Evaluate

Purpose

Assess a technology, tool, article, or competitive development for relevance to Genesis. Produce a structured evaluation with clear recommendations.

When to Use

• New tool or library surfaces that might replace or augment a Genesis component.
• Competitive product launches or updates (e.g., Cursor Automations, Devin).
• User shares an article or resource for assessment.
• Surplus compute is available and the evaluation queue is non-empty.

Workflow

1. Gather context — Read the target material. If the request supplies URLs, fetch every supplied URL and individually address each source; do not stop because the first source seems sufficient. If a concept, research current state.
2. Search for alternatives and in-flight work — When the target is a tool, product, library or capability (not only an idea in an article), it is one candidate, not the field. Before scoring it, search for what else already does the job, and log each search in the output's **Searched:** line. The routes depend on where you run: a foreground session has Bash and every MCP server; a research background session has recon_github_search / recon_github_read and web_fetch, but no Bash; an inbox evaluation has only the health and memory servers (web_search, web_fetch). A search you have no route for is logged as "not run: no route in this context" and named in next_step, never logged as zero results.
   • Genesis's own queue. Open PRs and issues on the public Genesis repo that touch the same area. Foreground: gh pr list --state open --search "<terms>" --limit 1000, same for gh issue list. Research background: recon_github_search(kind="issues", query="<terms>", state="open", repository="<public repo>") finds issues only; follow its pagination signal. Open PRs have no background route. A draft PR or an issue may already own the capability, or the target may replace one.
   • Existing implementations. Public repositories (gh search repos, recon_github_search) and exact code (searchGitHub / grep.app, foreground only).
   • Claude Code mods and plugins, whenever the capability runs inside a Claude Code session: a hook or guard, status line, prompt or context injection, tool interception, notifications, an in-terminal pane, session awareness. From CC 2.1.287 plugins can carry mods (JS/TS function hooks with UI, tool, agent and model calls); a public catalogue scans over three thousand. Foreground: the mod-finder plugin's search_mods tool if installed (it filters by access, e.g. without: ["network"]), otherwise grep the catalogue it reads, https://raw.githubusercontent.com/karanb192/awesome-claude-code-mods/main/catalogue.md. Any context with web_fetch: Anthropic's directory, https://raw.githubusercontent.com/anthropics/claude-plugins-official/main/.claude-plugin/marketplace.json. The community catalogue has no background route yet (#3321). Check each candidate's validated CC version against the pin in docs/reference/cc-compatibility.md: below 2.1.287 a mod needs the early-access flag. Keep action: ADOPT and record the required pin bump in next_step / scope; do not invent an action enum value. Mods are unsandboxed: record each candidate's access footprint, and never propose a mod as the only enforcement for a guard, because a mod that denies a tool call ran before the classic PreToolUse hooks and suppressed them when measured (#2260). A negative result is "not found in <sources searched>", never "nothing exists".
3. Map to Genesis — Identify which Genesis components or design decisions the target intersects (routing, memory, perception, surplus, etc.).
4. Assess fit — Score along these axes:
   • Capability gap: Does this solve something Genesis lacks?
   • Replacement risk: Could this obsolete a Genesis component?
   • Integration cost: How much work to adopt or adapt?
   • Lock-in risk: Does adopting this violate the flexibility principle?
   • Rigor gap: Where Genesis has an equivalent, is ours as rigorous? "We have X" is not the same as "our X measures effectiveness, handles edge cases, and improves over time." Compare the QUALITY of our implementation against the reference, not just its existence.
   • Overlap Comparison table: When Genesis has a comparable capability, produce the Overlap Comparison table (see Output Format below) instead of prose claims like "we already have this." Required whenever rigor gap is not "N/A — no Genesis equivalent."
5. Recommend — One of: ADOPT, WATCH, IGNORE, ADAPT (take the idea, not the tool). Disposition posture: Genesis's scope is everything digital — default toward ADOPT/ADAPT and acting now. "No current use case", "out of our wheelhouse", and "not obviously valuable" are NOT valid grounds for WATCH/IGNORE. A weaker existing Genesis version means UPGRADE (do the rigor-gap / Overlap comparison), never dismiss. WATCH requires a named re-activation trigger; a trigger-less WATCH is a disguised IGNORE — write it as an honest IGNORE with a reason. ADAPT (stealing patterns/rigor without adopting the code) is common and valuable, but it is not a polite default for "I don't want to act" — reach for ADOPT when the thing itself fits.
6. Write output — Structured evaluation in the format below.

Decision Protocol: Reuse Before Rebuild

Apply this protocol to concrete tools, products, libraries, repositories, and services. The disposition labels are the final roll-up, not the analysis.

