| name | skill-smith |
| description | Create, scaffold, or batch new Claude Code skills to a tested-real bar: research-first design, spec scaffold, eval+budget acceptance gate, then self-evolve iteration. |
skill-smith, a skill that creates skills
Governing principle (full text in the repo's PHILOSOPHY.md): a skill is done when it is
proven, not when it is generated. Research before you design (P1); generation != usable, so
nothing ships without passing the acceptance gate (P2); own the seam and delegate the engines (P3).
When to stop and route elsewhere (decide first)
- Improving / fixing an EXISTING skill -> this is
self-evolve's job. Route there, do not recreate.
- "Is there a ready-made skill for X?" ->
market-intel's ready-skills domain. Route there.
- Creating a NEW skill, or a batch of new skills -> continue here.
Workflow (thin, load the named reference/<shard>.md only for the step you are on)
| # | Step | Load | Delegates to |
|---|
| 0 | Research-first (MANDATORY), survey best references + frontier designs + anti-patterns before any design | reference/research-first.md | market-intel (deep scale) |
| 1 | Triage + dedup, overlap with existing library? single vs batch? | reference/scaffold.md §dedup + scripts/dedup_check.py | , |
| 2 | Choose generation backend | reference/generators.md | Skill_Seekers / official skill-creator |
| 3 | Scaffold Spec-v1 repo (add --with-config if config-bearing) | scripts/scaffold_skill.py + reference/config-spec.md | , |
| 4 | Draft SKILL.md + optimize triggering description | reference/triggering.md | official run_loop.py (60/40) |
| 5 | Acceptance gate (all must pass; any fail = explicit reject) | reference/acceptance-gate.md + budget_check.py + dedup_check.py + check_conformance.py + check_config_conformance.py (G8) | agent-skills-eval / scenario-eval |
| 6 | Hand off to auto-iteration | reference/iterate-handoff.md | self-evolve |
| 7 | Batch a series (if multiple) | reference/batch.md | Workflow + library-budget manager |
| 8 | Deploy (junction + GitHub publish; then MANDATORY set remote metadata + verify) | reference/deploy.md + scripts/set_repo_metadata.py + scripts/check_remote_conformance.py | npx skills / gh |
Config-bearing skills (need per-user keys / installed-tool registry / endpoints) are a
first-class case: decide at Step 1, scaffold with --with-config (Step 3), and they MUST pass the
config standard (E1 to E7) at the gate (Step 5, G8). Authority: reference/config-spec.md.
Hard invariants (never violate, these are P1/P2/P4/P5 in force)
- Research before design. Never scaffold a skill without the Phase-0 market-intel recon. If
market-intel is unavailable, fall back to
deep-research and SAY SO, never skip silently.
- No skill is "accepted" without passing the full gate. Eval lift vs baseline + held-out trigger
rate + library token budget + dedup + security audit + spec conformance + single-responsibility.
Failure is surfaced as an explicit gap, never a silent ship.
- The token budget is library-wide. A new/batch skill that pushes total descriptions past
~15k chars / ~4k tokens must be merged/pruned first, adding it would silently truncate the set.
- One skill, one job (<=3 modules). Sprawl is rejected and split.
- Conform or it is not a DaizeDong skill, locally AND on the remote. Output must pass
check_conformance.py (G6: local files, 7 files, philosophy-first bilingual README, badge block,
version four-source-synced, plugin fingerprint) and, once published, check_remote_conformance.py
(G6b: the live GitHub repo carries the base-9 + domain topics and a description). These are two
layers: a plain git push sets no remote metadata, so passing G6 alone is how the topics=null
incident happened. Deploy (Step 8) MUST run set_repo_metadata.py then G6b, neither layer is
optional.
- Config-bearing = configurable-by-anyone, or rejected. If a skill needs a companion config
(keys / registry / endpoints), it MUST pass the seven-element config standard (E1 to E7) via
check_config_conformance.py (G8): documented schema, env-var discovery mount, deterministic
init, a verify doctor, two configs hot-swappable by env var, secrets gitignored (Mode B), and
a README Config section. "Works on my machine" config is a reject. See reference/config-spec.md.
Progressive loading
This SKILL.md is the only always-loaded file. Read reference/<shard>.md on demand, one at a time,
for the step you are executing. Never read the whole reference/ directory at once.
