$npx -y skills add deepklarity/harness-kit --skill hk-scout-wikiRun a scout wave over outside sources (links, blogs, trending repos) into a progressive-disclosure research wiki in the current project, then distill findings into the project's ideas/backlog file. Generic across projects. Use when the user shares links/blogs to research, asks to
| 1 | # Scout wave → wiki → ideas (project-agnostic) |
| 2 | |
| 3 | Three phases. Phases 1-2 are cheap (haiku); phase 3 needs judgment (sonnet |
| 4 | or the main conversation). The wiki is evidence; the ideas are the product. |
| 5 | |
| 6 | ## Phase 0 — Discover the project (once per session) |
| 7 | |
| 8 | - Wiki root: existing `docs/wiki/` (or wherever a wiki README already |
| 9 | lives). None yet → create `docs/wiki/README.md` with: a category table, |
| 10 | the entry format below, and grep/tag conventions. Never a second format. |
| 11 | - Project context: read the project's CLAUDE.md / roadmap / ideas file to |
| 12 | learn (a) what the system IS in one line, (b) the named areas/buckets/ |
| 13 | themes work is organized under, (c) where distilled ideas belong (an |
| 14 | IDEAS/BRAINSTORM file, a backlog, or — if none — propose one). If the |
| 15 | project defines a writing tone doc, read it and obey it. |
| 16 | |
| 17 | ## Entry format (uniform, always) |
| 18 | |
| 19 | ```markdown |
| 20 | # <Title> |
| 21 | source: <url> |
| 22 | author: <who> | fetched: <yyyy-mm> | status: read | unreachable |
| 23 | tags: <comma, separated, lowercase-hyphenated> |
| 24 | |
| 25 | **One line.** <what you'd tell a colleague> |
| 26 | |
| 27 | ## TL;DR |
| 28 | 3-6 plain sentences. |
| 29 | |
| 30 | ## Key ideas |
| 31 | Self-contained bullets. |
| 32 | |
| 33 | ## Project hooks |
| 34 | Where this touches THIS project: named areas, files, contradictions worth |
| 35 | arguing about. Honest "nothing yet" allowed. |
| 36 | ``` |
| 37 | |
| 38 | ## Phase 1 — Scout (parallel haiku agents) |
| 39 | |
| 40 | 1. Partition sources into batches of 4-8 by theme; each theme is a wiki |
| 41 | category directory with its own `TOC.md`. Whole-blog surveys go in |
| 42 | `sources/` (survey = recurring theses + best posts, one line each). |
| 43 | 2. Launch ALL batches as parallel haiku general-purpose agents in ONE |
| 44 | message. Every prompt includes: the entry format verbatim, the fetch |
| 45 | list, the project-context line from Phase 0, the unreachable rule |
| 46 | (retry once, then write the entry with `status: unreachable` — NEVER |
| 47 | invent content), and the order to write the category `TOC.md` |
| 48 | (`- [title](file.md) — one line. tags: ...`). |
| 49 | 3. No explicit link list ("trending X")? One haiku agent with WebSearch |
| 50 | builds the URL list first; then partition. |
| 51 | |
| 52 | ## Phase 2 — Assemble (main conversation, minutes) |
| 53 | |
| 54 | - Spot-check 2-3 entries for format drift; fix entries, not the format. |
| 55 | - Ensure every category TOC exists and its links resolve; write any TOC a |
| 56 | scout couldn't own (shared directories). |
| 57 | - Commit the wiki separately from the distillation. |
| 58 | |
| 59 | ## Phase 3 — Distill (one sonnet agent, or main conversation) |
| 60 | |
| 61 | Read the new entries' TL;DR + hooks (not the sources again) and append to |
| 62 | the project's ideas file, in that file's existing structure and tone: |
| 63 | |
| 64 | - One paragraph per idea, graded honestly against where the project |
| 65 | stands, citing the wiki entry, naming the area it would move. |
| 66 | - Contradictions with the project's philosophy are wanted — state them |
| 67 | plainly and why the outside view might be right. |
| 68 | - Distillation means "what should WE do", never "what did they say". |
| 69 | |
| 70 | ## Cost discipline |
| 71 | |
| 72 | Scouts read the web, not the repo. Distillation reads the wiki, not the |
| 73 | web. Nothing here needs opus. ~30 links ≈ one mid-size task execution. |