design-doc-full-pass

Orchestrates parallel read-only agents to edit large design docs section by section.

Updated Jul 13, 2016
One-click install
npx skills add https://github.com/yang-l/configs --skill design-doc-full-pass-yang-l
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: design-doc-full-pass
Source: https://github.com/yang-l/configs/tree/main/.claude/skills/design-doc-full-pass
Command: npx skills add https://github.com/yang-l/configs --skill design-doc-full-pass-yang-l

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Editing a large or heavily-reviewed design doc, RFC, or spec single-threaded is slow and risks breaking headings, anchors, intra-doc links, and pinned review comments. This Skill fans out one read-only agent per section, collects verbatim old_string to new_string edits with rationale and risk tiers, and applies them under a whole-doc structural guard. ## Core Features & Use Cases - Parallel section editing: Partitions the doc on heading boundaries and dispatches up to 20 concurrent read-only section agents, each returning exact string replacements with a one-line rationale and risk tier. - Whole-doc integrity guard: The coordinator verifies every old_string is unique across the entire file, keeps headings, tables, code blocks, and anchors byte-identical, and downgrades cuts that hit link targets, table mirrors, or review-anchored passages. - Risk-tiered cut policy: Clean redundancies auto-apply after cross-ref checks, framing-only edits are flagged for the author, and anything touching rationale is discarded. - Use Case: Before shipping a 40-section RFC that is already under GitHub review, run a full pass so every section gets tightened in parallel while the verifier confirms all links, anchors, and reviewer-pinned passages still resolve. ## Quick Start Ask the assistant to do a full pass over your large design doc or RFC, reviewing every section and tightening it while preserving structure and anchors.

Frequently Asked Questions about design-doc-full-pass

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I review a large design doc or RFC section by section with AI?▼

Use a full-pass workflow that partitions the document on heading boundaries and dispatches one read-only agent per section. Each agent returns exact old_string to new_string edits with a rationale and risk tier, which a coordinator then applies under whole-doc structural checks.

When should I use design-doc-full-pass instead of the design-doc skill?▼

Use design-doc-full-pass only when the doc is large or heavily reviewed enough that parallel section agents justify their coordination cost. For a single section, one paragraph, or a quick tighten, defer to the plain design-doc skill, which owns the underlying editorial rules.

How does it avoid breaking links, anchors, and tables while editing?▼

The coordinator builds a cross-reference map of headings, anchors, intra-doc links, table mirrors, and review-comment anchors before dispatching agents. Every edit is applied as an exact string replacement verified unique across the whole file, and any edit that would alter structure is rejected.

Can it handle docs under external review on GitHub or Confluence?▼

Yes. Where the platform is reachable, it pulls review threads to map which passages reviewers have pinned. When threads cannot be fetched, it conservatively flags any passage a reviewer may have anchored to instead of auto-cutting it.

What happens to edits that touch the doc's rationale?▼

Edits that touch rationale are discarded outright, since rationale is the most protected content. Framing-only edits are flagged for the author's decision, and only clean redundancies that pass cross-reference checks are auto-applied.

What are the limitations of the multi-agent full-pass approach?▼

Section agents are read-only and blind to the rest of the doc, so the coordinator must own all cross-reference integrity. Fan-out is capped at 20 concurrent agents, and Workflow-tier orchestration requires explicit user opt-in plus a runbook author.