Forge the local circle

Forge the Local Circle

Freeze what the generic pack learned about this product into project-local skills — a fine-tune by forging, not by weights.

Gate

Requires gate: keystone = open and at least one completed cycle of real work in the ledgers (an experiment run, lore tested, or instrumentation verified). Forging locals from an empty circle just bakes in guesses — there must be something learned before it's worth freezing.

The one hard rule

Local adaptations may specialize prompts, embed data, shortcut discovery, and drop irrelevant branches. They may never weaken the refusals: the gate checks, the virality gate, the alienation test, the consent non-delegables, and the evidence-or-hypothesis rule survive verbatim in every derivative. The refusals are the product; everything else is packaging.

Inputs

• GROWTH.md (the learned circle), the generic pack's skills as base templates, and the SkDD skillforge contract (<name>/SKILL.md, YAML frontmatter, forged-* provenance).

Steps

1. Pick what's worth freezing. Candidates, by payoff: the loop (run-loop-<project> — cadence, experiment scaffolding, where flags and queries live), instrumentation checks (verify-<project>-events — exact event names, first-light procedure), the AEO probe suite (probe-<project>-channel — the intents, the assistants, the log format), context playbooks (grow-<project>-in-<context>). Skip what ran once; freeze what ran three times the same way.

2. Forge from the base, embedding the local truth. Start from the generic skill's structure; replace discovery steps with the discovered facts (keystone spelled out, queries inline or pathed, stack named). Keep each forged skill ≤200 lines and self-contained — a future agent with only this file and the repo should be able to act.

3. Carry provenance. Frontmatter on every local: forged-from: <generic-skill> @ 2026-07-growth-circle + GROWTH.md @ <commit> and a one-line forged-reason. The lineage is what makes later upstreaming possible.

4. Wire and register. Locals live in the project's colony (skills/), registered per SkDD convention; list them in GROWTH.md ## Team as operators, with their lanes and the human veto unchanged. If wind-the-clock runs, point the scheduled jobs at the forged locals — that's the intended pairing: the clock fires, the local skill knows exactly what to do.

5. Evolve in both directions. When a local hits an edge case: project-specific fixes stay local; generalizable lessons get translated back to the generic skill and pushed upstream to the commons (skdd push) with the edge case attached. When the keystone sunsets or a re-derivation lands, re-forge the affected locals — stale locals are lore in skill form.

Conventions

• Locals are named for the project and the job, not the ring — run-loop-acme, not acme-ring-3.
• A forged local that only restates the generic skill is noise; if nothing project-specific got embedded, don't forge it.
• Cheaper-model handoff is a first-class goal: write locals so a lighter agent can follow them mechanically — the judgment was spent at forge time.

Edge Cases

• The project's circle changes under the locals (new keystone, new stack). The re-forge in step 5 is mandatory, not hygienic — run it before the next scheduled job fires with stale facts.
• Temptation to fork the refusals ("our team is fine with viral mechanics early"). Refuse, citing the hard rule; if the team overrules, the derivative must drop the Growth Circle lineage from its provenance — it is no longer an adaptation of this pack.
• Secrets in the local truth (query credentials, internal URLs). Locals embed paths and names, never credentials; skills travel, and the colony is not a vault.
• No SkDD in the environment. The forge still works — a local skill is just a directory with a SKILL.md; register it in whatever the harness uses for skills and note the missing provenance tooling.