Get the design right before the first line of code.

The mechanisms below sound engineering-heavy. From your side, here is what it actually looks like.

What you walk away with

One design.md — a plan you can print and hand to your supervisor, your department, or an ethics committee, not a pile of configuration only a programmer can read. It contains your research proposition and its sub-questions, what evidence will answer each one, the figure plan, where the data comes from, which conclusions this design simply cannot reach, and fallback options for when a path does not work out. The methods sections are written to the reporting checklists collected by the EQUATOR Network — STROBE for observational designs, STROBE-MR for Mendelian randomization.

One thing that will feel obstructive but is protecting you

Before you start, it will tell you that with this design your conclusions can only say "associated with", not "causes". That sounds like cold water at first — but overclaimed conclusions are among the most common reasons reviewers reject a manuscript, and learning it before you write is far cheaper than learning it from a revision request.

The goal is not only to understand how a paper's analysis was done, but more importantly why the authors made those choices. If you want to try one thing first, this is the best entry point: drop in a reference paper from your own field and you will immediately see what it pulls apart.

1. Step 1 · Fast pre-screening: /bio-design scan judges in well under a minute whether a paper is worth a deep read, and supports recursive directory scanning. The chain includes deduplication, information extraction, rapid evidence-chain assessment, and an external citation-context search across PubMed and Crossref — checking whether others support or refute the paper.
2. Step 2 · Full decomposition: upload a PDF and type /bio-design. Each Results subsection becomes one module, annotated with its L1–L6 evidence level, scientific sub-question, methods and tools, key parameters, and inputs and outputs.
3. Step 3 · Three extra dimensions per module: the rationale for each method choice, its preconditions, and its limitations. When the paper does not state these, the system records "not stated by the authors" — that absence is itself information about the paper's rigour.
4. Step 4 · Honest downgrade without full text: when only a title or a DOI is available, a reconstruction mode kicks in with tiered confidence labels and explicit "I do not know" boundaries, preventing speculation from contaminating the knowledge base.

Supported formats: PDF, TXT, Markdown, DOCX, DOC, RTF, HTML and URLs. Decomposing a small reference set typically costs under 2 hours and pays for itself across the following 30 days of design work.

The L1–L6 evidence hierarchy is the plugin's shared language: it maps every analysis step you plan onto a specific evidence strength, in the same spirit as GRADE and the Oxford CEBM levels of evidence, but expressed in the vocabulary of a computational study. There is exactly one hard rule: claim strength must not exceed the strength of the most critical evidence layer.

L3, L4 and L5 are orthogonal dimensions rather than a progressive ladder — skipping L3 to go straight to L4 is a legitimate design choice.

One counterintuitive insight from the submission strategy map: impact factor does not correlate positively with the number of evidence layers — in the knowledge base a three-layer study published at IF 10.3 while a five-layer study published at IF 6, a gap of roughly 1.7x in the wrong direction. For a clinical researcher this matters: not being able to run wet-lab validation is not a reason to give up on a good journal.

All 546 trap patterns in the knowledge base come from real errors in published literature, and they are intercepted at the design stage:

• Data leakage: contamination between training and validation sets, feature selection performed on the full dataset
• Statistical artefacts: correlations propped up by outlier samples, uncorrected multiple comparisons, mismatched statistical units
• Overclaimed conclusions: causal language on L1 evidence — one of the most common reasons reviewers reject a manuscript, and a direct conflict with the ICMJE recommendations on accurate reporting
• Foreseeable reviewer objections: 413 reviewer-perspective patterns let you write the defence into the design before the objection arrives

A companion domain-aware transfer mechanism labels every piece of experience with an applicability scope and a disease-domain distance: immune-infiltration experience from reproductive medicine cannot be transplanted directly into cardiovascular research — transferable methodologically, not biologically. Knowledge temporality follows a "learn the strategy, re-verify the claim" principle: claims are tagged time-sensitive (3 to 10 years), stable (5 to 20 years) or permanent, and outdated claims are labelled rather than deleted because the strategic experience attached to them remains valid.

Data verified 2026-08-11 against measured bio-design knowledge-base values.

There is one main entry point, and the system routes you to the right workflow automatically:

/bio-design                          # main entry, auto-routing
/bio-design ferroptosis in myocardial ischemia using public GEO data
/bio-design scan papers/             # batch pre-screen a whole directory
/bio-design continue                 # resume your last topic
/bio-design review Phase_output/     # post-execution review, drives design iteration
/bio-design refresh                  # knowledge temporality check
/bio-design depth beginner           # guidance depth: beginner / working / expert

• Natural-language triggers: mentioning literature decomposition, analysis approach, study design, pipeline or "how did this paper do it" activates the skill automatically.
• Three guidance depths: beginner (for wet-lab or clinical backgrounds, explaining the reasoning in plain language throughout), working (default) and expert (minimal). Depth adjusts only how much is explained — it never reduces the safety checks.
• Practical tips: decompose one or two reference papers before designing; be honest about your constraints; trust the wording calibration; read the trap warnings carefully.
• Requirements: on the cc-bioinfo platform nothing needs installing — open a browser and start.

A typical first session runs 1 to 2 hours of dialogue and ends with a design document you can hand straight to the execution engine.