Skip to main content

Third-party skills gap analysis for the agent fleet

Date: 2026-10-07 · Access date for all citations: 2026-10-07 Prepared by the fleet research service: planned and verified by the lead researcher (maverick-muse-lead_researcher-001); executed by the research assistant (maverick-muse-research_assistant-001) using deep research against the live docsite index, the upstream repositories, and the public skills registries named below. Read-only throughout: nothing was installed, imported, or executed.

Question. The owner: "We have a lot of skills installed and I want to make sure we're using them effectively. I want to utilize some of the third-party skills like no-ai-slop and ponytail. Dispatch my research team to identify any skills we don't have available, but should, and get them into docsite and the agent runbook configurations." A mid-mission amendment from the owner added agile project management skills, aimed mainly at the scrum master (§6, rows 16–28).

Boundary. This report delivers the vetted list only. Import into the docsite and the runbook configurations happens after the chief's (and where needed the owner's) approval of the list; the chief routes that work to the sweeper. The research service imports nothing.

1. Bottom line​

The owner-named skills are already in the library, the assignment's starting sweep was wrong, as the lead's intake correction suspected and this mission verified from the live index: skill-no-ai-slop and a complete six-page ponytail family are in docsite's 429-skill index. Their real gaps sit in completeness, currency, and wiring: the no-ai-slop page's process references a companion eval.md that was never imported; the ponytail import is an untracked snapshot (upstream DietrichGebert/ponytail is at v4.13.0 (2026-10-05) and its recent releases have already changed /ponytail-gain's basis); the sibling writing skill humanizer is materially stale (fleet 2.5.1 vs upstream 3.1.0); and none of the sampled non-research role runbooks name any skills at all, no-ai-slop and ponytail sit in manifests that the chief/worker/reviewer runbooks never reference. Beyond the named families, the true acquisition gaps are few and specific: Anthropic's official frontend-design (absent; Apache-2.0, no scripts), a dedicated WCAG/accessibility audit method (absent, AccessLint/skills is the credible candidate, with license and execution caveats), and the systemic ui_ux coverage hole (49 creative-category pages, zero ui_ux tags, no ui_ux manifest). Superpowers (15/15) and the anthropics/skills set (17/19 in some form) are already essentially fully held, the fleet's problem there is duplication (eight skills are held twice, under both the anthropic and the claude categories) and wiring, not acquisition.

Agile (Amendment 1). The owner added agile project skills to the mission mid-flight, aimed at the scrum master. The internal scrum set is an operations set of 18 skills (retro, handoffs, to-tickets, triage, status reporting), and the RUNBOOKS.md pointer to a "planning & agile method set" names a set that does not exist as a set in the docsite. Against the fleet's settled method (sprints close on scope completion, never on a timebox; the in-repo llm-kanban board; Fibonacci points; two-stage intake; owner acceptance), exactly two external candidates clear the bar, both adapted: plinth's agile story trio and three plays from tronghieu's scrum-master set, plus a zero-cost wire-in (planning-and-task-breakdown into the scrum manifest, where the chief's disposition and the manifest on main currently disagree). Section 6 gives the per-sub-area evidence; rows 16–28 of the acquisition list carry the calls.

2. Baseline inventory + capability coverage map​

Baseline (verified live, docsite main): skills index docs/15-skills/skills-index.json blob 5f5777e794, 429 entries, the lead's intake numbers reproduce exactly. Index role tags: worker 185, chief_of_staff 126, planner 82, code_quality_reviewer 44, sweeper 36, security_reviewer 26, research_assistant 18, scrum_master 15, lead_researcher 8, finance 2, all 3. Per-role manifests on main (10 files) run +3 vs index tags (the three all skills): worker 188, chief 129, planner 85, CQR 47, sweeper 39, security_reviewer 29, research_assistant 21, scrum_master 18, lead_researcher 11, finance 5. No ui_ux manifest exists and the string ui_ux appears in zero entries' roles (index query: any role == ui_ux → 0 hits). Workspace skills on this host (ls ~/workspace/skills): exactly the five named, cloudflare, cloudflare-dedicated, fleet-sync, github-docsite, llm-kanban. Bundled platform catalog appendix: ~50 product skills (gmail, calendar, spotify, shopping, plaid, etc.) ride with the platform, a different layer from the docsite method library; full enumeration in my working notes on request, none are fleet method skills, all out of this mission's scope.

Coverage map (index-derived; method stated so counts are auditable):

AreaEvidence from the indexVerdict
Writing quality / anti-slopskill-no-ai-slop + skill-humanizer, both tagged chief_of_staff only; 33 purpose-text hits for writing/editing terms. Division of ground: humanizer = pattern catalogue for removing AI tells (Wikipedia-derived, 34 patterns as imported); no-ai-slop = an editor's process (preserve voice, minimum edit, detect-vs-edit modes, self-check vs eval.md). They overlap in aim, differ in method, but only the chief is tagged for either; worker/planner/research roles that also write have no tagged anti-slop skillPresent, narrowly tagged, one stale
Coding discipline / minimalismPonytail family (6) + skill-clean-code, skill-code-simplification, skill-simplify-code, skill-coding-behavior-rules, skill-lean-build, skill-minimalist-skillStrong
Code review16 ids containing "review" (incl. skill-code-review, skill-code-review-and-quality, skill-codebase-review, superpowers requesting/receiving)Strong
Planningskill-writing-plans, skill-executing-plans, skill-planning-and-task-breakdown, skill-plan; planner has 82 tagsStrong
Design / UI-UXcreative category = 49 pages; 17 ids containing "design"; strong exemplars: skill-ui-ux-pro-max, skill-design, skill-design-system, skill-emil-design-eng, skill-impeccable, skill-redesign-skill. BUT roles are worker/CQR/planner, ui_ux: 0 tags, no manifest, and (see §5) the ui_ux runbook names no skillsContent strong, role coverage ZERO, the map's biggest hole
Accessibility0 ids matching accessib*/wcag*; "accessibility" appears in the purpose text of only 5 entries (frontend-ui-engineering, react-patterns, react-testing, ui-styling, ui-ux-pro-max, incidental, incl. ui-ux-pro-max's searchable data). No WCAG audit method anywhereGAP, acquire
Testing13 ids containing "test" (TDD ×2 incl. an Osmani variant, webapp-testing ×2 duplicate, e2e, browser-devtools, language-specific)Adequate; duplicative in places
Security review5 ids containing "secur" + catalog-driven security testing, security-and-hardening; security_reviewer 26 tagsAdequate
Research methodresearch category 16 pages; 15 ids containing "research"; research_assistant manifest 21, lead 11Strong for research roles
Memory / context8 ids containing "context" + skill-memory-systems, skill-codebase-memory, context-engineering family, mostly chief-taggedAdequate
Agile project method (Amendment 1)Scrum manifest = 18 operations skills; vocabulary probe of all 429 entries: "sprint" 0, "groom" 0, "standup" 0, "velocity" 0, "agile" 0, "user stor" 0; planning-and-task-breakdown tagged [planner] onlyThin by design in places (estimation settled by charter; DoR/DoD enforced by the llm-kanban schema); genuine gaps in story craft and impediment handling, see §6
Documentation qualitydoc-coauthoring held twice (anthropic + claude duplicates), skill-documentation-and-adrs, changelog-generator; no dedicated technical-writing-quality skill beyond coauthoringAdequate, duplicative; thin at the quality end

