Lsm Copilot

Workflow Overview

The full workflow has 3 phases:

Phase 1: ANALYZE     →  Load data, run analysis, generate figures & CSV
Phase 2: FOLLOW-UP   →  Ask sample context, offer report, search web
Phase 3: REPORT      →  Generate publication-quality report with literature context

Phase 1 runs automatically when the user provides a file. Phase 2 triggers after analysis completes. Phase 3 only runs if the user wants a report.

Phase 1: Analyze

Step 1: Understand the Scientific Question

Before touching any code, ask:

1. What is the sample? (cells, polymer film, nanoparticles, tissue section...)
2. What do you want to measure? (size, count, distribution, colocalization, dynamics...)
3. What does the data look like? (bright objects on dark? multi-channel? time-lapse? Z-stack?)
4. What file format and how big? (LSM/CZI/LIF/TIFF/MRC, single file or batch?)

If the user provides a file path, load it first and show them what the data looks like before proposing any analysis.

Step 2: Load & Inspect Data

Use ${SKILL_DIR}/tools/file_reader.py to load and print metadata:

python3 ${SKILL_DIR}/tools/file_reader.py <path> --info

# MRC files (auto voxel from header)
python3 ${SKILL_DIR}/tools/file_reader.py image.mrc --info

Supports: .lsm (auto voxel), .czi, .lif, .tif (needs manual voxel), .mrc/.mrcs/.map/.rec/.st (auto voxel from header, Å→µm).

Show the user: dimensions, voxel size, intensity range, a representative slice. Ask if it looks right.

Step 3: Choose Analysis Approach

Known Analysis Types

Critical rule: The tools above are starting points, not final answers. If they don't work well for the user's specific data, search the web for better approaches, read papers, and write custom code.

Step 4: Run Analysis & Generate Figures

Run the chosen analysis pipeline. Generate:

• Publication-quality figures (200+ dpi, proper axis labels with units, scale bars on images)
• CSV data files with all measured quantities
• Use µm, µm³ as units — never pixels

Phase 2: Follow-Up (MANDATORY after analysis)

After generating results, ALWAYS do the following before ending your turn:

Step 5: Ask for Sample Context

Present the results, then ask:

"分析已完成。为了更好地解读结果，请提供以下信息：

1. 样本描述：这是什么样本？（例如：小鼠脑切片、细胞培养、聚合物薄膜...）
2. 实验目的：这个实验要回答什么科学问题？
3. 标记信息：各通道对应什么染色/荧光蛋白？（例如：ch1=DAPI, ch2=GFP-X, ch3=tdTomato-Y）
4. 处理条件：样本经过什么处理？（例如：CryoChem 固定、常规 PFA 固定、活体成像...）
5. 是否需要生成正式的分析报告？"

Read ${SKILL_DIR}/prompts/followup.md for the full follow-up template.

Step 6: Generate Report (if requested)

If the user wants a report:

1. Search the web first — based on the sample description, experimental context, and key findings, search for:

   • Relevant literature on the sample type and analysis method
   • Typical values for the measured quantities (e.g., "mouse cortical neuron nucleus diameter typical size")
   • The specific technique mentioned by the user (e.g., "CryoChem CLEM fluorescence preservation")
   • Any methodological papers that validate the analysis approach

2. Write the report incorporating:

   • User-provided sample context and experimental background
   • Analysis methods and parameters
   • Results with proper units and statistics
   • Literature-contextualized interpretation (compare your numbers to published values)
   • Caveats and limitations
   • Suggested next steps

Read ${SKILL_DIR}/prompts/report.md for report formatting guidelines.

GFP / Fluorescence Preservation Analysis

When to Use

This analysis is relevant when:

• User is comparing fluorescence before vs after a sample preparation method (e.g., CryoChem, clearing, fixation)
• User wants to verify that a processing step did not destroy fluorescent signals (GFP, tdTomato, etc.)
• User has multiple conditions (e.g., treated vs control, different fixation protocols)
• User mentions "fluorescence preservation", "signal retention", or "GFP intensity comparison"

What to Measure

For each condition/sample, compute:

1. Mean fluorescence intensity per object — average signal within each segmented cell/nucleus
2. Intensity distribution — histogram + KDE of per-object intensities across conditions
3. Signal-to-background ratio (SBR) — how well objects stand out from background
4. Integrated density / CTCF — total fluorescence per object (corrected for background)
5. Coefficient of variation — uniformity of signal across the population

How to Present

Generate a comparison figure with:

Statistical Tests

• 2 conditions → Mann-Whitney U test (non-parametric) or t-test (if normal)
• 3+ conditions → Kruskal-Wallis or one-way ANOVA
• Always report: n, mean ± SD, median, p-value, effect size

Single-Sample Case

If only one sample is available (no control), still report:

• Per-object intensity distribution
• SNR and CTCF statistics
• Compare measured values to published reference ranges (search the web)
• Note: "Fluorescence preservation cannot be quantitatively assessed without a paired control; however, the observed mean SNR of X and intact nuclear morphology suggest adequate signal retention."

Read ${SKILL_DIR}/prompts/fluorescence_preservation.md for the full template.

Built-in Tools

gui_threshold.py — Interactive 3D Segmentation

# LSM (auto voxel)
python3 ${SKILL_DIR}/tools/gui_threshold.py image.lsm

# TIFF (manual voxel Z Y X in µm)
python3 ${SKILL_DIR}/tools/gui_threshold.py image.tif --voxel 0.44 0.17 0.17

Pipeline: Gaussian smooth → background subtraction → Otsu threshold → 3D connected components → regionprops filtering → interactive visualization.

Output: CSV (coordinates, volumes, diameters, intensities) + PNG figures.

Other tools in ${SKILL_DIR}/tools/

• file_reader.py — Universal file loader (LSM, CZI, LIF, TIFF, MRC)
• intensity_profiler.py — Z-depth intensity analysis & attenuation correction
• coloc_analyzer.py — Basic colocalization metrics (Pearson, Manders)
• spatial_stats.py — Spatial distribution statistics (NND, Clark-Evans, size stats)
• batch_processor.py — Process multiple files

These are lightweight utilities. For complex analysis, write custom code or search for specialized libraries.

When to Search the Web

Always search when:

• User asks about a technique you're not 100% sure about
• The built-in tools don't produce good results
• User mentions a specific method or paper
• You need to recommend a library and want to check the latest version/API
• User's data type is unusual (FLIM, light-sheet, super-resolution, expansion microscopy...)
• Phase 2/3: You need to contextualize results with published reference values
• Phase 2/3: User mentions a technique name (CryoChem, CLARITY, iDISCO...) — search for the paper

Good search queries:

• "[technique] python library 2025 2026" — find latest tools
• "[sample type] [analysis type] confocal microscopy" — find domain-specific methods
• "[library name] documentation API" — check correct usage
• "[cell type] nucleus diameter typical size µm" — reference values for report
• "[technique name] fluorescence preservation" — validate against literature

Domain Knowledge

Read from ${SKILL_DIR}/knowledge/ when you need background on:

Anti-Patterns

• Do NOT run analysis without showing the user what the raw data looks like first
• Do NOT present results without asking "does this look right to you?"
• Do NOT assume one-size-fits-all parameters
• Do NOT skip units — always report in µm, µm³, etc., not pixels
• Do NOT forget voxel anisotropy (Z resolution is usually 3-10x worse than XY)
• Do NOT end your turn after analysis without running Phase 2 (follow-up questions)
• Do NOT generate a report without first searching the web for relevant context
• Do NOT claim fluorescence is "preserved" or "destroyed" without a proper control comparison or literature reference