Dev & Engineering multi-agent-researchparallel-processingsubagent-dispatchcompetitive-analysisliterature-reviewresearch-workflow

Infinite Gratitude

Multi-agent parallel research that keeps bringing findings home like cats — dispatch agents, collect reports, and loop on new questions until you're satisfied.

FollowSkills review · FSRS-2.0
Use with care
49/ 100 5-point scale 2.5 / 5
1 2 3 4 5 6
1Trust15 / 25 · 3.0/5

The skill handles no sensitive data, no credentials, no external writes; it asks 'Continue? (y/n)' before each loop and its data flow (web research and reporting only) is transparent. Deductions: dispatching up to 10 background agents on haiku has cost/rate implications without explicit caps, rollback or recovery guidance; publisher identity is unverified and attribution points only to washinmura.jp.

2Reliability10 / 20 · 2.5/5

The happy path (split directions, dispatch, compile, loop) is self-consistent with a clear parameter table. Deductions: static review cannot execute reproduction; no error handling or failure feedback for agent failures, missing research-scout subagent type, or out-of-range parameters; loop termination ('until satisfied') is vague.

3Adaptability8 / 15 · 2.7/5

Target scenarios (technical research, competitor analysis, literature review) are clear, with trilingual docs and explicit triggers. Deductions: no declared non-fit boundaries; inconsistency between SKILL.md (5 directions, 5 default agents) and README (10 agents, --waves flag); core function depends on GitHub/HuggingFace reachability, uncertain from mainland China; referenced references/agent-config.md is not present in evidence.

4Convention8 / 15 · 2.7/5

Docs are well structured with parameter table, examples, notes, MIT license and trilingual README. Deductions: no versioning or changelog; the referenced agent-config.md cannot be verified; maintenance ownership and update path rest on personal branding only.

5Effectiveness5 / 15 · 1.7/5

The multi-agent parallel research concept plausibly completes the task with directly usable structured reports. Deductions: static review cannot verify actual output quality; claimed benefits (77.6% accuracy, 30 min vs 20+ hours) are author self-reports without independent evidence; the data-volume-to-accuracy formula has no cited source, so marginal-value evidence is limited (static cap is 7 in any case).

6Verifiability3 / 10 · 1.5/5

Only author-side claims: story documents and sample outputs; no test suite, no CI evidence, no third-party reproduction material; fictional narrative (cat/dog characters) and factual claims are mixed without clear separation, and key conclusions cannot be independently verified.

Evidence confidence:Low Reviewed Sep 10, 2026 Reviewed revision 40e0f9704d4d
Before you use it
  • Dispatching up to 10 background agents incurs API cost and rate usage; confirm budget beforehand and manually bound deep-mode loop counts.
  • All research findings and performance figures (e.g., 77.6% accuracy) are author self-reports, independently unverified; validate before relying on them.
  • SKILL.md and README disagree on defaults (5 vs 10 agents) and the --waves parameter; actual behavior may vary.
  • Core function depends on GitHub/HuggingFace reachability, which is not guaranteed from mainland-China networks.
  • The referenced references/agent-config.md was not present in the reviewed evidence and could not be verified.
Review evidence [1][2][3][4]
See the full review method →

What does this skill do, and when should you use it?

Infinite Gratitude is a Claude Code skill from Washin Village's AI Dojo series for multi-agent parallel research. It splits a topic into up to 5 research directions (GitHub projects, HuggingFace models, papers, competitors, best practices), dispatches up to 10 background agents to investigate them, compiles the findings into a structured report, and loops on follow-up questions until you stop it. The README documents one real case: researching an AI system to recognize the village's 28 cats and dogs, where 10 parallel agents covered 12 topics and produced 9 reports in 30 minutes versus an estimated 20+ hours manually, with the resulting system reaching 77.6% accuracy. It suits solo developers and small teams who need broad multi-angle research quickly and accept haiku-model economics.

