Skip to content

ideas-appendix.md

Latest commit

 

History

History
65 lines (54 loc) · 3.82 KB

ideas-appendix.md

File metadata and controls

65 lines (54 loc) · 3.82 KB

Ideas appendix — ten more seeds, one paragraph each

Starters too good to leave unmentioned, too many to build. Each names the existing project to fork as a base. Difficulty in 🦞.

  1. roast-mastah (fun, L1, 🦞 — fork dungeon-mastah, drop the dice). Point it at a git repo; it reads git log --stat through one tool and lovingly roasts the commit history ("14 commits named 'fix', a Friday 17:58 force-push — bold"). The skill is a kind-comedy contract: punch at patterns, never people; end with one sincere compliment and one actually useful suggestion. Fully offline, very demoable.

  2. curb-your-enthusiasm (social good, L1, 🦞🦞 — fork qrious-citizen). The same Analyze Boston loop pointed at sidewalk/curb-ramp and accessibility-related 311 categories: "which blocks got wheelchair-access complaints and how long did closures take?" The civic-brief skill gains an equity lens section (response-time by neighborhood, with caveats).

  3. storm-ready + FEMA layer (social good, L1→L3, 🦞🦞 — extend storm-ready). OpenFEMA's keyless API (DisasterDeclarationsSummaries with $filter) adds "has this county been declared before, how often, for what" — turning the brief from weather into risk history. Natural second agent + skill if you want the Lane 3 bar.

  4. inbox-zero-coach (everyday, L1, 🦞🦞🦞). Triage an mbox/email export into reply-now / delegate / archive with drafted one-line replies. Do NOT wire live email send on day one — that's the lethal trifecta (private data + untrusted content + outbound channel) on a silver platter; the skill's Gotchas section IS the security slide.

  5. plant-doctah (everyday/fun, L1, 🦞🦞). Photo of the sad houseplant → vision-capable model (LLaVA-class local, or frontier) → diagnosis card skill: symptom, suspect, treatment this week, "am I overwatering" verdict (yes). Good first vision-tool project.

  6. bouncer-at-the-claw-b (wildcard/security, L3, 🦞🦞🦞 — the event doc's §3.2 seed). Two agents, capabilities split: one executes code only inside podman run --network=none --read-only --cap-drop=ALL, the other has search but no execution. A sandbox-discipline skill makes the separation procedural; the demo beat is a prompt-injected "exfiltrate the workspace" instruction failing architecturally. Cite ToxicSkills + OWASP ASTop10.

  7. chief-of-stuff + traces (chief-of-staff, L3, 🦞🦞 — extend the flagship). Wrap each specialist dispatch in OpenTelemetry GenAI spans, ship to a containerized Arize Phoenix next to your models, and demo the live trace waterfall — the hub-and-spoke fan-out is genuinely pretty in spans, and the per-agent cost table doubles as Model Selection evidence.

  8. commute-whisperah (everyday, L1, 🦞🦞). MBTA's v3 API (free key, generous limits) + the NWS tool from storm-ready → "leave at 8:12, Orange Line is degraded, here's the bike-vs-train call given the radar." The skill encodes the decision rules; the agent just fetches.

  9. skill-librarian (wildcard/security, L1, 🦞🦞 — fork skill-forge). Inverts the forge: scans a directory of installed skills and red-flags ToxicSkills patterns — scripts with network calls, base64 blobs, curl-pipe installs, descriptions that don't match bodies. Output: a graded inventory per the source-grading idea. The room full of freshly-installed registry skills is the live demo.

  10. translate-the-town (social good, L1→L3, 🦞🦞 — fork any brief-emitting starter). A plain-language-translation skill that re-renders civic/storm briefs at a 6th-grade reading level and in a second language, preserving every number and caveat (the gotcha: translation must not soften warnings). Composes as a second agent pass — instant Lane 3 shape.