Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Design & UI

tech-design

Take a PRD (or feature description) and produce a complete backend technical design document through a structured process: **understand → explore → grill → write → self-review**.

open-octo · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add open-octo/octo-agent --skill tech-design

About this skill


name: tech-design license: MIT description: Produce a complete backend technical design document from a PRD or feature description — understand, explore the codebase, grill on decisions, write, and self-review. Use when the user wants to write a tech design, technical proposal, or backend design doc, e.g. "写技术方案", "出个设计文档", "tech design", "technical proposal", "帮我写后端设计". For pressure-testing decisions first, use the grill-me skill; to build from a finished design, use the implement skill. metadata: origin: adapted for octo — generic stack, no org-internal KB/lint tooling

Skill: tech-design

Take a PRD (or feature description) and produce a complete backend technical design document through a structured process: understand → explore → grill → write → self-review.

Rules (the reason this skill exists)

These six rules run through the whole process. Follow them while writing (Phase 4), then grep for violations in the self-review. Most rework on a design doc traces back to breaking one of them.

R1. Never invent placeholders (API / field / service / URL / enum)

Every concrete technical name must be grepped from real code, looked up in docs, or asked of the user — never a plausible-sounding placeholder left for review to fix. Inventing svc.GetBookingNotes when no such method exists, or guessing a column is matter_type when it's really ticket_type, is the single most common source of production bugs that pass mocked tests.

R2. Never invent phasing

The design scope is strictly equal to the PRD scope. There are only two legal sources of phasing: (a) the PRD itself marks priority (P1/P2 in the stories), or (b) the user explicitly said "do X first, not Y" during grilling. Otherwise, if the PRD lists N sub-scenarios, design N.

Do not use P1/P2/P3 (or Phase 1/2/3) prefixes to label features, sections, or upstream links — even as "neutral numbering". A P-prefix is always read as priority/phasing and manufactures an ordering the PRD never stated. Use the PRD's own section names instead. Real execution order belongs only in a "release order" section, and only when driven by a dependency chain, not by scope-cutting.

R3. Technical facts must be traceable

Every concrete technical claim (DB column, URL prefix, field, API signature, enum value, cache key pattern, MQ topic) must come from one of: a code location (file:line), project docs, or a grill answer (quote the decision). Can't find a source → ask the user (they're online; asking is cheap). Never paper over a gap with a TODO.

Decision-level choices (sync vs async, new table vs extend column) can rest on a grill answer plus explicit trade-off reasoning — but factual assertions must be greppable.

High-guess areas (grilling usually misses these; handle by "check docs → check code → ask user"): compatibility design, security design, circuit-breaker thresholds, rollback plan, external dependency interface tables.

R4. The document must read context-free

Write for a future reader who wasn't in this conversation. Banned phrases: grill-question references (the Q3 decision, per the original Q3), grill-option letters (go with option A, path A — the reader has no idea what A was; inline the actual content instead), conversation references (after we discussed…, the earlier path), and editing-process narration (revised plan, originally we…).

Replace with the direct technical reason: "We use X rather than Y because … (name the actual X and Y, never bare option letters)."

Also context-bound: undefined jargon. Any domain term or coined abbreviation must be defined on first use or in a glossary. Test: would someone who only reads this doc, wasn't in the conversation, and isn't deep in this domain know what the word means? If not, define it.

Exception: a version-history changelog may say "X changed to Y", but must not reference a grill question.

R5. Examples must be empirical

A concrete example ("service X runs on cluster Y", "call chain A→B→C", "this field is usually N") must be grepped / measured / confirmed from logs. Ask "can I reproduce this example?" — if not, don't write it. An honest "needs verification" beats a wrong example that a later reader treats as truth. An old default config existing ≠ the current fact being correct; infrastructure migrates.

R6. Resolve every uncertainty now — never defer it to the implementation phase

The finished design must be accurate and complete: an implementer builds from it without discovering that a decision was left open. A "待实现阶段确认 / to be confirmed during implementation / TBD" item is a design defect, not a legal hand-off — it is exactly the gap this skill exists to close. Every uncertainty you hit while writing (Phase 4) or in self-review (Phase 5) must be closed before the doc is done. There are two kinds, and each has one correct way to close it:

  • A code-verifiable fact — does this RPC actually return field X? what is the struct's real shape? which column / tag / enum? → resolve it yourself by reading the code (grep / codegraph / trace the call chain to the source). Never turn a checkable fact into an open question for someone else.
  • A genuine decision — which of data-source A/B/C, sync vs async, extend an existing struct vs add a new RPC, each with different costs → re-enter the grill-me skill and drive it to a decision with the user, one question at a time. Then write the chosen approach in as settled fact (per R4 — the actual choice and its reason, never a bare option letter).

If you catch yourself writing an "open questions" / "实现阶段需确认" list that shifts a decision downstream: stop. Check the code, or grill the user, then write the answer. The only thing that may remain unresolved is something genuinely outside the design's control (e.g. an upstream team owes a field that does not exist yet) — and that is recorded as a **blocking external depen

Read the complete source ↗