agentsclimarketplace

Research harness

Skill Jpk0512/ryu-next/.claude/skills/research-harness

Three-stage research pipeline (collect → synthesize → analyze) for iterating on Ryu design decisions. Use when user drops URLs/repos/sessions for research, asks to "synthesize" findings, or asks for an "options analysis" / "recommendation". Defaults to non-saving until relevance is confirmed (approval scan). Lives at $PROJECT_ROOT/harness-research/. Never include in deployable.From its SKILL.md

Install
npx -y skills add Jpk0512/ryu-next --skill research-harness

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

7.1 KB, as published. Nobody here has run it

research-harness Skill

Three-stage research pipeline for iterating on Ryu design decisions:

URL / repo / session drop
         │
    ┌────▼────┐
    │ Stage 0 │  Approval scan — score relevance, ask if below threshold
    │  Probe  │
    └────┬────┘
         │ approved (relevance ≥4 AND fit=good → auto-save; else ask)
    ┌────▼────┐
    │ Stage 1 │  Collection — fetch, frontmatter, update INDEX.md
    │ Collect │
    └────┬────┘
         │ user: "synthesize X"
    ┌────▼────┐
    │ Stage 2 │  Synthesis — multi-pass, full team flow (Opus reviewers)
    │Synthesize│
    └────┬────┘
         │ user: "analyze" / "build options" / "recommendation"
    ┌────▼────┐
    │ Stage 3 │  Analysis — multi-pass, full team flow → options.md
    │ Analyze │
    └────┬────┘
         │ user picks direction
    decisions/YYYY-MM-DD-<topic>.md  (mini-ADR)

Workspace root: $RESEARCH_WORKSPACE (default: ./harness-research)


Stage 0 — Approval Scan (required before any save)

When the user drops a URL, repo, or session reference:

  1. Run scripts/fetch.sh --probe "$URL" to pull headers/abstract (first 2000 chars).
  2. Apply the prompt at prompts/approval-scan.md to score the item.
  3. Output the verdict in this exact format:
TYPE: <article|repo|paper|x-post|session>
TOPIC: <one-liner>
RELEVANCE: <1-5>
FIT: <good|marginal|poor>
REASON: <2 sentences max>
VERDICT: <SAVE | ASK | DROP>

Auto-save threshold: RELEVANCE ≥ 4 AND FIT = good → write to collection/ immediately.

Below threshold or marginal: present verdict to user, ask "Save anyway?" Default is NOT saved.

Probe lifecycle: store probe result ephemerally in inbox/ (no frontmatter, no INDEX entry). On approval → move to collection/<type>/YYYY-MM-DD-<slug>.md. On rejection/discard → delete from inbox/.


Stage 1 — Collection (lightweight capture)

URL routing (host-based)

HostMethod
x.com, twitter.comJina: r.jina.ai/$URL with Authorization: Bearer $JINA_API_KEY, X-Return-Format: markdown
github.comWebFetch + targeted reads (ANGLE-DRIVEN — see below)
arxiv.orgmarkdown.new
everything elsemarkdown.new; fallback to Jina if result is blank or < 200 chars

GitHub repos: ASK FOR ANGLE FIRST

Before fetching anything from a GitHub repo, ask the user:

"What aspect should I profile? (architecture / specific pattern / failure modes / API surface / other)"

Do not fetch blindly. The angle drives which files to read.

Session logs

Parse JSONL from ~/.claude/projects/<encoded-path>/<session-id>.jsonl. Extract user/assistant turns relevant to the research focus; summarize into a session template.

File conventions

  • Filename: YYYY-MM-DD-<slug>.md — ISO date prefix, kebab-case, sortable.
  • Frontmatter: YAML per the template for the source type (see templates/).
  • Tags: must come from tags.md. Add new tags to tags.md in the same operation.
  • Index: always update INDEX.md after writing a collection entry (run scripts/index.sh or append row manually).

Stage 2 — Synthesis (HIGH EFFORT)

Trigger phrases: "synthesize", "we have enough research on X", "digest"

Process:

  1. Inventory: list every collection/* markdown matching the topic (by tag or explicit filter).
  2. Extract: pull key_techniques from frontmatter + body of each entry.
  3. Group: cluster similar techniques into emergent themes.
  4. Write: for each theme, 200-400 word section. Every claim cites the collection entry by relative path.
  5. Surface contradictions and trade-offs across sources.
  6. Shortlist: end with "what's worth using" — 5-10 techniques ranked by relevance × maturity.

Team flow (mandatory):

Pre-impl reviewer (Opus) + critic (Opus) — parallel
          ↓
    Sonnet writer
          ↓
Post-impl reviewer (Opus) + critic (Opus) — parallel
          ↓
    Sonnet refinement pass
          ↓
  (repeat up to 3× until both reviewers approve)

Output: $RESEARCH_WORKSPACE/synthesis/YYYY-MM-DD-<topic>-digest.md with full frontmatter (use templates/synthesis.md).

Every claim must cite a collection/<path>.md by relative path from workspace root.


Stage 3 — Analysis (HIGH EFFORT)

Trigger phrases: "analyze", "build options", "give me a recommendation"

Process:

  1. Restate constraints: from decisions log + current user ask.
  2. Distill 3-4 named options. For each:
    • How it works
    • Pros / cons
    • Fit to constraints
    • What we'd build
    • Token budget estimate
    • Risks
    • Falsifiability test ("we'd know this is wrong if…")
  3. Explicit recommendation with rationale.
  4. Cross-reference to synthesis (preferred) or collection (allowed exception) by relative path.
  5. Phased rollout plan if recommendation is non-trivial.

Team flow: same full-team multi-pass as Stage 2.

Output: $RESEARCH_WORKSPACE/analysis/YYYY-MM-DD-<topic>-options.md with full frontmatter (use templates/analysis.md).


Decisions Log

After analysis review, when the user picks a direction, write a mini-ADR to $RESEARCH_WORKSPACE/decisions/YYYY-MM-DD-<topic>.md:

---
topic: <topic-slug>
status: accepted
date: YYYY-MM-DD
---
# Decision: <what we picked>

## Context
## Options considered
(link to analysis file)
## Choice
## Rationale
- bullet 1
- bullet 2
- bullet 3
## Consequences
## Supersedes
(links if any)

Change status: accepted → status: superseded when a later decision overrides this one.


Tag Canon

All tags must come from tags.md in this skill directory. Adding a new tag requires updating tags.md in the same operation (same file write, no separate step).

  • Lowercase kebab-case only.
  • 2-5 tags per entry.

Quality Notes

  • No token-economy constraints in harness-research/ — quality over economy. Use Opus liberally for Stages 2 and 3.
  • Never include harness-research/ in ryu-next-deploy-v1 or any deployable artifact.
  • This skill is reusable: when bootstrapping a new research effort, copy .claude/skills/research-harness/ into the target project and point RESEARCH_WORKSPACE at a fresh directory.

Keep looking

Skills are one crate of 608,666. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.