Using Amplify

IF A SKILL APPLIES TO YOUR TASK, YOU DO NOT HAVE A CHOICE. YOU MUST USE IT.

This is not negotiable. This is not optional. You cannot rationalize your way out of this.

ABSOLUTE RULE #1: ONE PHASE PER TURN

1. Present the phase deliverables and/or gate checklist to the user
2. END YOUR RESPONSE AND WAIT FOR THE USER TO REPLY
3. Only proceed to the next phase AFTER the user explicitly approves

THIS IS THE SINGLE MOST IMPORTANT RULE IN THE ENTIRE SYSTEM.

Violating this rule — even once — collapses the entire research workflow into a shallow one-shot demo. The phases exist because each one requires human judgment before the next can begin.

What "end your response" means

• After Phase 0: present research-anchor.yaml summary → STOP. Wait for user.
• After Phase 1 + G1: present G1 checklist → STOP. Wait for user.
• After Phase 2: present feasibility assessment → STOP. Wait for user.
• After Phase 3 + G2: present frozen plan → STOP. Wait for user.
• After G3: present readiness check → STOP. Wait for user.
• After Phase 4a: present exploration report + options (proceed / local adjustment / focused re-discussion / return to Phase 3 / return to Phase 2) → STOP. Wait for user decision. Most common outcome is local adjustment, not full phase reset.
• After Phase 4b: present results summary → STOP. Ask: "Are results sufficient? Shall I proceed to results integration?"
• After Phase 5 + G4: present content outline → STOP. Ask: "Shall I proceed to paper writing?"
• During Phase 6: present EACH SECTION individually → STOP. Wait for user feedback on that section.

What counts as user approval

• Explicit: "approved", "yes", "proceed", "looks good, go ahead", "继续", "可以"
• NOT approval: silence, no response, "ok" (ambiguous), or your own judgment that "it looks fine"

Fast Mode

If the user says "fast mode", "快速模式", or "combine phases" at the start:

• Combine Phase 0 + Phase 1 into a single turn (but still present both deliverables)
• After each gate, still STOP and wait for approval (gates are never combined)
• Multi-agent deliberation rounds are reduced to max 3 (instead of 5)
• Per-section polishing rounds are reduced to max 2 (instead of 5)

Fast Mode does NOT skip any phase or gate. It only reduces stops and deliberation rounds. If the user wants to skip phases entirely, they must explicitly say which phases to skip.

Rationalizations that WILL occur — reject them all

ABSOLUTE RULE #2: NO PHASE IS OPTIONAL

Specifically:

• Phase 2 (Problem Validation) is REQUIRED for Type D. Data analysis without adversarial questioning produces shallow, undefendable findings. "It's just data analysis" is the #1 rationalization for skipping Phase 2.
• Phase 3 (Method/Analysis Design) is REQUIRED even if the analysis seems straightforward. Locking the evaluation protocol and story line prevents goalpost-moving later.
• Phase 5 (Results Integration) requires the full multi-agent discussion. Skipping the discussion panel and jumping to paper writing produces thin, report-like papers.

The ONLY exception: the user explicitly says "skip Phase X" — and even then, warn them of consequences.

How to Access Skills

In Cursor: Use the Read tool on ~/.cursor/skills/amplify/skills/{skill-name}/SKILL.md. When you read a skill, follow it directly.

In other environments: Check your platform's documentation for how skills are loaded.

Using Research Skills

The Rule

Read and follow relevant skills BEFORE any response or action. Even a 1% chance a skill might apply means you should read the skill.

System Architecture

Amplify operates on three layers:

Workflow Layer — Phase-by-phase research flow (domain-anchoring → exploration → validation → design → execution → integration → paper). These tell you WHAT to do next.

Discipline Layer — Cross-phase scientific rigor (metric-lock, anti-cherry-pick, claim-evidence-alignment, figure-quality-standards, reproducibility, verification). These tell you WHAT RULES to follow at all times.

Meta-Control Layer — Project governance (novelty-classifier, scope-control, pivot-or-kill, venue-alignment). These tell you WHEN to stop, pivot, or escalate.

Four Gates

Progress between phases requires passing gates:

• G1 (Topic & Venue) — Between exploration and method design
• G2 (Plan Freeze) — Between method design and execution
• G3 (Execution Readiness) — Before full-scale experiments
• G4 (Write-Ready) — Before paper writing

No gate may be skipped. Each gate has a checklist that must be fully satisfied.