Duplication datum: docs/15-skills/anthropic/ and docs/15-skills/claude/ are the same 8 skills imported twice (algorithmic-art, canvas-design, claude-api, doc-coauthoring, internal-comms, mcp-builder, web-artifacts-builder, webapp-testing, anthropic-* and claude-* ids, verified by index lookup). The claude category holds three further skills (code, design, handoff) that are not duplicates. ~8 of 429 index slots are double-counts.

3. Named-skills dossiers​

3a. no-ai-slop, HELD; incomplete import; roles too narrow​

  • Docsite page (fetched in full): docs/15-skills/creative/no-ai-slop.md, blob 3b63bfa40d, 11,307 chars, a full skill text (edit + detect modes, editing principles, pattern list, process), NOT a summary. Footer attribution: source jknash/hermes-shared-skills branch hermes-jkdev001 @1d0d545c3970, imported 2026-10-04; "Supporting files (references, scripts) remain in the source repository."
  • Completeness gap (verified on the page): its Process step 4 says "check the edited draft against eval.md", that companion file is NOT in docsite (the footer confirms supporting files stayed behind). The self-check half of the method is unusable as imported.
  • True upstream (verified at repo): https://github.com/petergyang/no-ai-slop, Peter Yang, MIT License, ~12,004★, created 2026-07-07, 7 releases to v1.0.6; the skill lives at skills/no-ai-slop/{SKILL.md, eval.md, agents/} upstream, i.e. upstream ships exactly the eval.md the docsite copy lacks. Repo also carries scripts/, .codex-plugin, build_plugin.py (build/validate tooling), PRIVACY.md/TERMS.md, none of that needs importing; the skill + eval.md do. Whether the docsite text matches upstream v1.0.6 line-for-line was NOT diffed (gap, §8).
  • Role coverage: tagged chief_of_staff only (index + chief manifest). Plausible users who are untagged: planner, lead_researcher/research_assistant (long-form reports), worker (PR prose). Runbook wiring: none (§5).
  • Confidence: Settled (presence, page completeness, eval.md absence, upstream identity/license/version); Supported (role-extension judgement).

3b. Ponytail family, HELD, family COMPLETE vs upstream; currency untracked; executable ecosystem correctly NOT imported​

  • Docsite pages (all fetched in full): skill-ponytail (6,196 chars, blob 8e794238ac), ponytail-review (2,398), ponytail-audit (1,734), ponytail-debt (1,807), ponytail-gain (1,989), ponytail-help (3,058), full skill texts with the same internal-mirror footer (hermes-shared-skills @1d0d545c3970, imported 2026-10-04); ponytail main page carries "license MIT.". Roles: ponytail/gain/help → worker; review/audit → code_quality_reviewer; debt → CQR + sweeper.
  • Upstream (verified at repo): https://github.com/DietrichGebert/ponytail, Dietrich Gebert, MIT License (LICENSE file + sidebar), ~157k★, 311 commits, 22 releases, latest v4.13.0 (released 2026-10-05), repo created 2026-06-12, active Oct 2026. Upstream skills/ = exactly the same 6 + plugin.json, the fleet's family is complete; upstream has no seventh skill.
  • Currency: upstream moves fast and has changed around the skills: recent releases number review/audit findings and changes /ponytail-gain to agentic benchmarks, the fleet's gain page still describes the older published-medians scoreboard, so at least that page is behind upstream. The docsite import is a snapshot of an internal mirror, not tracked to any upstream release; exact per-page diff vs the current upstream release not performed (gap, §8).
  • Trust note (important): upstream is no longer markdown-only, it ships JS hooks (activate/config/mode-tracker/subagent), shell/PowerShell statusline scripts, a publish script, __init__.py, a pi-extension, and plugin manifests for ~20 agents. The docsite import took only the markdown skill texts. Recommend keeping it that way: the executable/plugin layer is a deliberate skip (see acquisition list).
  • Role coverage: worker/CQR (+sweeper for debt) is the right core; runbook wiring: none (§5). Confidence: Settled.

3c. Humanizer (writing-division comparator), HELD but STALE​

  • Docsite: docs/15-skills/creative/humanizer.md, blob 8318cefd68, 34,621 chars, roles [chief_of_staff]; page itself records: original author Siqi Chen (@blader), repo https://github.com/blader/humanizer version 2.5.1, MIT, Hermes additions = patterns 30–34 + a clichés list.
  • Upstream now: version 3.1.0 (release dated 28 Sep), ~54,533★, 97 commits, patterns rebuilt across 2.x→3.x (33→35→25→26) and a no-fabrication rule added in 2.9.0. The fleet copy is materially behind; an update must re-apply the Hermes additions on top of 3.1.0, not overwrite blindly. Upstream ships one script (scripts/validate-package.py, packaging validation only).

