Build Protocol

Rigorous workflow for producing long-form knowledge content (textbooks, knowledge bases, deep reports) with multi-agent collaboration. Use when the task requires >5,000 words, >2 parallel sub-agents, or >10 steps, or when the deliverable will be referenced long-term by the user. Enforces 8 iron rules (independent Audit unmissable / Plan before Execute / ≤2 parallel sub-agents for writing tasks / Why This Way reasoning / version + errata iteration / 3-layer consistency / anti-sycophancy content / independent review ≠ self-audit) and a 9-step standard workflow. Includes machine-verification scripts for PMID citations, red/green evaluation ratios, safety-keyword coverage, and H1-level checks. Triggers on explicit requests like 'build knowledge base / textbook / deep report' or implicitly on any long-form knowledge production.

Turn "write a knowledge product" into a rigorous 9-step workflow with machine-verifiable quality gates.

Distilled from 4 battle-tested projects: a multi-chapter AI/ML textbook, a tactical gear knowledge base, a supplements knowledge base, and a product design doc series.

When to Use

Triggers (any one):

• Task will produce > 5,000 words
• Requires > 2 sub-agents
• Requires > 10 steps
• User will reference the deliverable long-term (textbook / knowledge base / deep report)
• Explicit requests: "build knowledge base", "write textbook", "做知识库", "做教材"

Don't use for:

• One-off Q&A
• Simple documentation
• Engineering/code projects (use a separate build-protocol-engineering skill)
• Short reports (<5k words)

The 8 Iron Rules

#	Rule	Why
1	Independent Audit unmissable	LLM self-review has 30% lower accuracy (sycophancy)
2	Plan before Execute	Drift cost >> planning cost
3	≤2 parallel sub-agents	Conflict points grow as N(N-1)/2; 4+ concurrent = 75% failure
4	Why This Way	Force reasoning over authority ("best practice" = zero-cost lie)
5	Version + Errata iteration	Honest revision > pretending to get it right once
6	3-layer consistency	Naming ↔ Business ↔ Data
7	Anti-Sycophancy content	Must have 🔴 negatives; all-green = fake
8	Independent review ≠ self-audit	Writer and reviewer must be different agents

The 9-Step Workflow

1. Plan      → Outline + sub-agent allocation + time estimate
2. Prepare   → Extract content (≤3000 chars per sub-agent task)
3. Execute   → Dispatch sub-agents (≤2 parallel, jittered start)
4. Assemble  → Merge files, unify format
5. Review    → Lightweight review (format / dirty markers)
6. Audit     → Deep audit (L1 facts / L4 safety / L6 anti-sycophancy) ⚠️ unmissable
7. Fix       → Fix 🔴 > 🟡 > 🟢 in order
8. Publish   → Convert to docx + upload + VERIFY
9. Errata    → Subsequent issues → v1.x Changelog

Each step has a completion checkpoint — see references/workflow-checkpoints.md.

Machine-Verified Quality Gates

Before Step 8 (Publish), run audit_check.sh (see references/audit-script-template.sh) to verify:

• L1.1 Citation count: Each volume ≥ threshold (e.g., ≥5 PMIDs for medical)
• L1.2 Citation duplicates: Zero duplicates (sort | uniq -c | awk '$1>1' returns empty)
• L3 Length: Each volume ≥ 6,000 chars (for "full volume" style)
• L3 Gender balance: Male & female perspectives both present (if applicable)
• L4 Safety coverage: ≥3 safety keywords per volume (for health/medical content)
• L5 H1 count: Exactly 1 per file (no stray H1 from sub-agents' progress markers)
• L6 Red count: Each volume has ≥1 🔴 critical review; ideally 🔴 ≥ 20% of 🟢

🔴 fails block publishing. 🟡 can defer to next version.

Audit Fallback (when Audit sub-agent fails)

Scenario	Fallback
Single-model timeout	Retry with different model / region
3+ models all timeout	Main agent audits manually (use Opus-tier model + grep scripts)
Time-pressed / context >85%	Reduce scope: L1 + L4 only, skip L5/L6
Audit quality poor (<10 real issues)	Dispatch second audit agent for cross-validation

Main-agent audit has advantages: stronger model, machine-scripted checks, no idle-timeout risk.