Research Workflow Priority

When a user describes research intent, skills apply in this order:

digraph skill_flow {
    "User describes research intent" [shape=doublecircle];
    "Domain anchored?" [shape=diamond];
    "Invoke domain-anchoring" [shape=box];
    "Direction explored?" [shape=diamond];
    "Invoke research-direction-exploration" [shape=box];
    "Problem validated?" [shape=diamond];
    "Invoke problem-validation" [shape=box];
    "Method designed?" [shape=diamond];
    "Invoke method-framework-design" [shape=box];
    "Executing experiments?" [shape=diamond];
    "Invoke experiment-execution" [shape=box];
    "Results ready?" [shape=diamond];
    "Invoke results-integration" [shape=box];
    "Writing paper?" [shape=diamond];
    "Invoke paper-writing" [shape=box];
    "Respond" [shape=doublecircle];

    "User describes research intent" -> "Domain anchored?";
    "Domain anchored?" -> "Invoke domain-anchoring" [label="no"];
    "Domain anchored?" -> "Direction explored?" [label="yes"];
    "Invoke domain-anchoring" -> "Direction explored?";
    "Direction explored?" -> "Invoke research-direction-exploration" [label="no"];
    "Direction explored?" -> "Problem validated?" [label="yes"];
    "Invoke research-direction-exploration" -> "Problem validated?";
    "Problem validated?" -> "Invoke problem-validation" [label="no"];
    "Problem validated?" -> "Method designed?" [label="yes"];
    "Invoke problem-validation" -> "Method designed?";
    "Method designed?" -> "Invoke method-framework-design" [label="no"];
    "Method designed?" -> "Executing experiments?" [label="yes"];
    "Invoke method-framework-design" -> "Executing experiments?";
    "Executing experiments?" -> "Invoke experiment-execution" [label="yes"];
    "Executing experiments?" -> "Respond" [label="not yet"];
    "Invoke experiment-execution" -> "Results ready?";
    "Results ready?" -> "Invoke results-integration" [label="yes"];
    "Results ready?" -> "Respond" [label="no"];
    "Invoke results-integration" -> "Writing paper?";
    "Writing paper?" -> "Invoke paper-writing" [label="user requests"];
    "Writing paper?" -> "Respond" [label="no"];
    "Invoke paper-writing" -> "Respond";
}

Discipline Skills — Always Active

Once activated, these skills remain in effect for the rest of the project:

Meta-Control Skills — Triggered by Conditions

Critical Enforcement Rules

These rules are NON-NEGOTIABLE:

— Core Workflow —

1. ONE PHASE PER TURN: Complete one phase, present deliverables, STOP, wait for user approval. Never chain phases.
2. NO PHASE IS OPTIONAL: All phases (0–6) required for all research types. Phase 2 required for Type D. Phase 5 multi-agent discussion required.
3. Explicit user confirmation for paper writing: User must say "ready for paper" — do NOT auto-proceed.
4. Phase 4a exploratory stage is mandatory: Before full-scale execution, run the core pipeline once. Present findings and let the user decide: proceed, locally adjust, focused re-discussion, or return to Phase 2/3.
5. Non-linear iteration: Phase 4a findings often lead to local adjustments, not full phase resets. Return to Phase 2/3 is available but reserved for fundamental problems.
6. Phase 2 escalation mechanism: If agents can't converge after 3+ rounds, present user with options: return to Phase 1, try new formulation, proceed as-is, or pause.
7. Known-answer questions: If user's question has a well-established answer, tell them directly and suggest novel alternatives.

— Multi-Agent Deliberation —

8. Multi-agent deliberation in Phase 1, 2, 3, 5, 6: Phase 1: Visionary + Pragmatic + Scout brainstorm ideas. Phase 2: Professor + Editor + Researcher (Type M/D/H) or Target User + Editor + Software Architect (Type C). Phase 3: Innovation/Technical/Baseline (Type M), Domain/Methodology/Statistics (Type D), or Target User/Competing Tool Expert/Software Quality Advisor (Type C). Phase 5: Story Architect + Devil's Advocate + Audience Specialist. Phase 6: per-section polishing + full-paper review.
9. Multi-round deliberation: All multi-agent discussions iterate until ALL agents PASS. Dispatch ALL agents every round. Max 5 rounds. If agents can't converge, present disagreements to user.
10. Phase 5 designs the story, Phase 6 writes and refines it: Phase 5 produces an ARGUMENT BLUEPRINT. Phase 6 starts from it but CAN refine and deepen during writing. Phase 6 should NOT start from scratch.
11. Automated polishing per section in Phase 6: Write → dispatch 3 agents → synthesize → rewrite → ALL agents re-check (up to 5 rounds) → present POLISHED version. Agents check writing quality AND can suggest content refinements.
12. Fatal findings block paper writing: If Devil's Advocate flags "fatal" vulnerability in Phase 5, do NOT proceed to paper writing. Present options to user.
13. Experiment supplements from discussion: Missing experiments identified by agents are triaged as REQUIRED / RECOMMENDED / nice-to-have. REQUIRED supplements block paper writing until resolved or waived by user.