4. Ecosystem survey (each read live, access 2026-10-07)​

  • agentskills.io, the Agent Skills open standard home (originated by Anthropic; developed via agentskills/agentskills + Discord). Spec: folder + SKILL.md (required name/description; optional license/compatibility/metadata/allowed-tools; scripts/references/assets allowed; progressive disclosure). It runs a client showcase (~40 clients), NOT a skills directory, and states no skill count, a standard to conform to, not a source to import from.
  • anthropics/skills, Anthropic's official repo, ~180k★, updated 2026-10-05, no releases; no single repo license, many skills Apache-2.0, document skills (docx/pdf/pptx/xlsx) source-available, per-skill LICENSE.txt governs. 19 skills; index lookup shows the fleet already holds 17 of 19 in some form. Absent: frontend-design (the gap that matters), academy-guide (Anthropic-Academy-specific) and discernment-nudge (internal-facing). All 9 probed names (frontend-design, webapp-testing, doc-coauthoring, skill-creator, web-artifacts-builder, mcp-builder, internal-comms, canvas-design, algorithmic-art) are present upstream.
  • skills.sh, Vercel Labs' Agent Skills Directory, paired with the vercel-labs/skills CLI (MIT, ~33k★); leaderboard ranked by installs. Signals: find-skills #1 (3.7M installs), frontend-design 959K, skill-creator 399K, superpowers' brainstorming 383K / systematic-debugging 281K / writing-plans 266K, i.e. the ecosystem's centre of gravity matches skills the fleet mostly already holds.
  • GitHub Topics, agent-skills: 29,861 repos (top: anthropics/skills 180k, ponytail 157k, addyosmani/agent-skills 102k, nexu-io/open-design 100k, ComposioHQ 76k); claude-skills: ~10,254 repos. Discovery aids, star counts measure fame, not fit; every candidate below was verified at its repo.
  • ComposioHQ/awesome-claude-skills, Apache-2.0 badge (no root LICENSE file in the listing; per-skill caveat), ~76,643★, created 2025-10-17, claims 1000+ skills/plugins, ~22 bundled folders. The fleet already holds a 25-page composio category (incl. content-research-writer, lead-research-assistant), so this is a known, partially-imported source.
  • obra/superpowers, Jesse Vincent (obra) + Prime Radiant, MIT, ~296k★, 683 commits, 14 releases, latest v6.4.2 (2026-09-25). 15 skills, index lookup: all 15 are already in the docsite index (brainstorming, executing-plans, writing-plans, systematic-debugging, test-driven-development, requesting/receiving-code-review, verification-before-completion, dispatching-parallel-agents, subagent-driven-development, using-git-worktrees, finishing-a-development-branch, using-superpowers, writing-skills; and skill-diagnosing-bugs as the counterpart of upstream diagnosing-superpowers). Trust note: the repo ships hooks/scripts and its brainstorming visual companion loads a Prime Radiant logo remotely (telemetry, opt-out via SUPERPOWERS_DISABLE_TELEMETRY), again, markdown-text import only.
  • Others verified: VoltAgent/awesome-agent-skills (MIT, ~35k★, actively curated, claims 1000+), a good watch-list source; addyosmani/agent-skills (~102k★), the fleet already has a sweeper-tagged cron skill skill-cron-sync-addyosmani-agent-skills-230068, i.e. a sync relationship already exists; travisvn/awesome-claude-skills (~15k★, NO license shown, README stale since Nov 2025), skip as a source; hesreallyhim/awesome-claude-code (~55k★, license Other/NOASSERTION, broader than skills), watch only.

5. Utilization findings (Q6), the wiring gap is the finding​

Sampled live from jknash/agent-orchestration main: chief-of-staff.md (blob 8d6b7fd66d), worker.md (e2e24c1e05), code-quality-reviewer.md (9582617321), ui-ux.md (0796aea20b), planner.md (b034f27037), security-reviewer.md (0403637b72), lead-researcher.md (c64076858b), research-assistant.md (5da9930208). Greps run on the fetched texts (auditable):

  • ponytail in all 8 runbooks: 0 hits. no-ai-slop / ai-slop / humanizer: 0 hits.
  • The word "skill" outside the two research runbooks appears only once, in an unrelated sense (CQR's "a manifest of files and digests"). Only lead-researcher.md and research-assistant.md have an Artifacts "Skills of the role" section listing their manifest skills; chief, worker, CQR, ui_ux, planner, and security runbooks name no skills at all, they rely on bootstraps pointing at the skills library in general.
  • Divergence, stated plainly: index tags say no-ai-slop→chief and ponytail→worker/CQR; manifests agree (both skills are in those roles' manifest files); runbooks say nothing. An agent working strictly from its runbook would never discover either named skill. The ui_ux case is triple: the desk's runbook names no skills, the index tags it zero skills, and no ui_ux manifest exists for the AF-46 machinery to hand it.
  • Secondary: the index's ~8 anthropic/claude duplicate slots (§2) mean manifest counts overstate distinct capability by ~2%.

6. Agile project management skills (Amendment 1, added 2026-10-07)​

The owner added this area mid-mission, aimed mainly at the scrum master (planner and chief of staff secondary). Every candidate below was run through one fit test: the fleet's method closes sprints when committed scope completes and never on a timebox, runs back-to-back including weekends, boards work on the in-repo llm-kanban, sizes stories in Fibonacci points under a two-stage backlog-to-story intake, and reserves acceptance to the owner. A candidate earned acquire only if it strengthens that method. Candidates built on timeboxed cadence or on Jira/Linear ceremony were marked acquire-adapted (adaptation named) or skip (conflict named).

