| name | docs-as-deliverable |
| description | Use when shipping or handing off code — treat docs as a first-class deliverable: present-tense, diagrammed-as-code, verified against reality, reader-tested. |
Docs as a First-Class, Diagrammed, Tested Deliverable
Documentation is a deliverable, not exhaust. Hold it to the same bar as the code: tight present-tense prose, diagrams expressed as code, every signature and flag verified against the implementation, and comprehension proven by readers — not assumed.
When to use
Apply whenever you ship code, hand a project to another team, write or revise a README/ARCHITECTURE/runbook, or change behavior that an existing doc describes. Documentation work belongs in the same change as the code it documents — not a follow-up "docs pass" that never comes.
Red-flag thoughts that mean STOP and apply this skill:
- "I'll update the docs later / in a separate PR."
- "The diagram is close enough — the code only moved a little."
- "This describes how it should work" (instead of how it does).
- "Everyone on the team already knows this, so I don't need to spell it out."
- "I'll keep the old design doc around as-is; people will figure out it's outdated."
- "I'll hand-type the config table into the docs so it reads nicely."
- "The doc's example commands are close enough — nobody runs them literally."
The rule
- Write to a reader, in present tense, about what is. Describe the system as it actually behaves now. Cross-link instead of repeating; one authoritative statement per fact. Cut hedging and filler.
- Verify every concrete claim against the source. Open the code/config and confirm each function signature, CLI flag, endpoint path, env var, and default. If you cannot verify it, either verify it or do not claim it. Commands are claims too: any command a reader will copy-paste must be run, as written, against the current repo — a doc telling the reader to build a venv with an interpreter the package no longer supports fails in the first second of actually trying it.
- Express architecture and flows as diagrams-as-code. Embed parseable diagrams (e.g. Mermaid) directly in the markdown: a context/container view of the system, a sequence diagram for any multi-step or pause/resume flow, a state machine for lifecycle, and a swap map showing which interfaces have which interchangeable implementations. For graph-shaped diagrams, put a small node/edge table beside the diagram so prose, picture, and code can be checked against each other. Diagram support is toolchain setup: if the docs pipeline cannot render the diagram fence (a Sphinx build with no Mermaid extension, a wiki that strips code fences), provision the renderer before writing the doc — otherwise authors quietly fall back to the ASCII art this skill's anti-patterns forbid.
- Change the diagram in the same commit as the code. When you alter a flow, a boundary, or a state, update its diagram in that same change so it tracks implemented reality and never silently drifts.
- Generate authoritative tables; don't hand-copy them. Tables of flags, routes, config keys, or metrics should be produced from the code/config (a script, a
--help dump, a schema export) rather than maintained as a parallel hand-edited copy that rots. The same goes for embedded excerpts of machine artifacts — a workflow file, a config block: link the canonical file or generate the excerpt, because a hand-copied sketch drifts until it describes runners and job names that no longer exist.
- Fence off superseded docs; never silently delete or silently keep them. Preserve historical/design artifacts for provenance, but put a one-line banner at the top pointing to the current authoritative doc, so no reader mistakes a stale snapshot for truth.
- Test the docs with context-free readers. Have several reader agents or people, each holding only the documentation (no codebase, no tribal knowledge), try to act on it from a distinct persona — new developer, on-call/SRE, end user, skeptic. Log every point where they get stuck, guess, or are misled as a documentation defect, and fix it like a bug.
- When automation implements a procedure, the doc is its operator's guide — the automation is the source of truth. A runbook narrating manual steps beside a pipeline that performs them will eventually contradict it, and not cosmetically: a manual publish instruction next to a tag-triggered publish workflow is a double-publish. Document the trigger, what the automation does, and how to recover when a stage fails; demote any manual path to an explicitly-labeled fallback that is mutually exclusive with the automated one.
Why
Docs are the interface a stranger meets before the code. If they are aspirational, unverified, or drifted, the cost is paid later and larger: a new hire burns a day on a flag that was renamed, an on-call engineer follows a runbook into an outage, a client integrates against an endpoint that no longer exists, a reviewer trusts a diagram that lies about the data flow. Verifying against source, generating tables, and binding diagram changes to code changes remove whole classes of drift at the source. Reader-testing converts "I think this is clear" into evidence that someone with no prior context can actually act — which is the only definition of clear that matters.
In practice
On the project this was distilled from — a hexagonal, human-in-the-loop AI agent deployed to a cloud runtime — the docs were rewritten as a coherent client-facing deliverable in one focused pass: present-tense, with every function signature, CLI flag, and endpoint re-checked against the live code before being written down. That same pass also deleted the development-history docs (BRIEF, DESIGN, SPECS, PAPER, ...) — the very anti-pattern this skill warns against — and the correction was a delete→restore→banner arc: the artifacts were restored, then marked as historical with a dated one-line banner pointing at the current authoritative doc, so provenance survived without anyone mistaking a stale snapshot for truth. The architecture doc gained four embedded diagrams in a single change: a context view of the whole system, a sequence diagram of the pause/resume handoff to a human, a map of which ports were backed by which adapters, and a container view; the agent-framework view additionally carried a node/edge table beside it so prose, picture, and code stayed mutually checkable. The operations runbook generated its table straight from config rather than hand-copying it — "the table below is generated from it" (docs/OPERATIONS.md). Finally, the docs were graded by context-free reader agents given only the documentation, each reading as a new developer, an SRE, an end user, and a skeptic; their comprehension failures were filed and fixed as documentation defects (recorded in the evaluation write-up). You can apply all of this without knowing that project: write present-tense and verified, embed your diagrams as code and update them with the code, generate your tables, restore-and-banner old docs rather than deleting them, and let fresh readers prove the result.
Anti-patterns
- Aspirational docs: describing intended or hoped-for behavior instead of what the code actually does today.
- Drifting diagrams: ASCII art or external diagram files (or screenshots) that no longer match the implemented graph because nothing forces them to move with the code.
- Unmarked or deleted history: erasing development/design docs (losing provenance) — or leaving them unbannered so they get mistaken for current truth.
- Untested docs: treating documentation as an afterthought that is never checked against whether a newcomer with no context can act on it.
- Runbooks that fight their automation: manual procedure steps documented beside a pipeline that already performs them — followed literally, the two paths collide (a double-publish, a duplicate release).
Enforcement
What a machine can check: drift, in every direction it happens. Docs built in CI with warnings-as-errors plus linkcheck; diagram fences proven rendered — the renderer declared in the docs dependency group, and the built output containing one rendered node per fence; generated tables regenerated in CI and diffed against the committed copy; copy-paste commands smoke-run where feasible. Reader-testing stays human — schedule it like a gate, don't simulate it with a linter.
Related skills:
