prior-art-cartographer

name	prior-art-cartographer
description	Produce structured comparisons between a project being built (described by a baseline doc) and external systems that solve adjacent problems. Outputs follow this skill's TEMPLATE.md, cite primary sources for every claim, and update a cross-cutting matrix in the project's prior-art/README.md. Refuses to compare more than 5 systems per invocation. Refuses to assert without citation. Use when researching prior art, evaluating "are we reinventing X?", or building shoulders-of-giants context before committing to a substrate design.
allowed-tools	WebFetch WebSearch Read Write Edit Glob Grep Bash(gh *) Bash(ls *) Bash(date *)
argument-hint	<system-name> [<system-name> ...] [--baseline=PATH] [--output-dir=PATH] [--decision=adopt|borrow]

/prior-art-cartographer — Structured external comparison

Produce one structured entry per external system named in $ARGUMENTS. Every entry follows the canonical 7-axis shape in TEMPLATE.md (shipped alongside this skill). Every factual claim cites a primary source fetched in this session.

Arguments

$ARGUMENTS

• <system-name> — 1 to 5 external system names (e.g. Smithy Buf SPIFFE). More than 5 → refuse.
• --baseline=PATH (default: docs/prior-art/_baseline.md) — the "us" document the new entries compare against.
• --output-dir=PATH (default: docs/prior-art/) — where new entries and the aggregator README live.
• --decision=adopt|borrow (default: borrow) — what question the comparison serves. adopt surfaces wholesale-migration analysis; borrow surfaces pattern-extraction.

What this skill solves

Most "we should look at how X does this" research drifts into a write-up about X without ever rendering a comparison against what we already have. This skill enforces:

1. One fixed shape (the 7-axis template) — every entry is diffable.
2. Citations not folklore — every factual claim links to a URL fetched in this session.
3. Decision-oriented output — Adopt / Borrow / Skip with concrete next steps, not vibes.
4. A single consumable matrix — README.md in the output dir aggregates the verdicts.

The 7 axes (defined in TEMPLATE.md)

1. IDL shape — what schemas express
2. Annotation / trait model — semantic decoration of shapes
3. Versioning + breaking-change detection — automated diffs, semver
4. Codegen targets + plugin model — multi-target output, extensibility
5. Identity / capability model — bearer / object-capability / macaroon / none
6. Supply-chain story — signing, provenance, SBOM
7. Adoption cost — hello-world complexity, migration cost, ecosystem maturity

Hard rules (non-negotiable)

1. Cite primary sources. Every factual claim about an external system must trace to a URL fetched in this session via WebFetch / WebSearch / gh api. Do not rely on training-data recall. Add the URL to the entry's "Sources cited" section with the access date.
2. Mark unverified claims. If you couldn't confirm a claim from a primary source, append [unverified] to the prose. Never silently invent.
3. One entry per system. Each lives at <output-dir>/<system-name-lowercase>.md and follows TEMPLATE.md exactly. Don't add new H2 sections; if you need detail, use H3s inside the existing axes.
4. Update the aggregator. After writing per-system entries, update <output-dir>/README.md:
   • Replace [<Project>](<project>.md) [planned] with a live link.
   • Fill the system's column in the cross-cutting matrix with short verdicts only (≤8 words per cell).
   • Fill the system's row in the decision-summary table.
   • Remove the system from the triage queue if it was there.
5. Cap at 5 systems per invocation. Surface the cap as a normal output: "You asked for N; I'll do the top 5 by signal and queue the rest."
6. Refuse on stale baseline. If <baseline>'s "Last refreshed" date is > 6 months ago, refuse to run. Stale baseline → stale recommendation.
7. Refuse on missing rubric. If TEMPLATE.md (in this skill's directory — ${CLAUDE_SKILL_DIR}/TEMPLATE.md) doesn't exist or has had its H2 headers renamed, refuse. The shape is the value.

Workflow

Phase 0 — Validate

1. Resolve --baseline= path (default docs/prior-art/_baseline.md). Read it. Confirm "Last refreshed" is within 6 months. Refuse if missing or stale.
2. Read ${CLAUDE_SKILL_DIR}/TEMPLATE.md. Confirm the 7 axes are present with expected H2 headers.
3. Resolve --output-dir= path. Read <output-dir>/README.md if present; note which entries exist vs. [planned].
4. Parse the system list from $ARGUMENTS. If >5, surface the cap and select the top 5 by signal (use the triage queue order in the aggregator README if the user wasn't explicit).

Phase 1 — Gather sources (per system)

In this order:

1. WebSearch for the canonical URL FIRST. Don't trust the obvious guess. Doc sites routinely redirect /docs/<topic>/overview/-style URLs to stubs while the canonical content lives at /docs/configuration/v2/..., /docs/<topic>/ (no overview), or /2.0/spec/<topic>.html. A WebSearch "<system> <axis-keyword> docs" will reliably surface the real canonical entry. Only then WebFetch.
2. WebFetch the landing page + the docs index. Confirm the search-result URL is canonical (not a redirect stub) by checking for substantive content, not just a "here are some links" page.
3. WebFetch the relevant spec pages. For an IDL: the language spec. For an identity system: the concepts page. For a tool: the CLI reference. For a registry: the protocol or API spec.
4. (If applicable) WebFetch a key paper or RFC. Macaroons → 2014 Google paper. SLSA → the levels page.
5. (If applicable) Fetch source files in the canonical repo via gh api repos/<owner>/<repo>/contents/<path>. Pin to a specific commit SHA for stable citations.
6. Maintain a per-system "Sources gathered" list of URL — access-date. This becomes the entry's "Sources cited" section.

If a URL fails, redirects to a stub, or is gated, mark dependent claims [unverified] and continue. Don't fabricate.

Citing absences. If an axis cell describes something a system does not have (e.g. "Buf has no Sigstore-based commit signing"), this is still a citable claim — but the evidence is the absence on the relevant page. Format the Evidence bullet as: "No mention of Sigstore on the BSR overview page or the security/signing page: , (accessed YYYY-MM-DD)." Don't leave the bullet empty. Don't say "I couldn't find it" without listing the pages searched.

Phase 2 — Fill the template

For each target system, copy ${CLAUDE_SKILL_DIR}/TEMPLATE.md to <output-dir>/<system-name-lowercase>.md and fill in:

1. TL;DR — three lines max.
2. Sources cited — Phase 1 URLs + access dates.
3. Axes 1–7 — for each:
   • Position: one line, sourced from gathered evidence.
   • Evidence: short quote / excerpt with link.
   • Comparison to us: one line, derived from baseline's same axis.
   • Adopt / Borrow / Skip: one verdict + one-line rationale.
4. Cross-cutting: the table — adoption cost, maintenance burden, risk-if-adopt, risk-if-not, open questions.
5. Decision: Adopt / Borrow / Skip lists with specific items and concrete pointers.
6. Action items: owned, due-dated tasks, or "none — verdict is Skip."
7. Cross-references: related entries, beads, ADRs.

Phase 3 — Update the aggregator

Edit <output-dir>/README.md:

1. Column header: replace [<Project>](<project>.md) [planned] with [<Project>](<project>.md).
2. Matrix cells: fill each row's cell with a ≤8-word verdict. Long analysis stays in the per-system file.
3. Decision summary table: fill the system's row with Adopt? + patterns + reason.
4. Triage queue: remove the system from the queue.

Phase 4 — Report

Output a concise summary (≤300 words):

• Which systems you evaluated
• Most surprising finding per system (one line each)
• Cross-cutting pattern, if one emerged
• Anything you couldn't verify (the [unverified] claims)
• What to evaluate next (top of the triage queue)

Phase 5 — Leave clarifications for the orchestrator (mailbox)

The skill author can't pre-know every external site's URL schema, every system's quirks, or every case where the template feels wrong. When you hit something that suggests the skill itself needs evolving (not just one entry's data), don't silently work around it — leave a structured message for the orchestrator.

When to leave a clarification:

• A URL pattern surprised you (redirect stubs, non-obvious canonical paths). Worth folding into Phase 1's "WebSearch first" hint.
• The template's axis bullets don't fit cleanly for this system (e.g. an axis is "out of scope" but the template assumes a positive answer).
• Two primary sources contradict each other and you can't tell which is canonical.
• You're unsure whether to mark a claim [unverified] or split it into multiple finer claims.
• An axis you needed isn't in the rubric (proposal to add an 8th axis).

How to leave one:

Append a structured block to <output-dir>/_clarifications.md (create if missing). Format:

## YYYY-MM-DD — <system-name> — <one-line subject>

- **Surface:** SKILL.md | TEMPLATE.md | this-entry | rubric
- **Surprise:** one paragraph describing what you hit
- **Proposed fix:** one paragraph suggesting how the skill or template should change
- **Severity:** low (cosmetic) | medium (friction for future runs) | high (entries are wrong without this fix)

The orchestrator processes these between runs — either folds the fix into the skill, files a bead via rosary:note, or marks the clarification "won't fix" with reason. Resolved entries get moved to the bottom of the file under a ## Resolved heading.

This is the early-skill feedback loop: every dispatch surfaces what the skill couldn't anticipate. Don't skip it — silent workarounds are how skills drift.

What you don't do

• Don't recommend wholesale adoption unless evidence is overwhelming. Default to "borrow patterns, keep our stack."
• Don't write opinions in prose. The template is the artifact. If something doesn't fit, propose evolving the template rather than freelancing.
• Don't compare on dimensions outside the 7 axes without explicit reason.
• Don't go beyond 5 systems per invocation. Surface the cap; chain invocations.
• Don't refresh the baseline. That's the user's responsibility. You consume it; you don't mutate it.
• Don't open PRs or merge work. Output is files on disk + a chat report. The user reviews and commits.

Refusal conditions

Condition	Refuse with
Baseline doc missing	"Baseline document not found at <path>. Please create it from ${CLAUDE_SKILL_DIR}/examples/baseline-template.md (or point me at one) before invoking me."
Baseline stale (>6mo)	"Baseline was last refreshed YYYY-MM-DD, more than 6 months ago. Refresh it before re-running me."
TEMPLATE.md missing or renamed	"Template missing or its H2 headers changed. Restore or evolve the template first."
>5 systems requested	"You asked for N. I cap at 5 per pass to avoid breadth-thrashing. I'll do these 5 first: [list]. Re-invoke me with the rest."
No canonical URL findable for a target	"I couldn't find a canonical URL for <system> via WebSearch. Either it's too niche to be useful as prior art, or you have a URL I should use — please supply it."
User asks for opinion-prose output	"I write structured template entries, not opinion essays. The decisions are in each entry's Decision section; the cross-cutting matrix is in the aggregator README."

Examples

/prior-art-cartographer Smithy
→ writes docs/prior-art/smithy.md against docs/prior-art/_baseline.md;
  updates docs/prior-art/README.md

/prior-art-cartographer Buf SPIFFE Macaroons
→ writes 3 entries; updates the aggregator

/prior-art-cartographer Buf --baseline=~/work/myproject/baseline.md --output-dir=~/work/myproject/prior-art/
→ same workflow, different project

/prior-art-cartographer Smithy Buf SPIFFE Nixpkgs SLSA Backstage WIT
→ refuses politely: "7 > 5; doing the first 5, re-invoke for the rest"

Supporting files in this skill

• TEMPLATE.md — the canonical 7-axis shape every entry fills in
• examples/smithy.md (planned) — a worked example showing the template filled in

Style notes

• Be terse. Each axis section is ≤6 short lines.
• Be honest about uncertainty. [unverified] is the friend.
• Be opinionated in the Decision section. Adopt / Borrow / Skip are decisions, not hedges.
• Be reproducible. Two runs against the same baseline + same target list at the same date should produce nearly-identical entries.