Baseline, as found​

  • Scrum manifest on docsite main (docs/15-skills/manifests/scrum_master.json, blob b0cb84e235): exactly 18 entries, skill-retro, skill-handoff, skill-claude-handoff, skill-to-tickets, skill-triage, skill-github-issues, skill-meeting-action-items, skill-sdlc-review, skill-verification-before-completion, skill-changelog-generator, skill-fleet-build-status-reporting, skill-filesystem-context, skill-using-agent-skills, skill-using-superpowers, skill-setup-matt-pocock-skills, skill-contract-registry-work-package-remediation, skill-live-delivery-reconciliation, skill-repo-and-tracker-consolidation. Character: fleet-operations skills (handoffs, triage, tickets, board/status reporting) plus one ceremony skill (retro). It is an operations set, not an agile-method set.
  • Discrepancy, named: skill-planning-and-task-breakdown (docs/15-skills/engineering/planning-and-task-breakdown.md) is tagged [planner] only in the live index and is absent from the scrum manifest on main, notwithstanding the chief's disposition this morning adding it to the scrum set. State as found: disposition recorded, manifest not yet regenerated. Wire-in row 16 below carries the fix.
  • Vocabulary probe (redone properly; query = term searched across id+title+purpose of all 429 index entries): "sprint" 0, "groom" 0, "standup" 0, "velocity" 0, "agile" 0, "definition-of" 0, "scrum" 0, "user stor" 0. "retro" 2, skill-retro plus one incidental purpose-text hit (skill-banner-design). "estimat" 4, skill-planning-and-task-breakdown is the only method-shaped hit; the rest are incidental. "backlog" 2, both incidental (contract-registry remediation; a launcher-recovery skill). The lead's probe reproduces exactly.
  • RUNBOOKS.md pointer, resolved: ~/workspace/fleet-agents/RUNBOOKS.md (scrum master table) points "Sprint / retro method" at "docsite docs/15-skills/ planning & agile method set". No such set exists as a named grouping. A full docsite git-tree scan finds zero paths containing "agile" under docs/15-skills and only two containing "planning" (engineering/planning-and-task-breakdown.md, productivity/weekly-review-planning.md); there is no planning or agile category folder among the 15-skills subdirectories. What the pointer can honestly resolve to is three scattered singles: skill-retro (ceremony), skill-planning-and-task-breakdown (task breakdown, planner-tagged), skill-weekly-review-planning (a review cadence, not sprint method). Judgement: the pointer overstates the depth behind it, it names a "method set" where the library holds, at most, one directly on-point skill (retro) and two adjacents. Recommend the pointer be corrected or the set be built by the wire-in/acquire rows below; that decision is the chief's.

Coverage by sub-area​

Sub-areaInternal state (index/manifest evidence)External supply (verified at repos)Verdict
Sprint planningNo method skill. 0 "sprint" hits in 429 entries. The scrum 18 cover board mechanics around a sprint (to-tickets, triage, status reporting), not how to plan/commit one; the method actually lives in fleet charters + the lead's/chief's practice, not in a skillCandidates exist but are built on the rejected premise: phuryn sprint-plan takes duration "[2 weeks / 1 week / etc.]" and sizes by historical velocity over the last 3 sprints; anthropics/knowledge-work-plugins sprint-planning asks sprint length; borghei/alirezarezvani assume cadences + capacityGap real; nothing clears the bar unadapted. The fleet's sprint-planning act (propose committed scope, chief approves, close on completion) has no external counterpart worth importing
Story writing & splittingThin. skill-to-tickets (intake mechanics), skill-planning-and-task-breakdown (task breakdown, mentions estimating scope, not story craft, not INVEST, no splitting method)The strongest external cluster: deanpeters user-story / user-story-splitting / user-story-mapping / epic-breakdown-advisor (quality, 3.9K installs for user-story), but CC BY-NC-SA 4.0; jabrena/plinth 012-agile-epic / 013-agile-feature / 014-agile-user-story (Apache-2.0, INVEST check, Gherkin output)One acquire-adapted clears the bar: plinth (row 17). The deanpeters set is better-known and is skipped on license alone
EstimationSettled internally by charter, not by skill: Fibonacci points + the estimate-at-pull-in rule (S05 retro DQ-01 disposition). planning-and-task-breakdown is the only estimat-* method hitEssentially no supply. skills.sh "estimation" returns no agile estimation skill; "story points" returns only xgent-ai/story-points (113 installs) + a 19-install Jira one. No maintained planning-poker/estimation skill exists to acquireNothing clears the bar, because there is nothing to clear it with, and the internal rule already answers the question. A finding, not a failure
Backlog grooming / refinementOwned by fleet process, not a skill: two-stage backlog-to-story intake is codified in the llm-kanban board contract + charters; "backlog" index hits are incidental; "groom" 0 hitsNo substantive backlog-refinement method skill found in any source swept (skills.sh "backlog" → no method skill)Nothing clears the bar. External supply is absent; the fleet's intake contract is the method
RetrospectivesCovered. skill-retro is IN the scrum 18. Depth judgement: the fleet's retro practice is real and evidenced, sprint retros are published artifacts whose action items become backlog stories (Justin's methodology rule; S05/S07 retros on record). skill-retro carries the ceremony; the fleet's follow-through discipline around it is stronger than the skill textphuryn retro and BMAD's bmad-retrospective (6.8K installs) offer alternative templates; BMAD's license is flagged NOASSERTION + TRADEMARK.md at GitHub despite an MIT claim in its READMECovered internally; no acquisition. External templates add format, not method. BMAD = watch (row 24) on license clarity alone
Standup / flow facilitationReplaced, deliberately, by fleet machinery: skill-fleet-build-status-reporting is in the scrum 18; the hourly fleet update + board-as-record is the fleet's flow mechanism, asynchronous, artifact-based, no daily human ceremonyanthropics/knowledge-work-plugins standup (5.0K installs, Apache-2.0) generates standup notes for human teams; skills.sh "standup" is dominated by tool automationsNothing clears the bar, conflict named: a standup-note skill serves a ceremony this fleet abolished on purpose. Skip (row 23)
Definition of Ready / DoneEncoded operationally, not as a skill: the llm-kanban story schema (required fields, acceptance criteria, claim-by-commit protocol) + stakeholder-acceptance gate ARE the fleet's DoR/DoD, enforced by tooling (kanban.py validate) rather than by checklist proseOnly micro-supply: three "definition of done" micro-skills on skills.sh at ≤80 installs; addyosmani/agent-skills carries a pack-level definition-of-done.md checklist (not a skill)Nothing clears the bar. The fleet's enforcement-by-schema is stronger than the checklist skills on offer
Velocity / forecastingUndefined by design. "velocity" = 0 index hits, and that is correct for this method: with no fixed timebox there is no denominator, so points-per-sprint-period is not a fleet metric; the fleet forecasts by committed scope vs completion (sprints close at e.g. 13/13) and reports throughput through board stateThe forecasting candidates all forecast the missing metric: borghei's scrum-master does Monte Carlo velocity forecasting (50/70/85/95% confidence) over per-member capacity; phuryn sizes by trailing velocity; igta-tech (index-only, unverified) also forecasts velocityNothing clears the bar, on evidence: every external forecasting skill assumes a timeboxed cadence and human capacity, the two things this method does not have. The honest fleet-shaped metric (stories/points completed per calendar week across back-to-back sprints) is derivable from the fleet's own board history; no external skill computes it. Reported as a finding, as instructed, not pre-decided

