claude-adaptive-research

Research that adapts to YOUR projects. Autonomous loops that get smarter with every run.

Set a topic. Walk away. Come back to a quality-gated report — with findings mapped directly to your projects, role, and goals.

One command. Personalized research. Compound learning.

What it does

Most AI research workflows are manual: you ask a question, read the answer, ask another. claude-adaptive-research automates the entire loop. You give it a topic, it researches autonomously across the web, writes a structured report, scores it for quality, and adapts every finding to YOUR projects and goals.

The plugin learns your context once (a 2-minute setup) and then every report speaks directly to your work — not generic advice, but specific adaptations.

/auto-run "How do ant colony optimization patterns apply to database sharding?"

Claude researches, analyzes, writes a report, checks quality, and delivers — all without you touching the keyboard.

Quick Start

# Install the plugin
claude plugins install primeline-ai/claude-adaptive-research

Then in any Claude Code session:

/auto-run

First run triggers a guided setup:

1. See examples of what's possible (domains, presets, free-text)
2. Choose your research domains (e.g., psychology, biology, finance)
3. Quick profile interview (your projects, role, goals)
4. Done — start researching

Features

Autonomous Research Loop

Claude researches your topic independently — searching the web, analyzing sources, synthesizing findings. No babysitting required. The loop continues across multiple iterations until the report meets quality standards.

Personalized Adaptations

Every report includes an Adaptations section that maps findings to YOUR projects. A biology finding about swarm intelligence doesn't just explain the concept — it shows how it applies to your specific SaaS architecture or your open-source library.

Quality Gate

Reports are scored on 4 criteria (structure, depth, originality, findings count). Score below 50? Claude automatically improves the report before completing. No half-baked outputs.

Premium rubric (v2, optional): For stricter premium-by-default scoring, run python3 scripts/quality_gate_v2.py path/to/report.md. v2 checks 5 metrics (citation density, DSV evidence, gap disclosure, ECP section, cross-track convergence) and assigns a tier: Premium (5/5), Standard (3-4/5), Reject (<3/5). See knowledge/quality-gate-v2.md. Requires Python 3.9+.

Cross-Track Aggregation (multi-track runs)

Run multiple research tracks in parallel on the same problem (one track per angle: technical, market, prior-art, etc.). After all tracks finish, scripts/cross_track_aggregator.py detects where independent tracks converged on the same idea (Jaccard similarity over content tokens, Union-Find clustering). Convergence across independent observers is a stronger signal than a single track. See knowledge/cross-track-aggregation.md. Requires Python 3.9+.

Compound Learning

Each run makes the next one smarter. Keywords, patterns, and follow-up questions discovered during research are saved and injected into future runs. Run 1 finds keywords → Run 2 searches deeper → Run 3 connects cross-domain. This is what makes it adaptive, not just autonomous.

Compound Score

Track your research progress: total runs, findings discovered, streak days. Research becomes a habit with visible momentum.

Research Domains

Organize your research into knowledge areas. Pick from examples or create your own:

Presets

Pre-configured research strategies for common needs:

Rate Limit Resilience

Three-layer protection keeps your research running:

1. In-prompt retry — waits and retries on transient errors
2. Stop Hook — detects rate limits, pauses without losing progress
3. Watchdog — monitors sessions, resumes after cooldown

tmux Batch Mode

Run multiple research topics overnight:

# From the scripts/ directory
./scripts/start-loop.sh --preset technique-scout

Usage

Free-text research (any topic)

/auto-run "What can distributed systems learn from how mycelium networks share nutrients?"

Presets

/auto-run --preset technique-scout
/auto-run --preset cross-domain
/auto-run --preset trend-radar

Re-run setup

/auto-run --setup

Cancel a running loop

/cancel-loop

Check loop status

./scripts/start-loop.sh status

How it works

/auto-run "topic"
     |
     v
[Setup check] ─── no config? ──→ guided setup (domains + profile)
     |
     v
[Build prompt] ─── load profile + feedback context from previous runs
     |
     v
[Create loop state] ─── _autonomous/loop.state.md
     |
     v
[Research loop] ←──────────────────────────┐
     |                                      |
     v                                      |
[Claude researches] ── web search,          |
     |                  read sources,       |
     |                  analyze             |
     v                                      |
[Write report] ── _autonomous/results/      |
     |                                      |
     v                                      |
[Quality Gate] ── score >= 50? ── no ──────→┘
     |
     yes
     |
     v
[Save feedback] ── keywords + topics for next run
     |
     v
[Kairn?] ── if installed: save top findings to memory
     |
     yes
     |
     v
[<promise>DONE</promise>] ── loop ends, report ready

Output

Reports are saved to _autonomous/results/{domain}/{date}.md:

_autonomous/
  results/
    psychology/
      2026-03-30.md    ← today's research
    biology/
      2026-03-29.md
    cross-domain/
      2026-03-28.md
  config.yaml          ← your domains
  profile.yaml         ← your projects & goals

Configuration

_autonomous/config.yaml

Your research domains — created during setup, editable anytime.

_autonomous/profile.yaml

Your projects, role, and goals — used to personalize the Adaptations section in every report.

Cost Awareness

Each research loop uses multiple API calls for web search and analysis.

The plugin shows a cost reminder on first use each session.

Pro Tip: Deeper Analysis & Persistent Memory

Quantum Lens — Perfect your findings

For perfected results, run Quantum Lens on your reports. 7 cognitive lenses analyze findings from fundamentally different perspectives — catching blind spots that single-perspective research always misses. The Solution Engine then turns insights into engineered solutions with feasibility scoring.

# After auto-run completes, deepen the best findings:
/quantum-lens "analyze _autonomous/results/biology/2026-03-30.md"

Kairn — Remember across sessions

Install Kairn for persistent knowledge across sessions. When Kairn is detected as an MCP server, the plugin automatically:

• Saves top findings from each run to your knowledge graph
• Recalls relevant past findings when starting a new run
• Prevents re-discovering what you already know

Without Kairn: reports live as markdown files. With Kairn: findings become searchable memory that survives across sessions and projects.

pip install kairn-ai  # That's it — the plugin detects it automatically

Requirements

• Claude Code >= 2.1.80
• Claude model: Opus recommended (best research quality). Sonnet works well too. Haiku is not recommended (may struggle with quality gate).
• tmux (optional): Required only for persistent/batch runs
• Firecrawl MCP (optional): Recommended for web research. Without it, Claude uses built-in WebSearch/WebFetch.

License

MIT — free to use, modify, and distribute.

Credits

Built by PrimeLine AI. Extracted from a production AI orchestration system with months of daily autonomous research.

Part of the PrimeLine Ecosystem

@PrimeLineAI · primeline.cc · Free Guide