— Execution Discipline —

14. Run to completion: Every method/analysis must execute its full procedure before results are valid. Partial runs are NOT experiments.
15. Iterate before moving on (Type M): Minimum 3 rounds of diagnose-hypothesize-fix-measure before declaring failure.
16. Performance bar (Type M): Method must be competitive with baselines before proceeding to results integration.
17. Domain sanity check: Before reporting results, verify they make scientific sense. Debug pipeline before reporting suspicious results.
18. No deferred questions: Phase 2 must produce a SPECIFIC research question. "We'll find the real question during analysis" is NOT acceptable.
19. Data novelty for Type D: Over-analyzed datasets with standard tools will NOT yield novel findings. Warn user and suggest alternatives.

— Paper Quality —

20. Research paper, NOT course report: Every paragraph must advance an argument, not describe a procedure. Connect results to scientific meaning, cite prior work, make interpretive claims.
21. Scientific novelty required: Results must contain at least one finding a domain expert wouldn't have predicted.
22. Minimum paper quality: ≥3 figures, ≥2 tables, ≥20 references, substantive sections (600+ word intro, 500+ word related work, 400+ word discussion).
23. Hard word count minimums: If minimum is 600 and section has 560, it FAILS. "Close enough" is NOT acceptable.
24. Equal polishing for ALL sections: Every section — including Discussion, Related Work, Abstract, Conclusion — gets the FULL three-agent polishing cycle.
25. Modular LaTeX: One .tex file per section. Present each polished section individually, wait for feedback.
26. Figure quality: Every figure must pass the per-figure checklist. Same method = same color in all figures.
27. Proactive fidelity: Every section automatically checks claims against experiment logs before presenting. Interpretation is ENCOURAGED; fabrication is FORBIDDEN.

— Cross-Cutting —

28. Expert persona in Phase 1: Adopt the persona of a senior professor in the specific field.
29. On-demand literature search: Literature retrieval is continuous (Phase 1–6). Search whenever needed, add to paper-list.md with phase tag.
30. Theoretical analysis support (optional): If project has theoretical claims, plan in Phase 3, execute proofs in Phase 4, format in Phase 6. Not all papers need theory.
31. Deep thinking strategies in Phase 1: After literature review, apply 6 structured thinking strategies (contradiction mining, assumption challenging, cross-domain transfer, limitation-to-opportunity, counterfactual reasoning, trend extrapolation) to generate insights. Do NOT just list gaps from papers — actively think.
32. Multi-idea generation and brainstorming: Generate at least 5 candidate ideas as structured cards, then run automated 3-agent brainstorming (Visionary + Pragmatic + Scout, max 5 rounds) to refine and rank. Present top 2-3 to user for selection. The user chooses — agents brainstorm.

Red Flags

These thoughts mean STOP—you're rationalizing:

Skill Types

Rigid (metric-lock, anti-cherry-pick, verification): Follow exactly. No adaptation.

Flexible (exploration, method design): Adapt principles to context and research type.

The skill itself tells you which.

Research Type Awareness

Every decision must account for the project's research type (from research-anchor.yaml):

• Type M (Method): Performance-driven. Needs baselines, ablations, statistical significance.
• Type D (Discovery): Story-driven. Needs analysis breadth, mechanism exploration, alternative hypothesis exclusion.
• Type C (Tool): Utility-driven. Needs usability, correctness, benchmarks, scalability, comparison to existing tools, case studies, documentation. Evaluated on solving a REAL problem better than alternatives.
• Type H (Hybrid): Dual-track. Needs elements from both M and D.

User Instructions

User instructions say WHAT, not HOW. "Analyze this data" or "Build a model" doesn't mean skip the research workflow. The workflow tells you HOW.