desplega-ai/ai-toolbox/design-docs

Design Docs

A design doc is the durable statement of intent for one system: what it's for, what words mean, what must always hold, and what it deliberately doesn't do. It lives at thoughts/<username|shared>/design-docs/<system-slug>.md — no date prefix; it's a living doc, amended in place, never superseded by a new dated file.

It is normative: the desplega:code-reviewing Spec axis checks diffs against the Invariants section, and researching/planning/one-shot must read-and-abide (below). A design doc nobody enforces is a wish; this one is enforced at review time.

Structure (in this order — template at cc-plugin/base/skills/design-docs/template.md)

1. Purpose — one paragraph. What the system is for and for whom. If it takes two paragraphs, the system boundary is probably wrong.
2. Glossary — the ubiquitous language: each core term, one-line definition, plus "avoid:" aliases (terms people misuse for it). Zero implementation detail — a glossary entry naming a class or table has leaked.
3. Invariants — numbered (I1, I2, …), each one testable — a reviewer must be able to look at a diff and answer "does this hold?" with yes/no. These are spec, not style; vague aspirations ("should be fast") don't belong here.
4. Boundaries & Non-goals — what the system explicitly does NOT do, including rejected scope with the reason it was rejected (so the next person doesn't re-propose it from scratch).
5. Interfaces / Seams — how other systems touch this one, high level only (names and responsibilities, not signatures).
6. Decision log — one-paragraph ADR entries, newest first: context → decision → consequence. Only decisions that are hard to reverse, surprising, or a real trade-off — routine choices are noise here.
7. Amendment log — one line per change to this doc: date, what changed, which plan/session drove it.

Creating one

On demand, when a system is worth pinning down. Grilling-style Q&A (same rules as desplega:brainstorming):

• Facts vs decisions — never ask what you can look up. Current behavior, existing terms, actual boundaries: spawn quick background sub-agents (routed per desplega:delegate-work) and draft from findings. Only genuine decisions — naming calls, invariant strength, scope rejections — go to the user, via AskUserQuestion with a recommended answer on every question.
• Draft the doc from the template, section by section, filling what research settled and asking only what it couldn't.
• Finish with /file-review:file-review <path> so the user can mark up the draft.

Amending

Design docs change through work, not drift: when a plan or one-shot intentionally changes the design, updating the doc is part of that work — append the Decision log entry (if the change was a real decision) and the Amendment log line in the same session. There is no periodic drift audit by design; enforcement is read-and-abide plus the review Spec axis.

Read-and-abide (enforcement)

For any skill touching a system that has a design doc:

• Read it first. The Glossary is canonical vocabulary for anything you write about the system.
• Don't violate it. Invariants and Boundaries constrain plans and code alike.
• Flag conflicts, never silently work around them. If the task requires breaking an invariant or crossing a boundary, surface it to the user as an explicit decision (which, if taken, amends the doc). If existing code contradicts the doc, report that as a finding — doc says X, code does Y — and let the user pick which one is wrong.

Wired into: desplega:researching and desplega:planning (read-and-abide rules), desplega:one-shot (Step 2), desplega:code-reviewing (Invariants as Spec-axis source).