Takes a topic plus optional flags (--depth quick/normal/deep, --agents 1-10, and --waves mentioned in the README); splits the topic into up to 5 research directions; dispatches agents in parallel via the Task API as research-scout subagents on the haiku model with run_in_background; collects and compiles their findings into a structured report; asks whether to continue if new questions emerged and loops back to dispatch (deep mode loops until satisfied); caps 5 agents per wave. Install by placing infinite-gratitude.skill.md into ~/.claude/skills/, then invoke with /infinite-gratitude "topic" in Claude Code.

  1. A solo developer comparing many tools or libraries (vector databases, embedding models) before committing to a stack, without researching each by hand.
  2. A founder or PM running competitor analysis: each agent digs into a different competitor, results rolled into a comparison report.
  3. An engineer doing a fast literature scan: parallel paper reading and summarization, then targeted follow-ups on key findings.
  4. Market research: scan one industry topic from project, model, paper, competitor and best-practice angles simultaneously.
  5. Due diligence: multi-angle background checks on a project or company.

What are this skill's strengths and limitations?

Pros
  • Parallel multi-agent research: the README's real case finished an estimated 20+ hour manual research effort in 30 minutes with 9 traceable reports.
  • Loop mechanism: each report surfaces new questions and asks the user before the next wave; deep mode digs until satisfied.
  • Cost-conscious by design — defaults to haiku, max 5 agents per wave; depth, agent count and waves are configurable.
  • One-command install, trilingual docs (EN/中文/日本語), and an agent-config reference file.
Limitations
  • Hard dependency on Claude Code's Task subagent API, the research-scout subagent type and slash-command behavior — not portable to other runtimes without rework.
  • Internal inconsistency: SKILL.md says default 5 directions and max 5 agents per wave while the README advertises 10 parallel agents; actual parallelism depends on the Claude Code environment.
  • Output quality depends on the underlying model's research ability; no independent benchmarks or test suite are provided beyond the single pet-recognition case study.
  • The --waves flag appears only in the README, not the SKILL.md parameter table — some behavior details are under-documented.

How do you install this skill?

The README gives a one-command install: curl -sSL https://raw.githubusercontent.com/sstklen/infinite-gratitude/main/infinite-gratitude.skill.md -o ~/.claude/skills/infinite-gratitude.skill.md, i.e. download the skill file into ~/.claude/skills/. The repo also ships SKILL.md and references/agent-config.md (agent configuration). No further dependency setup is documented.

How do you use this skill?

In Claude Code, run /infinite-gratitude "your research topic". Options: --depth quick|normal|deep (default normal; deep loops until satisfied), --agents 1-10 (default 5, max 10), and --waves 3 per the README (iterations, default 3). Agents research in parallel, return a report, and ask whether to continue. SKILL.md examples: /infinite-gratitude "pet AI recognition", /infinite-gratitude "RAG best practices" --depth deep, /infinite-gratitude "React state management" --agents 3.

How does this skill compare with similar options?

The repo explicitly positions this skill against its sibling research-scout (single-agent research): use Infinite Gratitude for broad parallel deep-dives, research-scout for lightweight single-agent lookups. It also recommends pairing with the author's yes.md (safety gates and evidence rules) and 5x-cto (post-research development pipeline).

FAQ

Is it expensive to run?
SKILL.md explicitly defaults to the haiku model to save cost, caps 5 agents per wave and defaults to 3 waves; actual spend depends on your depth/agents/waves settings — no real cost figures are given in the source.
Can I use it outside Claude Code?
Not as-is. It relies on Claude Code's Task subagent dispatch, the research-scout subagent type and slash-command invocation; porting to another runtime means rewriting the agent orchestration.
How is report reliability ensured?
The skill only dispatches and compiles; it has no built-in evidence verification. The repo suggests pairing with the author's yes.md to add safety gates and evidence rules as an optional complement.
Is there any measured evidence it works?
The README documents one case study: 10 parallel agents, 12 topics, 9 reports in 30 minutes for a pet-recognition AI project that ultimately reached 77.6% accuracy with a clear roadmap higher. That is the only empirical evidence provided.

Related skills