Ecosystem sweep for agile​

  • skills.sh, no term returned zero results (fuzzy search), but the near-negatives are the signal: "estimation" → no agile estimation skill; "backlog" → no method skill; "story points" → only xgent-ai/story-points (113 installs) + a 19-install Jira skill; "definition of done" → 3 micro-skills ≤80 installs; "standup" → tool automations dominate. Top scrum hit: alirezarezvani's scrum-master (795 installs), Jira-wired (row 21).
  • GitHub repo/topic searches ("scrum-master" SKILL.md is NOT a negative search, it surfaces borghei, alirezarezvani, tronghieu, vampayr, igta-tech and more; topic agent-skills+scrum/agile surfaced nothing beyond the candidates in §A4). Registry mirrors (majiayu000, learn-skills.dev) are copies, not independent sources.
  • ComposioHQ/awesome-claude-skills, checked: its project-management material is tool integrations (Jira/Linear/Asana automations), no agile method skill.
  • VoltAgent/awesome-agent-skills, DOES index agile entries (its own business-product/scrum-master; deanpeters story set; phuryn retro/sprint-plan/user-stories), a discovery path, each verified at its own repo below.
  • anthropics/skills (19 skills), none is an agile/scrum method skill (verified, as expected).
  • anthropics/knowledge-work-plugins (distinct repo, 26,919★, Apache-2.0), sprint-planning (2.9K installs) + standup (5.0K): standalone-capable, network only via optional MCP connectors, and both fail the fit test on method assumptions (row 23).
  • obra/superpowers (15 skills), none is an agile method skill; writing-plans/executing-plans are per-task process skills (already held, main draft §4).
  • addyosmani/agent-skills (102,401★, MIT, 25 skills), no agile skill; its planning-and-task-breakdown sizes by T-shirt/file-count (a variant of the skill the fleet already holds), and its pack includes a definition-of-done.md checklist (a pack document, not a skill).
  • mattpocock/skills (278,909★, MIT), to-tickets / triage / handoff confirmed (the fleet's scrum 18 already carry all three via setup-matt-pocock-skills lineage); no estimation, story-splitting, sprint-planning, or sprint retrospective, his "retro" is an agent-environment retrospective, not a sprint retro. Tracker-agnostic, which is why his set fits this fleet where others don't.
  • BMAD-METHOD (bmad-code-org, 53,889★), a full agent framework with bmad-retrospective (6.8K installs) and an SM agent persona; adjacent, not a drop-in skill set; license flagged NOASSERTION by GitHub with a TRADEMARK.md despite an MIT claim in the README, unverified for vendoring until resolved.

The amendment's question, answered​

What external agile skills offer that the internal 18 (plus planning-and-task-breakdown) lack is narrower than the sub-area list suggests. The genuine items are (i) story craft, an INVEST-checked method for writing and splitting stories, which the internal set only touches through ticket mechanics and task breakdown; (ii) an impediments/health-scan discipline with named plays, where the fleet currently relies on escalation habit; and (iii) a close-sprint play written for sprints that end by finishing, a concept only tronghieu's set even names. Everything else on the amendment's list is either already owned (retrospectives, via skill-retro plus a proven retro-to-backlog pipeline), owned in a stronger form (Definition of Ready/Done, enforced by the llm-kanban schema and validator rather than by checklist prose; status/flow, replaced by board-plus-hourly reporting), settled by charter where the ecosystem has no supply at all (estimation, Fibonacci plus estimate-at-pull-in, against an ecosystem with no maintained estimation skill to buy), or defined away by the method itself (velocity forecasting, whose denominator this fleet deliberately does not have). Of the candidates verified, exactly two clear the bar, both adapted: plinth's agile story trio (row 17, Apache-2.0, format retargeted to llm-kanban) and tronghieu's impediments/health-scan/close-sprint plays (row 18, MIT, script and cadence language stripped), plus one wire-in that costs nothing: planning-and-task-breakdown into the scrum manifest (row 16), where the chief's disposition and the manifest on main currently disagree.

7. Acquisition list​

Format: numbered rows in the packet's column order, name | source URL (verified live) | what it does | fleet roles served | license & trust notes | overlap (existing docsite ids) | disposition + reason.

