build-simple

---

name: build-simple description: Build a wiki end to end with the simplified conventional-RAG pipeline (the simple path, not the primary researcher-style wikify builder): extractor-driven concept suggestions, optional author/person seeding, evidence gathering, writing, validation, commit, render, eval, inspection, and iteration without hand-curated concept lists. allowed-tools: Bash(wikify *) Task mcp__wikify__context_set mcp__wikify__context_show mcp__wikify__corpus_find mcp__wikify__corpus_show mcp__wikify__corpus_sample mcp__wikify__corpus_schema mcp__wikify__corpus_traverse mcp__wikify__corpus_citation_walk mcp__wikify__corpus_similarity_walk mcp__wikify__wiki_find mcp__wikify__wiki_show mcp__wikify__wiki_traverse mcp__wikify__wiki_schema

build-simple

Simplified conventional-RAG build path. Owns sample count, concept extraction, evidence quotas, writer tier, concurrency, retry policy, render/eval closure, and stop conditions. Core capability skills explain mechanics.

Workflow (5 phases)

P1 — Bootstrap

wikify run init --bundle <bundle> --corpus <corpus> --strategy baseline --target-haiku-eq 25000000

Capture corpus_fingerprint from stdout — pin it as the manifest hash in run notes. Then bind MCP and sample:

mcp__wikify__context_set(corpus_path="<corpus>", bundle_path="<bundle>")
mcp__wikify__context_show()
mcp__wikify__corpus_sample(strategy="diverse", max_docs=16, pagerank_weight=0.7)

context_show().health carries doc/chunk counts, detected field, AND rank_metrics.{source,author} — use the author list when seeding persons (P2). No separate corpus_schema round trip needed.

P2 — Concept extraction + optional persons

Map-reduce over the sampled docs. Full contract in ../reference/references/exploration/concept-extraction.md.

• Map: one haiku Task per sampled doc, in parallel. Each Task fetches its own doc body via MCP and emits 0–8 candidates from its doc only.
• Reduce: one sonnet Task (or orchestrator) dedupes by canonical title, scores by cross-doc frequency, trims to target.

Append staged concepts and consolidate:

wikify work add feedback concept --record <bundle>/scratch/concepts_staging.jsonl
wikify work tend

Staging path MUST be outside work/inbox/ (pointing --record at the inbox file re-appends it to itself).

Optionally seed person pages. Pick a populated author rank metric from context_show().health.rank_metrics.author, then:

mcp__wikify__corpus_find(by="author", rank="<metric>", top_k=10)

Add 3–5 valid authors:

wikify work add concept "<Display Name>" --kind person --aliases '["author:<key>"]'

P3 — Evidence (vetter per cluster)

wikify work cluster-concepts --by auto --run <bundle> --format json

--by auto picks seeds pre-evidence and evidence post-evidence; the response's mode_selected reports which it chose.

For every cluster (any size), spawn one gather-evidence Task with cluster_slugs=<all slugs in cluster>. The supervisor plans one shared query set, fans out haiku judges, and commits one ledger per slug. Returns a per-slug envelope dict. Singletons use the same contract — no fan-out savings but the judge discipline guard still applies.

Run cluster Tasks in parallel; each holds its own MCP session.

Rescue wave (after each cluster Task). For any slug with stop_reason == "pool_exhausted" AND appended < 10 in the supervisor envelope, spawn another single-slug gather-evidence Task as a top-up. The new supervisor sees the existing evidence.jsonl and sizes its remaining quota; build-evidence dedups by chunk_id so duplicates cannot land. Run rescues in parallel.

Targets per slug (after rescue): ≥10 records across ≥5 distinct docs, at least one definition chunk. Persons need quoted research contributions; never invent biography. If a Task returns stop_reason="error" or appended < 6, mark the slug failed — at most one retry.

P4 — Write + commit

Re-cluster (chunks now drive grouping) via wikify work cluster-concepts --by auto. Per slug: wikify work claim then wikify draft build with --task create --tier M --with-adjacent (loads ord-1/ord+1 flanking chunks for synthesis). Spawn one write-page Task per cluster.

Per page, run wikify draft finalize <slug> --owner baseline — the single-shot normalize → check → commit → release chain, short- circuits on first failure. If check fails, inspect validation.json and re-invoke the writer. Do not patch committed markdown.

P5 — Finalize

wikify work tend
wikify wiki check
wikify wiki rebuild --run <bundle>
wikify wiki navigation-context --run <bundle> --out <bundle>/derived/navigation_context.json
# Invoke organize-wiki to write and apply derived/navigation.json.
wikify render --bundle <bundle> --format html
wikify run close --status completed --run <bundle>
wikify eval --bundle <bundle> --corpus <corpus>

wiki rebuild runs vectors → indexes → graph in one process; use --skip <kind> (repeatable) when only some projections changed.

