vamseeachanta OSS-Wiki Development Arc — Three-Phase Methodology
Empirical basis: Traced across 38 iterations of llm-wiki (iter-22 → iter-59, V1 → V17 audit lineage). Phases were observed, not designed; this skill codifies them so future wiki projects can pre-load the pattern.
The 3-Phase Pattern
| Phase | Audit-version range | Iters observed | Marginal-cost signature | Exit criterion |
|---|---|---|---|---|
| 1. Substrate | V1 → V11 | ~27 | Cheap breadth wins | substrate-fill marginal-cost > marginal-value |
| 2. Depth | V12 → V15 | ~7 | Compounding per-page value via 4-element recipe | median per-iter line-growth <50% + thin-starter cohort exhausted |
| 3. Quality | V15 → V17 | ~2 | Diagnose-then-execute closure | 4-criterion publication-readiness gate met |
Each phase has a distinct iter-shape, distinct success metric, and distinct saturation signal. Mixing phases (e.g., depth-expanding while substrate is incomplete) wastes effort because depth-content depends on substrate-resolvers being citable.
When To Use This Skill
- Bootstrapping a new engineering/marine/scientific OSS wiki (greenfield)
- Resuming work on a stalled wiki (which phase are we in?)
- Diagnosing wiki-development thrash ("we keep adding pages but nothing improves")
- Auditing a wiki for publication readiness
When NOT To Use
- Closed/internal documentation projects (audit-trail, governance constraints differ)
- Single-page reference docs (no substrate/depth/quality distinction)
- Wikis whose authoritative content is mirrored from elsewhere (use sync tooling, not this arc)
Phase Invocation
Read the phase-specific reference before kicking off iter-N+1. Each reference is self-contained — a future agent can apply it without reading the others.
- Phase 1 — Substrate:
references/phase-1-substrate.md- 4-agent fanout (audit + 2-3 content lanes + bridge)
- Goal: every concept page links to a resolver page; every entity has a stub; substrate-completeness measured per-wiki
- Phase 2 — Depth:
references/phase-2-depth.md- Top-N inbound-citation candidate selection
- 4-element recipe per expanded page (worked example + comparison table + ≥30 outbound intra-wiki links + cross-link to entity/source)
- 3-cohort growth-profile (thin-starter +200-300% / mid-tier +100-130% / already-expanded +35-50%)
- Phase 3 — Quality:
references/phase-3-quality.md- Diagnose-then-execute 2-iter cycle
- 4-criterion gate: orphans ≤3 + unidir-bridges=0 + frontmatter consistency + link-integrity ≥99%
Supporting references applicable to all phases:
references/audit-template.md— audit-doc structure + V1→V17 measurement-discipline progressionreferences/iter-shape-recipes.md— 4-agent fanout default; bundled-sequential exception (W231 race-block lesson)references/cross-wiki-bridges.md— sister-pair pattern; sibling-template recognition; substantive (not "see also") bridge content
Saturation-Detection Signals
Before kicking off another iter in the current phase, verify the saturation signal hasn't fired:
| Phase 1 → 2 transition | Phase 2 → 3 transition | Phase 3 → publish |
|---|---|---|
| Substrate-fill audit identifies <3 high-value gaps per wiki | Median per-iter line-growth across top-N pages drops <50% | All 4 closure criteria green simultaneously across 1 audit cycle |
| Cross-wiki edge-density plateaus | Thin-starter cohort exhausted (no <100-line top-citation pages remain) | Frontmatter mismatch count = 0 |
| Resolver-coverage ≥95% on canonical concept set | Already-expanded cohort dominates the next-target list | unidir-bridges = 0 (every cross-wiki link is reciprocated) |
If a saturation signal fires while in-phase, transition rather than spend another substrate-iter or depth-iter; the marginal value has collapsed.
Reference Exemplar
vamseeachanta/llm-wiki — public OSS wiki where this arc was empirically traced. Repository structure, audit-doc lineage (docs/audits/V1.md … V17.md), and iter-shape decisions are all visible.
Critical Cost-Avoidance
A future wiki using this skill should:
- Start with the audit-doc skeleton (see
audit-template.md) on iter-1, not iter-15. Most of llm-wiki's measurement-discipline gain happened V12→V15; loading it pre-baked saves ~10 iters of audit-format churn. - Adopt the 4-agent fanout default from iter-1. Solo-author iters were the dominant low-throughput pattern in llm-wiki's V1-V8.
- Track inbound-citation count per page from iter-1 (substrate-phase metric is just count; depth-phase ranks by it). This avoids retroactive instrumentation.
- Resist the urge to depth-expand before substrate is saturated. Depth-content references resolvers; if resolvers are stubs, the cross-links don't compound.