  1. frontend-design | https://github.com/anthropics/skills/tree/main/skills/frontend-design | Anthropic's official design skill: distinctive, non-templated UI design, a two-pass plan (palette/type/layout/principles reviewed against the brief) then code, named AI-default clusters to avoid, and a quality floor incl. focus/reduced-motion. Not a WCAG audit skill | ui_ux (primary), worker | Apache-2.0 (per-skill LICENSE.txt); maintainer Anthropic, repo updated 2026-10-05; SKILL.md + LICENSE.txt ONLY, no scripts, no network | NONE in the index (lookup "frontend-design" → 0 exact ids; adjacent: skill-frontend-patterns, skill-frontend-ui-engineering, skill-imagegen-frontend-web, skill-ui-ux-pro-max, method design guidance is the missing piece, implementation patterns are held) | ACQUIRE, the one true official-skill hole; fills the ui_ux desk's flagship method need.
  2. AccessLint accessibility skills (scan/inspect/audit/fix/diff, WCAG 2.2, AA default) | https://github.com/AccessLint/skills | Dedicated WCAG accessibility scanning/auditing method for web UI, the capability the index lacks entirely | ui_ux, code_quality_reviewer | README asserts MIT but no LICENSE file at repo root, obtain the license text before vendoring; ships HEAVY execution (npx @accesslint/cli, bundled @accesslint/mcp server, Chrome/CDP auto-launch, optional hosted connector) | NONE (lookup accessib*/wcag* → 0 ids; 5 purpose-text mentions are incidental) | ACQUIRE-ADAPTED, import the markdown audit method only; gate or drop the CLI/MCP/hosted parts pending a security review.
  3. no-ai-slop, WIRE-IN + COMPLETE. Already held (skill-no-ai-slop, chief). Import the missing upstream eval.md (petergyang/no-ai-slop, MIT, v1.0.6), diff the page text vs v1.0.6, wire into the chief runbook's Artifacts, consider extending tags to planner + lead_researcher.
  4. ponytail family (6), WIRE-IN + CURRENCY CHECK. Already held, family complete vs upstream (DietrichGebert/ponytail, MIT, v4.13.0). Diff pages vs v4.13.0 (ponytail-gain at minimum is behind), wire ponytail into worker runbook and ponytail-review/audit/debt into CQR runbook Artifacts. Do NOT import upstream hooks/scripts/plugins (skip that layer, JS hooks, shell/PowerShell scripts, publish tooling).
  5. humanizer, UPDATE (wire-in family). Already held at 2.5.1; upstream blader/humanizer at 3.1.0 (MIT). Update with Hermes additions (patterns 30–34) re-applied; wire into chief runbook alongside no-ai-slop, with the division of ground documented (humanizer = pattern catalogue, no-ai-slop = editor's process + eval).
  6. ui_ux role coverage pass, WIRE-IN (systemic). No new skill: tag the existing design exemplars (skill-ui-ux-pro-max, skill-design, skill-design-system, skill-emil-design-eng, skill-impeccable, skill-redesign-skill, skill-apple-design, frontend-design once acquired) with ui_ux, extend the roles vocabulary/manifest generator, generate ui_ux.json, and give ui-ux.md an Artifacts skills section like the research runbooks have.
  7. Superpowers method set, WIRE-IN. Already held 15/15 (skill-systematic-debugging, skill-test-driven-development, skill-writing-plans, skill-executing-plans, skill-requesting-code-review, skill-receiving-code-review, skill-verification-before-completion, etc.; obra/superpowers, MIT, v6.4.2). Wire the named ones into worker/CQR runbook Artifacts, they are tagged and manifested but, like ponytail, unreferenced where the work happens.
  8. anthropic/claude duplicate sets, WIRE-IN (consolidate). 8 skills held twice (skill-anthropic-* ≡ skill-claude-*). Recommend the sweeper consolidate to one canonical copy per skill and regenerate index/manifests; frees ~8 slots and ends double-maintenance.
  9. find-skills (vercel-labs/skills) | https://github.com/vercel-labs/skills/tree/main/skills/find-skills | Skill discovery: search skills.sh, install via the skills CLI | chief_of_staff, sweeper (hypothetically) | MIT; SKILL.md only, but its workflow is network/CLI by design (skills.sh leaderboard + npx skills find/add, i.e. it installs things) | No discovery skill in the index (lookup "find-skills" → 0); fleet-side, skill-agent-skill-bundle-governance + the sweeper's sync skills cover adjacent ground | WATCH, the fleet's sweeper sync machinery already covers discovery; do not wire auto-install behaviour in. Revisit if the chief wants a sanctioned discovery front-end.
  10. AccessLint hosted MCP connector (mcp.accesslint.com), SKIP. External hosted service in the appraisal path of Leg A2's family; the adapted markdown method stands without it.
  11. 84emllc/claude-wcag-skill | https://github.com/84emllc/claude-wcag-skill | Alternative single-skill WCAG audit | ui_ux, code_quality_reviewer (hypothetically) | SKILL.md claimed MIT but repo license is mixed Other/NOASSERTION with W3C-licensed material; 9 commits, 0★; ships scripts (build-reference.sh curls W3C, validate.mjs) | Same zero-overlap as row 2 | WATCH, immature and license-messier; AccessLint (row 2) is the stronger candidate if the license text resolves.
  12. mlwelles/technical-writing-density | https://github.com/mlwelles/technical-writing-density | Technical-writing density/quality editing in a single SKILL.md | chief_of_staff, planner (hypothetically) | MIT; single SKILL.md, no scripts in the skill (an install.sh exists in the repo) | Adjacent: skill-anthropic-doc-coauthoring / skill-claude-doc-coauthoring (held twice), skill-documentation-and-adrs | SKIP, 1★/4 commits is too thin to carry the documentation-quality gap; revisit only if it matures.
  13. anthropics academy-guide + discernment-nudge | https://github.com/anthropics/skills | Anthropic-product onboarding (Academy) and an internal-facing discernment nudge | none (fleet method roles) | Per-skill LICENSE.txt in an Apache-2.0/source-available mixed repo | Not in index (verified) | SKIP, product/internal-facing, not fleet method skills; out of the mission's scope.
  14. travisvn/awesome-claude-skills | https://github.com/travisvn/awesome-claude-skills | Curated Claude-skills list | (source, not a skill) | No license shown; README updates stop Nov 2025, stale | Overlaps ComposioHQ/VoltAgent lists the fleet already mines | SKIP as a source.
  15. VoltAgent/awesome-agent-skills + addyosmani/agent-skills | https://github.com/VoltAgent/awesome-agent-skills ; https://github.com/addyosmani/agent-skills | Large curated skill collections (VoltAgent claims 1000+; Osmani's is an engineering-skills repo, ~102k★) | all roles, per-skill | VoltAgent MIT; Osmani license not verified this mission (gap §8.3) | Fleet already holds a 25-page composio category, the full superpowers set, and skill-cron-sync-addyosmani-agent-skills-230068 (sweeper) evidencing an existing Osmani sync | WATCH as sources, mine per-skill later against named gaps rather than bulk-importing.

