| name | hk-scout-wiki |
| description | Run 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 "scout", "build a wiki from these", or wants outside ideas funneled into planning. Triggers on - 'scout wave', 'wiki scout', 'research these links', 'update the wiki', or /hk-scout-wiki. |
Scout wave → wiki → ideas (project-agnostic)
Three phases. Phases 1-2 are cheap (haiku); phase 3 needs judgment (sonnet
or the main conversation). The wiki is evidence; the ideas are the product.
Phase 0 — Discover the project (once per session)
- Wiki root: existing
docs/wiki/ (or wherever a wiki README already
lives). None yet → create docs/wiki/README.md with: a category table,
the entry format below, and grep/tag conventions. Never a second format.
- Project context: read the project's CLAUDE.md / roadmap / ideas file to
learn (a) what the system IS in one line, (b) the named areas/buckets/
themes work is organized under, (c) where distilled ideas belong (an
IDEAS/BRAINSTORM file, a backlog, or — if none — propose one). If the
project defines a writing tone doc, read it and obey it.
Entry format (uniform, always)
# <Title>
source: <url>
author: <who> | fetched: <yyyy-mm> | status: read | unreachable
tags: <comma, separated, lowercase-hyphenated>
**One line.** <what you'd tell a colleague>
## TL;DR
3-6 plain sentences.
## Key ideas
Self-contained bullets.
## Project hooks
Where this touches THIS project: named areas, files, contradictions worth
arguing about. Honest "nothing yet" allowed.
Phase 1 — Scout (parallel haiku agents)
- Partition sources into batches of 4-8 by theme; each theme is a wiki
category directory with its own
TOC.md. Whole-blog surveys go in
sources/ (survey = recurring theses + best posts, one line each).
- Launch ALL batches as parallel haiku general-purpose agents in ONE
message. Every prompt includes: the entry format verbatim, the fetch
list, the project-context line from Phase 0, the unreachable rule
(retry once, then write the entry with
status: unreachable — NEVER
invent content), and the order to write the category TOC.md
(- [title](file.md) — one line. tags: ...).
- No explicit link list ("trending X")? One haiku agent with WebSearch
builds the URL list first; then partition.
Phase 2 — Assemble (main conversation, minutes)
- Spot-check 2-3 entries for format drift; fix entries, not the format.
- Ensure every category TOC exists and its links resolve; write any TOC a
scout couldn't own (shared directories).
- Commit the wiki separately from the distillation.
Phase 3 — Distill (one sonnet agent, or main conversation)
Read the new entries' TL;DR + hooks (not the sources again) and append to
the project's ideas file, in that file's existing structure and tone:
- One paragraph per idea, graded honestly against where the project
stands, citing the wiki entry, naming the area it would move.
- Contradictions with the project's philosophy are wanted — state them
plainly and why the outside view might be right.
- Distillation means "what should WE do", never "what did they say".
Cost discipline
Scouts read the web, not the repo. Distillation reads the wiki, not the
web. Nothing here needs opus. ~30 links ≈ one mid-size task execution.