cadence is a predicate string consumed per-round — not a scheduled job, not an adherence ledger. The only launchd here is machine-health; bin/dogfood-mark/tally is a maturity cycle-counter (collapses time; no should denominator; subject is a tool). Its format is reusable if longitudinal tracking is ever built. It is not one today. |

Anti-patterns (do NOT)

  1. ❌ Ship a plan as a "system" — a multi-step plan is not minimal (P1 gate). A system is one recurring step.
  2. ❌ why_minimal as unfalsifiable prose ("it's small") — state the cost + why it can't be refused, else it is R2 theater.
  3. ❌ A wish as a trigger ("when I have time") — P2 needs a concrete recurring cue.
  4. ❌ Activity-count as the signal — Metron vanity blacklist. Use fires + moves.
  5. ❌ Blame the driver when adherence is high and nothing moved — REV falsifies the vehicle.
  6. ❌ Grow a loop/gate/verifier in here — that is ooda-loop/gap-loop re-implemented → cut (Gordian).
  7. ❌ Invent a driver flag for the recurring action — none exists; use the positional (anti-theater R4).
  8. ❌ Auto-derive a system for the operator's personal goals — HUMAN_DOMAIN; on request only.
  9. ❌ Ceremony on a trivial action — its system IS the single step (Gordian floor).

Protocol Rules (integrity invariants)

  1. One-shot → N/A (P0.5): never demand a recurring cadence from a bounded goal; exit without an envelope — the plan IS its system.
  2. Thin deriver: no execution loop, no convergence gate, no verifier in here — if one appears, it is gap-loop/ooda-loop re-implemented → cut (Gordian).
  3. HUMAN_DOMAIN carve-out: never auto-derive for the operator's personal goals ([C17] §2) — on request only.
  4. Metron-admissible signal only: adherence = "fires + still moves the goal"; bare activity counts are blacklisted.
  5. Validator-gated emission: every system-as-prompt passes skills/goal-recovery/bin/validate_envelope.py before the DECIDE gate consumes it.
  6. Additive to the 8Q core: an R9 REFINE never overrides an upstream R2/R4/R6/R8 REJECT — this skill builds vehicles, it does not launder theater.

Quality Tests (6/6 self-validity)

  1. Self-Application — this skill's own goal ("close the vehicle gap") has a system: R9 fires → derive → envelope → ORIENT-b → driver. ✅
  2. Non-Contradiction — composes goal-recovery/Prisma/ooda-loop/gap-loop/Metron without duplicating; the Metron tension is resolved explicitly (§Caveats), not glossed. ✅
  3. Survival — applied to itself it demands a minimal recurring vehicle; it is one step in an existing loop, not a new loop. ✅
  4. Bounded-Responsibility — derives only (no loop/gate/verifier) · Gordian floor · HUMAN_DOMAIN carve-out · law's §8 DUED. ✅
  5. Explicit-Exception — the 4 not-use cases + [C17] §0 SER. ✅
  6. Utility-Sunset — inherits the law's DUED: retire when ooda-loop ORIENT absorbs system-construction natively (E6), or agents internalize build-the-vehicle-first (E3, ≥10 zero-regret). ✅

Refs

  • Law (SSOT): akasha ~/.claude/docs/derive-system-from-goal.md ([C22] pointer in CLAUDE.md) · fire-point anti-theater-grounding-protocol Layer-5 R9 (additive · recurring-only).
  • Siblings: skills/ooda-loop (the conductor — ORIENT-b is here) · skills/goal-recovery (+ its bin/validate_envelope.py) · skills/decompose-abstract-to-measurable (Prisma — ORIENT-a) · skills/gap-loop · skills/quiesce.
  • Sources: transcript ~/Downloads/transcricao_pessoas_inteligentes_sistemas.md (Lucas Yano 2026-07-05, read in full — {ts:411,451,511,518,554,604-617}) · ~/Downloads/framework_12_etapas.csv.
  • Evidence: Gollwitzer & Sheeran (2006) · Fogg Tiny Habits · Lally et al. (2010) · Michie BCT taxonomy · Nat Eliason (the counter) · Huang et al. 2310.01798 (why the verifier stays independent, and elsewhere).
  • Etymology: μέθοδος = μετά + ὁδός; ὁδός + μέτρον = ὁδόμετρον (odometer) — Hodos the road, Metron the measure.
  • Naming: [[anima]] per [C-naming] — 12/12 + 4/4, conf 0.97. Registered: bin/artifact-registry (name + create). Rejected: minimum-viable-system (taken — SYSTEMology) · poros (pt-BR fluency).
  • Cross-link: [[derive-system-from-goal]] · soul-name Hodos (ὁδός).