Agile rows (Amendment 1)​

  1. skill-planning-and-task-breakdown → scrum_master | source: live index, docs/15-skills/engineering/planning-and-task-breakdown.md | Breaks a spec/requirements set into ordered, estimable, implementable tasks | scrum_master (new), planner (held) | Internal fleet-imported page (Hermes import lineage, as the other 429); no new trust surface, a retag + manifest regeneration only | Already held as skill-planning-and-task-breakdown, tagged [planner] only; absent from scrum_master.json on main despite the chief's disposition | WIRE-IN, regenerate the scrum manifest with it included (or record why the disposition differs); also the only estimat-* method skill the fleet owns, so the scrum desk should actually hold it.
  2. plinth agile story trio, 012-agile-epic / 013-agile-feature / 014-agile-user-story | https://github.com/jabrena/plinth | Interactive Q&A that produces an epic/feature/user story with an INVEST check and Gherkin acceptance criteria | scrum_master, planner | Apache-2.0 (quoted at repo); maintainer jabrena, 445★, topic agent-skills; the skills invoke the terminal date command, trivial, no network, no credentials, no other executables reported | No story-craft skill held (nearest: skill-to-tickets = intake mechanics; skill-planning-and-task-breakdown = task breakdown) | ACQUIRE-ADAPTED, the one agile candidate that clears the bar. Adaptation, named exactly: (i) retarget the output from standalone .feature files to the llm-kanban story format (frontmatter + acceptance criteria in the board's schema, Gherkin kept inside the AC section where it helps); (ii) drop any cadence language; the INVEST check and the splitting dialogue transfer unchanged and strengthen the two-stage intake's story stage.
  3. tronghieu/agent-skills, scrum-master (selected plays) | https://github.com/tronghieu/agent-skills | Scrum-master play set: plan-sprint, pulse, close-sprint, retro, impediments, health-scan, coach; metrics must trace to a tracker/export/stated assumption | scrum_master | MIT (LICENSE at repo); 72★, 97 commits, v1.0.0; ships scripts/init-scrum.sh (setup script, do not import); no tool hard-wiring found; hybrid stance, agent may be Scrum Master of record or copilot | Nearest held: skill-retro (retro play overlaps), skill-fleet-build-status-reporting (pulse-adjacent); the fleet has NO named impediments or health-scan method | ACQUIRE-ADAPTED, import only the impediments, health-scan, and close-sprint plays: close-sprint maps directly onto this fleet's scope-completion closing (no other candidate even has the concept), and impediment tracking is a genuine internal gap. Adaptation: strip the init script and any fixed-cadence wording from plan-sprint (not imported); pulse is not imported (fleet status reporting covers it).
  4. vampayr/scrum | https://github.com/vampayr/scrum | An agent-native scrum design: agents fill PO/SM/Developer roles, the human is stakeholder, state in .scrum/ markdown, evidence-gated Definition of Done, throughput + DoD pass-rate instead of velocity | scrum_master, chief_of_staff (design reference) | MIT; 2★, created 2026-07-28, negligible adoption; 5 optional stdlib Python scripts, no network | No overlap by import; strong conceptual overlap with the fleet's own settled method (which predates and out-evidences it) | WATCH, the closest design relative in the ecosystem (sprints "usually hours rather than weeks"; stakeholder = human), and precisely because the fleet already HAS this method, importing a 2-star rival framework would create a second source of truth. Its evidence-gated DoD phrasing is the one transferable idea; the chief or sweeper can note it in the llm-kanban runbook without any import.
  5. borghei/Claude-Skills, scrum-master | https://github.com/borghei/Claude-Skills | Sprint health scoring, retro analysis, per-member capacity, Monte Carlo velocity forecasting (50/70/85/95%) | scrum_master (hypothetical) | LICENSE HAZARD: "MIT + Commons Clause" (GitHub: NOASSERTION), Commons Clause bars selling/repackaging and makes the text non-open for vendoring purposes; 875★, 8 releases, skill v2.0.1 (2026-06-15); ships 4 stdlib Python tools, no network | Method overlap with skill-retro + fleet status reporting | SKIP, two independent reasons, either sufficient: the Commons Clause license, and the method core (per-member human capacity + velocity forecasting) assumes exactly the timeboxed human team this fleet is not.
  6. alirezarezvani/claude-skills, scrum-master | https://github.com/alirezarezvani/claude-skills | Scrum-master skill in a 27,805★, MIT, 6-PM-skill collection; skills.sh's top scrum hit (795 installs) | scrum_master (hypothetical) | MIT (clean), the problem is architecture, not license: hard-wired to Jira + Confluence via the bundled Atlassian Remote MCP (https://mcp.atlassian.com/v1/sse), live network + OAuth; scripts folder present | skill-github-issues / skill-to-tickets cover the fleet's actual tracker (in-repo board + GitHub) | SKIP, conflict named: this fleet's board is the in-repo llm-kanban; a Jira/Confluence-MCP-dependent skill cannot run its method here, and wiring a hosted Atlassian MCP for it is a non-starter.
  7. phuryn/pm-skills, sprint-plan / retro / user-stories | https://github.com/phuryn/pm-skills | PM skill set: sprint planning by trailing velocity, retrospectives, user-story writing | scrum_master, planner (hypothetical) | MIT; 26,817★, real maintenance and adoption | retro overlaps skill-retro (held); user-stories overlaps the story-craft space row 17 fills with a cleaner license-to-format fit | WATCH, sprint-plan is a hard fit-test fail (sprint duration input + historical-velocity capacity + 15–20% buffer = textbook timebox planning), so the set cannot be acquired whole; its retro/user-stories components are the only transferable parts and both are covered (skill-retro; row 17). Revisit only if row 17 disappoints in use.
  8. anthropics/knowledge-work-plugins, sprint-planning + standup | https://github.com/anthropics/knowledge-work-plugins | Official-plugin sprint planning (asks sprint length; capacity in points/hours) and standup-note generation | scrum_master (hypothetical) | Apache-2.0 (clean); 26,919★; network only via optional MCP connectors | skill-fleet-build-status-reporting + skill-retro cover the fleet's actual flow/retro practice | SKIP, conflicts named: sprint-planning's first question is the timebox this method abolished, and a standup-note generator serves a daily human ceremony the fleet replaced with asynchronous board + hourly status reporting. Clean license, wrong method.
  9. BMAD-METHOD, bmad-retrospective / bmad-agent-sm | https://github.com/bmad-code-org/BMAD-METHOD | Full agent-delivery framework incl. a retrospective skill (6.8K installs) and a Scrum-Master persona agent | scrum_master (hypothetical) | LICENSE UNRESOLVED: README claims MIT; GitHub flags the repo NOASSERTION and a TRADEMARK.md exists, do not rely on either until verified at the repo's license files | skill-retro (held) covers the retro ceremony | WATCH, license clarity is the gate, and it is also a competing framework, not a skill: nothing piecemeal until the license question is settled and the chief wants a framework comparison.
  10. deanpeters/Product-Manager-Skills, user-story family | https://github.com/deanpeters/Product-Manager-Skills | user-story (3.9K installs), user-story-splitting, user-story-mapping, epic-breakdown-advisor, the best-known story-craft set in the sweep | scrum_master, planner (hypothetical) | LICENSE HAZARD: CC BY-NC-SA 4.0, non-commercial + share-alike: vendoring its text into the docsite library would encumber the library and misstate the fleet's usage terms; 7,186★ | Same story-craft space as row 17 (plinth, Apache-2.0) | SKIP, on license alone, despite quality. Row 17 carries the same capability with a clean license, which is why this trade is recorded.
  11. djalmajr/skills, agile- set (14 skills)* | https://github.com/djalmajr/skills | An intake→sprint→review→retro agile pipeline of 14 skills | scrum_master (hypothetical) | No LICENSE found at repo root, nothing to quote, nothing to vendor on; 4★, 232 commits | Broad overlap with the scrum 18's mechanics | SKIP, no license, and a direct method conflict: the set deliberately refuses story-point sizing ("Do not reference size codes") where this fleet's method is Fibonacci points by charter.
  12. jwynia/agent-skills, agile-workflow / agile-coordinator / github-agile | https://github.com/jwynia/agent-skills | Multi-agent agile coordination over git + the gh CLI | scrum_master (hypothetical) | No LICENSE found; 165★; stale ~225 days; requires git + gh CLI at runtime | skill-fleet-build-status-reporting + the fleet's own dispatch machinery cover coordination | SKIP, unlicensed and stale; the fleet's coordination method is its own.
  13. skills.sh micro-skills (xgent-ai story-points; definition-of-done micros) | https://skills.sh | Single-purpose estimation / DoD checklist micro-skills (≤113 installs; DoD set ≤80) | scrum_master (hypothetical) | Licenses not verified at repos (install counts this low do not justify the pass) | Estimation and DoR/DoD are settled internally (Fibonacci + estimate-at-pull-in; llm-kanban schema + validate) | SKIP, negligible adoption against settled internal practice; listed to show the estimation/DoD shelves were actually looked at, not overlooked.

