Forging

Forging

Idea → airtight decision artifact → plan. This skill routes the work through superpowers:brainstorming and grilling (invoke them via the Skill tool — do not paraphrase them from memory) and decides whether an existing artifact is reused or rebuilt.

REQUIRED SUB-SKILLS: superpowers:brainstorming, grilling, superpowers:writing-plans.

Fase 0 — Detect state (before any question to the user)

Two observable predicates:

1. Is the approach OPEN or SETTLED? SETTLED if a codebase convention covers it (N shipped modules with the same pattern), the user names the pattern, or the request is pure decision-closing ("cero decisiones abiertas", "poke holes"). OPEN if multiple viable architectures exist or the user signals vagueness ("no sé, algo tipo…").
2. Does a prior artifact exist? (spec / decision-log, typically docs/superpowers/specs/). If yes, run the Quality Gate BEFORE asking the user anything.

Routing

State	Route
Approach OPEN, no artifact	brainstorming (ONLY until approach chosen) → grilling → writing-plans
Approach SETTLED, no artifact	grilling → writing-plans. Skip brainstorming: proposing 2-3 approaches for a pattern the user just named is theater
Artifact exists, gate PASS	writing-plans directly; re-open ONLY what the gate flags
Artifact exists, gate BURNED	Rebuild Protocol below
Multiple independent subsystems	brainstorming's decomposition first; each sub-project re-enters Fase 0

Harmonization when both run: brainstorming stops at purpose, constraints, success criteria, chosen approach — every implementation decision belongs to grilling; do not write the design doc between them. One artifact: a spec with the grilling decision-log embedded (Elegido / Por qué / ⚠ supuesto per entry) — never two parallel documents. Question format throughout is grilling's: options, recommended first, tradeoff per option.

User pushes to skip ("haz el plan de una vez"): rapid sweep, defaults pre-chosen; whatever the user waves off takes the recommended option marked ⚠ supuesto in the log. Cede visibly, never silently.

Quality Gate (existing artifact)

Audit against ALL six: (1) placeholders — TBD, "decidir luego", empty sections; (2) internal contradictions; (3) ambiguity (a requirement readable two ways); (4) dimension coverage per grilling's sweep list, or explicit N/A; (5) unresolved decisions; (6) staleness — commits since the artifact touching what its decisions assume.

Verdicts:

• PASS — reuse. Never re-ask what the artifact answers. Enter the pipeline after it.
• LOCAL DEFECT — cosmetic gaps only, AND every load-bearing decision (keys, data model, contracts, authz) intact and mutually consistent → fix inline, note it in the log.
• BURNED — any of: a contradiction in a load-bearing decision, an entire dimension absent, or multiple unresolved decisions → Rebuild Protocol. A load-bearing contradiction is never a "local defect": you cannot know which other sections were written under which side of it, so the whole document loses authority as a source of truth.

Rebuild Protocol (burned artifact)

The old artifact is INPUT, never TEMPLATE.

1. Extract to a raw list ONLY the decisions the user explicitly confirmed and the hard constraints. Nothing else survives.
2. Demote to "re-validate" every extracted decision that was answered under either side of the contradiction.
3. Mark the old file SUPERSEDED BY <new> in its header. Do not delete it. Set it aside — do not re-open it while rebuilding.
4. Re-enter Fase 0 with the extracted list as pre-answered inputs: confirmed answers are not re-asked; everything else is asked fresh through the normal route.
5. Produce the new artifact from zero. Do not reuse the old document's structure, section order, or prose.

Patching a burned artifact inline is prohibited. No exceptions:

• Not because "the defects are localized and enumerable" — the visible defects say nothing about the invisible ones a structural contradiction implies.
• Not because "it respects the hours the user invested" — those hours live in the extracted decisions (step 1), which survive intact; the prose is not the work.
• Not because "token-aware-authoring says Edit over Write" — that rule optimizes healthy documents. A burned artifact is not a revision target; extraction + rebuild IS the token-cheap path. Patching produces a poisoned doc v2 plus a second audit.

Rationalization table

Excuse	Reality
"Defectos localizados y enumerables, no estructurales"	A contradiction in the model's key is structural by definition: it poisons every section that assumed it.
"Reescribir destruye las 3 horas del usuario"	The hours live in the extracted decisions, not the prose. Step 1 preserves them whole.
"Edit > Write, token-aware"	Applies to healthy documents. A burned doc patched is a burned doc v2.
"Brainstorming es obligatorio (regla 1%)"	With the approach settled by convention, the user's specific trigger beats the generic mandate. Grilling IS the design closure.
"El artefacto estaba completo cuando se escribió"	Completeness does not survive commits. Staleness check before reuse.
"El usuario dijo 'haz el plan ya'"	Sweep with ⚠ defaults — visible concession, never silent decisions inside the plan.

Red flags — STOP

• About to propose 2-3 approaches for a pattern the user just named → grilling directly.
• About to write a plan containing TODOs or open decisions → grilling first (or ⚠ defaults).
• About to "fill in the TBDs" of a spec with a structural contradiction → Rebuild Protocol.
• About to produce spec AND decision-log as separate documents → one artifact.
• About to re-ask something a PASS artifact already answers → reuse it.

Terminal state

Sweep dry or gate passed → invoke superpowers:writing-plans. Never an implementation skill, no code, no scaffolding (brainstorming's HARD-GATE carries through this skill).