| name | common-design-docs-k |
| type | knowledge |
| description | The agent documentation system — doc types, folder structure, frontmatter conventions, agent roles, lifecycle |
| user-invocable | false |
Knowledge skill — Design documentation system: doc types, folder structure, frontmatter conventions, agent roles.
Agent Documentation System
Agent docs are design documentation that agents create and maintain — plans, living docs, architecture maps, and principles. They live in design-docs/ folders, separate from external docs (API references, generated documentation, user-facing guides).
Core Doctrine
Visual-first, prose-second. Diagrams, tables, and code blocks over paragraphs. Prose explains why; visuals show what and how. Every sentence must be precise and information-dense.
Watch doc size. When a single doc exceeds ~1000 lines, it's a sign it should be split. Monolithic specs that grow to thousands of lines exceed agent read limits and mix concerns that belong in separate files. Split by logical boundary — per-block plans, per-component living docs, per-concern sections into their own files.
Code in plans — minimize.
- Design plans: no code snippets. Describe interfaces, contracts, data shapes — not implementation.
- Implementation plans: code is expected but surgical. Show signatures, key types, critical algorithms. Not full implementations.
Documentation is every agent's duty. Not the planner's job. Not "someone else will update this." Every agent — planner, builder, verifier, gardener — treats documentation health as a core output of their work. The next session's agent inherits what you leave behind. Stale docs, missing context, undocumented decisions — these are bugs you're shipping to the next version of yourself. Treat them with the same severity as broken tests.
Reflect before you finish. Before completing any documentation task, step back: Does this doc make the system more legible to an agent starting cold? Would you understand this if you had no prior context? If not, fix it.
Foresight vs Hindsight
Plans are foresight. Written before building. "Here's what we're going to do."
Everything else is hindsight. Written after building. "Here's how it works now."
Never write living docs about code that doesn't exist yet. Never plan in hindsight.
Doc Types
Plans
Bounded work with a lifecycle: draft → active → complete. Plans have milestones, acceptance criteria, and a defined end state. Once complete, they become historical records.
There are two kinds of plans:
- Design plans — system design, architecture, how components interact. Higher-level. How things should work.
- Implementation plans — code-level. What the code looks like, what to build, in what order. Milestones, steps, acceptance criteria.
Small features may combine both in one document. Larger efforts separate them — a design plan at a higher level, implementation plans nested beneath for each buildable piece.
Plans live centralized at design-docs/plans/. Plans can nest — a parent plan defines vision and ordering, child plans are independently buildable. Sibling plans are parallel efforts (different services, different components). Nested plans are phases of the same effort (blocks within a service).
Plans are never condensed, summarized, or rewritten. Every diagram, rejected alternative, and reasoning chain is part of the decision trail. When reorganizing or moving plans, use cp — not "summarize and rewrite." An agent's instinct to tidy by condensing destroys the very context that makes plans valuable as historical records.
Living Docs
Describe how things work now. Staleness is a bug. Living docs emerge from implementation — created after building, not before.
- System docs — how a service/module/component works
- Integration docs — how things connect, contracts, boundaries
- Data docs — models, schemas, state machines
- Test docs — test strategy, fixture guides, integration setup
Living docs live near what they describe — distributed, not centralized:
- System-level:
design-docs/ at repo root
- Package-level:
{package}/design-docs/
- Test-level:
{test-dir}/design-docs/
ARCHITECTURE.md
The entry point. System map at design-docs/ARCHITECTURE.md. Links to deeper living docs. An agent starting a fresh session reads this first.
PRINCIPLES.md
Golden rules at design-docs/PRINCIPLES.md. Short, opinionated, evolved through experience.
Folder Structure
design-docs/ # System-level (repo root)
├── plans/ # ALL plans — centralized
│ └── {name}/
│ ├── {name}_plan.md # The plan (foresight)
│ ├── {name}_decisions.md # Decisions made during execution
│ ├── {name}_verification.md # Verifier's report
│ └── {sub-plan}/ # Nested child plans
│ └── {sub-plan}_plan.md
│
├── ARCHITECTURE.md # System map — the entry point
├── PRINCIPLES.md # Golden rules
└── {topic}.md # System-level living docs (hindsight)
{package}/
├── design-docs/ # Package-level living docs
│ └── {topic}.md
└── tests/
└── design-docs/ # Test-level living docs
Use semantic file names — every file should be identifiable by name alone. No generic plan.md or decisions.md that require reading the parent folder to understand.
Frontmatter
Plans
---
status: draft | active | complete
created: YYYY-MM-DD
updated: YYYY-MM-DD
scope: system | package-name
parent: parent-plan-name
related:
- other-plan-name
---
Living Docs
---
created: YYYY-MM-DD
updated: YYYY-MM-DD
plans:
- plan-that-created-this
- plan-that-modified-this
---
Decision Logs
---
plan: plan-name
created: YYYY-MM-DD
---
Entries within docs use [YYYY-MM-DD] or [YYYY-MM-DD HH:MM] timestamps. Update the updated field on meaningful changes.
Agent Roles in Documentation
Every agent owns documentation quality. The table below shows the minimum — go beyond it when you see gaps.
Agent │ Creates │ Maintains │ Flags
──────────┼──────────────────────┼────────────────────────────┼──────────────────────
Planner │ Plans, decisions │ Seeds ARCHITECTURE.md, │ Missing context,
│ │ PRINCIPLES.md if absent │ stale entry points
──────────┼──────────────────────┼────────────────────────────┼──────────────────────
Builder │ Living docs after │ Updates plans (progress, │ Drift between plan
│ implementation │ surprises), ARCHITECTURE │ and reality
──────────┼──────────────────────┼────────────────────────────┼──────────────────────
Verifier │ Verification report │ Validates doc accuracy │ Stale docs, missing
│ │ alongside code quality │ coverage, doc drift
──────────┼──────────────────────┼────────────────────────────┼──────────────────────
Gardener │ — │ Audits all docs against │ Orphaned plans,
│ │ reality, prunes, cross- │ broken links, gaps
│ │ links, marks complete │
If you see a stale doc while doing other work, fix it or flag it. Don't walk past it.
Lifecycle
- Planner creates plan (draft → active)
- Builder implements, updates plan (progress, surprises, decisions), creates living docs
- Verifier reviews, writes
{name}_verification.md
- Plan marked complete — historical record
- Gardener maintains living doc accuracy over time