Sub-agent Specification Template

Every sub-agent task must include:

1. Goal:        What to do (one sentence)
2. Output:      Which file, what format
3. Content:     Key info provided directly (don't make sub-agent hunt through large files)
4. Constraints: What NOT to do (as important as "what to do")
5. Acceptance:  Measurable completion criteria

Single-file content in task instructions: ≤3,000 chars (just-in-time context principle).

Version Numbering

• +0.1 → minor revision (≤3 small fixes per ~5k words)
• +1.0 → major version (≥50k words or ≥3 new chapters or architecture change)

Keep only latest version in the delivery folder; archive old versions to HistoryBackup/.

Each Release Requires a Changelog

## v1.1 · YYYY-MM-DD
- 🔴 Fixed: [blocker issue X]
- 🟡 Added: [improvement Y]
- 🟢 Deferred to v1.2: [minor issue Z]

Anti-Patterns (DO NOT)

❌ Don't	Why
"I already reviewed it"	Implementer = LLM. Must independently verify.
"Time is tight, ship it"	Time tight = reduce scope, never reduce standards
Let writer sub-agent self-review	Sycophancy guarantees overestimation
Skip Audit because sub-agent failed	Use fallback path, don't skip
Version-bump without Changelog	Silent changes destroy trust
"Mostly correct, minor issues"	Details = professionalism; 🟡 issues are tracked
Copy existing review grades (all 🟢)	Real world has 🔴; all-green = fake
4+ parallel sub-agents	75% failure rate; use 2+2+1 batching instead

Gotchas

1. Sub-agent progress markers leak to output — check for # progress: step X/N in deliverables; grep and remove before Audit
2. Same-region concurrent dispatch causes throttling — jitter second sub-agent by 1-3s; use different Bedrock regions if available
3. Fix step can introduce new duplicates — always re-run audit_check.sh after Step 7; don't skip "verify after fix"
4. uniq -c must reveal zero duplicates — don't just count totals; count uniques and compare
5. Upload/move can hit resource contention — retry individually; don't parallelize moves >2 at once
6. Single docx upload loses folder_token parameter — use 3-step: upload (to root) → move (to folder) → verify

References

• references/workflow-checkpoints.md — Detailed checkpoint for each of the 9 steps
• references/audit-script-template.sh — Bash template for machine-verified Audit
• references/subagent-spec-examples.md — Example sub-agent task specs for different domains
• references/case-study-supplements.md — Full case study: Supplements Knowledge Base v1.0 → v1.1 (includes the 5 blind spots found via dogfood)

Related Skills

• skill-distiller — This skill was created by distilling BUILD-PROTOCOL.md (see that file in workspace)
• pptx-presentation-builder — For PPT-specific knowledge products
• Future: build-protocol-engineering — For software/design projects (separate rulebook)

Why This Way (Meta)

Why not just write a long prompt each time? LLM behavior degrades with context length and prompt complexity. A rules-as-file approach + machine verification scripts catches more issues than any one-shot prompt. Distilling this as a skill means: future-me gets automatic loading when the OpenClaw matcher sees trigger words, without having to re-invent the workflow.

Why independent Audit is unmissable? Research: LLMs self-reviewing their own output show ~30% sycophantic bias. The writer agent cannot objectively grade itself. An Audit agent (or main-agent with different context) operating on the deliverable alone — without seeing original requirements — produces harder judgments.

Why ≤2 parallel sub-agents (for writing tasks)? Multi-agent conflict points grow as N(N-1)/2. For long-form writing tasks specifically, where output coherence matters and sub-agents share file regions, 4+ concurrent sub-agents hit 75% failure (timeout / throttling / thundering herd on same API region). 2 parallel is the sweet spot; 2+2+1 batching for 5-agent workloads.

For non-writing agentic tasks (independent file outputs, less coherence-sensitive), the cap can be relaxed up to 4. See trinity-harness for the general-purpose ceiling.

Why quantify anti-sycophancy? Soft rules ("have some negatives") get ignored. Hard quantitative rules ("🔴 count ≥ 20% of 🟢 count") can be machine-verified. grep -c "🔴" file.md enforces the rule; grep finds what opinions miss.

---

Distilled: 2026-04-29 · Source: BUILD-PROTOCOL.md v1.1 (refined through dogfood) Validated on: AI/ML textbook, tactical gear KB, supplements KB (v1.0 → v1.1 dogfood)