Then run the Inspection Loop and write the Final Report.

Subagent contracts

role	tier	skill	inputs	return
extractor-map	haiku	this skill (P2)	doc_handle, corpus, bundle	≤8 candidates JSON (≤400 tok)
extractor-reducer	sonnet	this skill (P2)	all map arrays	staging JSONL path
supervisor	sonnet	gather-evidence	cluster_slugs, run, corpus, quota_per_slug=12, max_query_rounds=3	per-slug envelope dict (≤600 tok)
chunk-judge	haiku	gather-evidence (judge role)	sibling slugs + batch of ≤8 chunks with text	per-chunk routing+score+quote JSON
writer	sonnet M	write-page	cluster slugs + dossier paths	per-slug response.json paths

Telemetry

Token counts come from the harness <usage> payload at each Agent tool boundary, NOT from subagent self-reports (subagents undershoot their tool-result intake by 5-10x). Orchestrator records via:

wikify run record-calls --from-stdin --run <bundle> --format json <<'EOF'
{"role":"vetter","model_id":"...","tier":"M","tokens_in":12000,"tokens_out":300,"stage":"evidence"}
... one line per Task ...
EOF

Stages: extract, evidence, write. Recording is mandatory — wikify run close warns on stderr if no call events exist, and cost curves in wikify eval are invalid without them.

Defaults

• Sampled documents: 12–16; corpus sample --strategy diverse, pagerank_weight=0.7.
• Target articles: 10–20. Target people: 3–5 when author metadata exists.
• Vetter quota: 12 records, max 3 gap-driven query rounds.
• Retrieval: MCP corpus_find(rank="all", include_text=True). The include_text flag inlines chunk bodies in the search response — vetters skip the per-candidate corpus_show follow-up. Bash wikify corpus find is the fallback only when the MCP server is unbound.
• Writer tier: M. Extractor tier: S.
• Concurrent writers: default 2 (rate-limit safe on Sonnet); raise to 4 only after confirming no rate-limit warnings.
• Claim owner: baseline. Claim TTL: 1800 seconds.
• Retry: one evidence/writer repair, then mark failed.
• Budget: ~22M haiku-eq on a 12–16-doc build (dominated by ~28 sonnet vetters). --target-haiku-eq 25000000 tracks headroom; informational, not enforced.

Hard Rules

• Do not bypass concept_suggestions.jsonl for the initial concept set.
• Do not call wikify work build-evidence <slug> directly (without --from-ids). That is the pre-vetter seed-first-N-by-ord path and ships byline / references-list / off-topic noise. Evidence gathering goes through the vetter subagent.
• Do not repair committed pages with ad hoc scripts. Fix the skill, evidence, draft, or writer output and re-run the gate.
• Do not call cost curves valid while type="call" events are absent.

Common mistakes

• Choosing the article list by reading sampled titles only (bypasses the extractor).
• Using graph-RAG, summaries, rerankers, or query refinement (off-protocol for this simple-RAG path).
• Falling back to bash wikify corpus find while MCP is bound — pays ~3.6 s cold-start per query.

Inspection Loop

After render, inspect ≥5 pages:

page kind	what to check
strong central article	lead quality, coverage breadth, citation density
weak / thin article	how the writer handled sparse evidence
person page (if any)	quoted contributions, no invented biography
index / sidebar / search	navigation correctness, link integrity
page with equations / chemistry	math and chemical notation rendering

Iterate (add evidence, refine, add missing concepts; re-render and re-eval) when weak.

Stop Conditions

Success criteria pass; budget exhausted; or a deterministic blocker is documented.

Final Report (checklist)

• Bundle + corpus path + corpus_fingerprint
• Sampled doc count, extractor output count, promoted concept count
• Committed / failed article and person pages
• Rendered index.html path, eval report path
• Per-page evidence (active records, distinct source docs, used references)
• Eval metrics M1, M3, M5, M6; GT-P / GT-C availability
• Call-cost telemetry status (must be non-empty)
• Qualitative site judgment; deterministic blockers + next fix

References

• ../search-corpus/SKILL.md
• ../bundle/SKILL.md
• ../gather-evidence/SKILL.md
• ../write-page/SKILL.md
• ../organize-wiki/SKILL.md
• ../reference/references/exploration/concept-extraction.md
• ../reference/references/exploration/sampling-patterns.md
• ../reference/references/writing/escalation.md