8. Source list (selection; full list in the working research report)​

9. Open gaps​

  1. No line-by-line diff was run between docsite skill texts and upstream current versions (no-ai-slop vs petergyang v1.0.6; ponytail pages vs DietrichGebert v4.13.0; superpowers pages vs v6.4.2). Settles by: a diff pass at import time, a sweeper task once the list is approved.
  2. AccessLint's license: README asserts MIT but no LICENSE file was found at repo root; the actual license text must be obtained/verified before any vendoring.
  3. addyosmani/agent-skills: its license and agile inventory were settled during the Amendment 1 sweep (MIT, verified at the repo by the lead; 25 skills, none of them an agile method skill). Its per-skill inventory beyond agile was not checked.
  4. Exact last-commit dates for ComposioHQ, VoltAgent, AccessLint, vercel-labs/skills were not verified (activity assessed from releases/stars/README state).
  5. The bundled platform catalog (~50) was enumerated from my own available-skills context as a draft appendix per the plan, not independently audited; nothing in the acquisition list depends on it.
  6. Whether role runbooks should enumerate skills (vs. bootstraps pointing at manifests) is a design judgement for the chief, §5 reports the divergence and the research desks' Artifacts pattern as the working precedent; it does not decide the convention.
  7. Three index-only candidates were not opened at their repos (j4flmao agile-scrum-kanban; kyqurikan github-scrum-flow; igta-tech scrum-master-agent, the last is integration-heavy with velocity forecasting, so its fit-test outlook is poor on its face). If the chief wants the shelf exhaustive, a short verification pass settles them; nothing in §A5 depends on them.
  8. Exact last-commit dates for several candidate repos were not displayed by the fetching route; activity is reported via stars/commits/releases and dated signals, none invented.
  9. BMAD's license (MIT claim vs NOASSERTION flag + TRADEMARK.md) is unresolved, a hard gate before any future reliance, stated in row 24.
  10. AccessLint-style depth check was not repeated for agile repos' script folders beyond what their pages list (e.g. alirezarezvani's scripts/ rendered partially); the dispositions that turn on scripts (rows 18, 21) rest on the documented init/MCP facts, which are sufficient for the calls made.

10. Lead verification, and the edit pass on this report​

Verification ran against the plan, not against the draft's own claims. The lead re-fetched the docsite index and the scrum manifest from main, recounted every index-derived figure in §§2 and 6 (role tags, the review/test/context/security/research/design id counts, the superpowers 15, the vocabulary probe), and checked the upstream facts at the GitHub API for every repo a disposition turns on: licenses (petergyang/no-ai-slop MIT; DietrichGebert/ponytail MIT; blader/humanizer MIT; jabrena/plinth Apache-2.0; tronghieu/agent-skills MIT; AccessLint/skills with no license file at its root; borghei/Claude-Skills and bmad-code-org/BMAD-METHOD both flagged NOASSERTION; deanpeters under CC BY-NC-SA 4.0; djalmajr/skills and jwynia/agent-skills with no license at all), releases (ponytail v4.13.0 of 2026-10-05; humanizer v3.1.0; no-ai-slop v1.0.6; superpowers v6.4.2), and the contents of Anthropic's frontend-design folder (SKILL.md and LICENSE.txt only). Two runbooks (worker, chief of staff) were re-fetched and grepped to confirm §5. Corrections the lead carried into this final: the ponytail version is stated as v4.13.0, one release past the draft's read; the duplication finding counts eight shared skills, with the claude category's three extra skills (code, design, handoff) named as non-duplicates; and the addyosmani license question the first draft left open is closed (MIT, settled during the agile sweep and confirmed at the API).

What changed in the no-ai-slop edit pass (first use, pilot). The verified text went through the skill's workflow: full read, minimum effective edit, self-check, fix, re-check. The draft used the em dash as its default separator (160 instances); the pass replaced them with commas, sentence breaks, or parentheses as each sentence required, and cut the pattern-level faults the scan found (one binary-contrast sentence in §1, three colon reveals in prose, zero banned words outside quotation). A mechanical scan of the published text for the skill's banned list returns exactly one hit, a banned word inside the owner's verbatim question in the front matter. That word stays. It is a quote, and quotes are evidence, not prose. One limit of the pass, stated plainly: the skill's workflow checks a draft against a companion file, eval.md, which the fleet does not hold (§3a and row 3). The self-check here ran against the skill page's own rules, pattern list, and banned-word list instead.

Published by Muse · 2026-10-07.