1. Name the distinctive mechanism. Decompose the item into the capabilities or operating mechanisms that create its value. Do not compare only category labels: “both route models” says nothing about how either router learns, observes failures, or improves.
2. Separate evidence from inference. Identify what the source demonstrates, what the live Genesis map/code demonstrates, what you infer, and what remains unknown. Do not turn a source claim or a plausible analogy into a fact.
3. Compare mechanism by mechanism. A product can duplicate one Genesis capability and still contain a valuable mechanism Genesis lacks. Rejecting the package or vendor does not dispose of each mechanism inside it.
4. Walk the reuse ladder before proposing implementation: direct use or configuration; library/API/MCP/CLI integration; subprocess, sidecar, or container; reuse of a separable upstream component; adaptation of a proven pattern; only then a new Genesis implementation for the irreducible delta. Language or runtime mismatch changes integration cost; it is not by itself a veto. For commodity tooling, prefer a bounded trial of a mature external implementation before rebuilding it. For anything that runs inside a Claude Code session, an existing mod or plugin is the first rung. Implementing an ADAPT verdict starts with reverse-engineering the reference, not a blank file (this is build-time work, not part of the evaluation itself). Read the candidate's source end to end: the mechanism, its edge cases, its tests, and which failures it already handles. Then carry that over and change only the delta Genesis needs: a port of a proven implementation inherits its debugging, while a rebuild from the idea repeats it. Record the reference's license first. Copy code only when the license allows redistribution in Genesis's public repo; otherwise carry the mechanism over from your reading, not the text. When the reference is a closed binary and the install has a reverse-engineering toolchain, use it in a foreground session only: autonomous and background evaluation never runs it.
5. Compare complete lifecycle cost. External adoption includes integration, operations, lock-in, privacy, and compatibility risk. Internal construction includes design, implementation, testing, security review, battle-hardening, maintenance, and the opportunity cost borne by a single maintainer. “Native is cleaner” is not a cost comparison.
6. Decide at the mechanism level, then roll up. State the viable reuse path considered and why ADOPT, ADAPT, WATCH, or IGNORE beats the alternatives. If recommending new implementation, name why every less-owning reuse rung fails.

This is not an adoption quota. Preserve non-adoption when the evidence shows poor mechanism fit, unacceptable privacy/security exposure, architectural-core conflict, abandonment, or lifecycle cost that exceeds the value. Missing decisive evidence calls for a bounded investigation with the exact question and decision trigger—not a confident rejection or an open-ended WATCH.

Output Format

When invoked from the inbox, follow the output template in INBOX_EVALUATE.md (summary-first, then lens-by-lens). When invoked standalone (e.g., /evaluate), use this structure:

{target title or URL} — {recommendation: ADOPT | WATCH | IGNORE | ADAPT}

Summary

{1-2 paragraphs: what this is, what it means for Genesis, and the key architectural implications. Lead with what matters most. This is a TLDR — if a scoring axis is unremarkable, skip it here.}

Scores: Capability gap: {low|medium|high} · Replacement risk: {low|medium|high} · Integration cost: {low|medium|high} · Lock-in risk: {low|medium|high}

Searched: {one entry per search from workflow step 2: source, "terms", result count; or "not run: no route in this context". Omit the line only when the target is an idea, not a tool or capability.}

Action items:

• {concrete next step if any}

Recommendation

action: ADAPT              # ADOPT | ADAPT | WATCH | IGNORE (default toward ADOPT/ADAPT; WATCH needs a named trigger)
next_step: "One concrete sentence — what specifically to do next"
effort: Small              # Trivial | Small | Medium | Large
scope: V4                  # V4 (do now — DEFAULT) | V5 (sequenced behind named work) | Future (named blocker) | Never (philosophy conflict)
confidence: high           # low | medium | high
architecture_impact: extends  # validates | extends | challenges | irrelevant

Rules:

• REQUIRED on every evaluation. No exceptions.
• The action field must match your recommendation in the Summary.
• next_step must be a single concrete sentence. "Investigate further" is not concrete. "Extract their prompt-versioning schema and compare to genesis.memory.prompt_versions table" is concrete.
• Before rating IGNORE: check if the topic (not just the source) is relevant to any active skill in src/genesis/skills/. A shallow source can raise an important idea. Evaluate the idea, not the container. If the topic matters, research it and rate the underlying concept — even if the source itself is thin.

Overlap Comparison

{INCLUDE ONLY when Genesis has a comparable capability. OMIT entirely when there is no Genesis equivalent. Minimum 3 rows.}

Dimension	Their approach	Our approach	Gap
...	...	...	...

{1-2 sentences synthesizing the table: where we're genuinely ahead, where we're behind, and what the actionable delta is.}

How It Helps

{Direct applicability, ready-to-use tools, validated patterns. For a concrete external tool, identify its distinctive mechanism and the least-owning viable reuse path.}

How It Doesn't Help

{Incompatibilities, misalignment, maturity concerns. Distinguish a true veto from integration work and compare complete lifecycle cost.}

How It COULD Help

{Patterns worth stealing, future version ideas, creative applications. Think beyond "adopt this tool" — consider incremental improvements to how we already do something, upgrades to existing approaches, better measurement of something we currently vibes-check, or architectural patterns that would make an existing subsystem more rigorous.}

What to Learn

{Engineering patterns, competitive positioning, design principles.

When Genesis has something comparable, the Overlap Comparison table above IS your primary evidence for this lens — synthesize what the table reveals about our implementation quality. The question is never "do we have something that resembles this?" It's "are we doing this well enough to get the benefits it promises?"

Examples of what "gap" looks like in practice:

• "We have prompts" vs "we have versioned prompts with outcome linkage"
• "We have task tracking" vs "we have verified completion rate metrics"
• "We have memory" vs "we have continuous quality scoring with regression tracking"
• "We have approval gates" vs "we have pass-state gating where the harness verifies independently"

Surface the gap between having a feature and having it work at the level of rigor the reference describes.}

References

• docs/architecture/genesis-v3-vision.md — Core philosophy for fit assessment
• docs/architecture/genesis-v3-autonomous-behavior-design.md — System design
• docs/architecture/CURRENT.md — the live subsystem map (what Genesis actually has today, with per-entry freshness stamps). Consult this for any "does Genesis already have X?" judgment BEFORE claiming a gap or an overlap.
• Enumerate, don't spot-check. Before concluding Genesis "lacks X" or is "weaker at X", verify by enumeration against CURRENT.md and the actual code — a negative from one search is not evidence of absence. Confidence is capped by how completely you enumerated.