Create executable documents that demonstrate an agent's work
verify command re-executes every code block and diffs against recorded output, turning documentation into a live test suite; the tool hit 1,145 stars in under six weeks, authored by Simon Willison (Datasette, LLM).| Rating | Summary | |
|---|---|---|
| Quality | excellent (20/24) | Well-maintained by a prolific open-source author, substantive README with full command reference, CI, typed language — minor gaps in external docs link and a missing limitations section. |
| PAI Relevance | integrate (0.88) | Fills a clear gap in agent accountability; subprocess-ready Go CLI with structured markdown output maps directly onto PAI's CLI-first architecture and has no equivalent in the current capability manifest. |
20/24 — actively-maintained / adequately-documented / high-discipline
Failed:
Passed:
Failed:
Passed:
Failed:
Passed:
go install github.com/simonw/showboat@latest and go run github.com/simonw/showboat@latest both require go.mod; its existence is definitively implied by the build and distribution evidence| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 2 | The build→exec→verify pattern with output diffing is novel for PAI; it directly informs a future agent-accountability subsystem and could enrich the existing Evals skill with a structured proof-of-work document format. |
| Integration Readiness | 2 | Go binary available via uvx showboat with no installation required; produces structured markdown output to files; fully offline and subprocess-friendly — PAI skills can invoke it directly without adapter code. |
| Overlap Risk | 1 | Partial conceptual overlap with PAI's Evals skill (both address agent output verification), but the specific executable-document format, pop undo semantics, and verify diff command are not covered by any current PAI capability. |
| Gap Fill | 2 | PAI has no capability for producing structured, re-executable, verifiable records of agent work sessions; showboat fills a clear gap in agent accountability and reproducible documentation distinct from the WORK/ file pattern. |
Composite: 0.88
Capture-to-Knowledge Pipeline / clean-room Haiku validation stage: Run pip install showboat and configure the Haiku validation agent to write each validation run as a showboat .md file — code blocks that call the validation logic, with output captured inline. The artifact from each run is then independently verifiable: showboat verify <run>.md re-executes the same checks and diffs against recorded output, replacing "the agent said this passed" with a reproducible proof that can be re-checked weeks later as the pipeline matures.
Fabric Recommender (fab) regression coverage: Create a demo.md showboat document with fab invocations on a representative set of content/intent pairs — run showboat run demo.md once to capture live output. Add showboat verify demo.md as the pre-release check before any future fab version bump. Any silent change to pattern-matching or ranking that alters recommendations surfaces immediately as a diff, without writing a separate test suite.
Agent-assisted work on Conservancy or Aphoria: When delegating dev tasks to Claude Code on either project, require the agent to deliver a showboat .md as its work artifact — with code blocks that demonstrate the change (a build invocation, a content generation call, a deploy step) and captured output proving the result. This gives each completed task a human-readable summary that is also a re-executable verification check, replacing ad-hoc "it worked when I ran it" review with a durable, re-runnable proof.
Category: Developer Tools & CLIs
In this category: casey--just (excellent, 21/24)
Standing: Showboat is the only agent-documentation CLI in the category; where just provides a stable subprocess interface for development recipes, showboat provides an executable-document format purpose-built for AI agent proof-of-work — the two are complementary rather than competitive.
Density: 8/10 — Available: full README (~8KB), release metadata (v0.6.1), star and fork counts, creation date, last commit date, open issue count, license, language, CI badge. Missing: go.mod dependency manifest (contents not surfaced), full contributor list, test suite details.
Simon Willison shipped v0.6.1 within 17 days of the initial commit, suggesting the core feature set was intentionally scoped and completed quickly rather than abandoned. The gap between the last release (2026-02-23) and last commit (2026-03-14) indicates a brief maintenance window followed by apparent stabilisation — not decay. The Go-binary-on-PyPI distribution pattern (also used by his other tools) is a deliberate choice to reach Python-ecosystem users without requiring a Go toolchain. The extract command — which emits the CLI commands that would recreate a document from scratch — is a sleek self-documenting property worth studying as a design pattern for any PAI skill that produces structured artefacts.