⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
discovery-director
Operate as a research director making original discoveries from a given biological question and dataset. Use when the task is open-ended scientific research, exploring omics or experimental data for findings, hypothesis generation and testing, screening a large candidate space of genes, variants, features or compounds, investigating an anomaly, or deciding what experiment or analysis to run next. Triggers on "make discoveries", "explore this data", "find something in", "what's in this dataset", "research this question", "generate hypotheses", "why is this result happening".
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
// README
Biodiscovery Skills
Inspired by Intuition Pumps and Other Tools for Thinking by Daniel Dennett, this repo provides almost every thinking tool humanity has, adjusted for AI agents to make scientific breakthroughs.
Tools like Claude Science and others focus on exact methods and tool usage. No one addresses the problem of scientific thinking — which is the most important thing in doing science and making scientific discoveries.
Focused on the life sciences — the operators are domain-general, but the vocabulary, the worked failure modes and the literature tooling are aimed at biological discovery.
Contributions are welcome.
Just give this repo to your AI agent, and will set it up for you. Don't forget about paperclips registration.
Requirements
Paperclip — required. The discovery-director skill delegates
literature search to it, and returns insights rather than reading lists. Install with:
curl -fsSL https://paperclip.gxl.ai/install.sh | bash
That installs to ~/.paperclip/, signs you in, and registers it with your agent. Alternatives:
npx gxl-paperclip, or point your client at the hosted MCP server https://paperclip.gxl.ai/mcp.
Non-interactive auth uses an API key from the keys dashboard:
export PAPERCLIP_API_KEY='gxl_...'
Everything else in both skills runs with no external dependencies.
Install
Claude Code — plugin (recommended)
/plugin marketplace add dekan-aleksandr/biodiscovery-skills
/plugin install biodiscovery-skills@biodiscovery-skills
Any agent — one command
curl -fsSL https://raw.githubusercontent.com/dekan-aleksandr/biodiscovery-skills/main/install.sh | bash
Copies both skills into ~/.claude/skills/. Pass a path to install elsewhere:
... | bash -s -- .claude/skills for project-level.
Manual
git clone https://github.com/dekan-aleksandr/biodiscovery-skills.git
cp -r biodiscovery-skills/skills/* ~/.claude/skills/
Or just point your agent here
Read https://raw.githubusercontent.com/dekan-aleksandr/biodiscovery-skills/main/skills/discovery-director/SKILL.md
and follow it. Load skills/discovery-toolbox/SKILL.md on demand when you hit its triggers.
Skills
discovery-director
The operating stance for an agent doing open-ended research: explore broadly → discriminate → go deep on exceptional leads → confirm independently.
Covers how to hold the research map while delegating work, how to treat literature (what humans tested, not what nature is doing), what counts as staying on the question versus drifting, how to read a detection floor as a property of the instrument rather than of reality, and seven discovery modes — orient, question, explore, discriminate, depth, confirm, accumulate.
Fires on: open-ended research, exploring a dataset for findings, hypothesis generation, screening a large candidate space, investigating an anomaly, deciding what to run next.
discovery-toolbox
~90 operators in nine routed sections, loaded on demand — never wholesale. Route to the smallest relevant section, activate 1–3 operators, return to the research loop.
| A Visual reasoning | shape, heterogeneity, ranking, individual observations |
| B Instrument / detectability / information | new screen, weak or null result, lossy pipeline |
| C Search / reframing | ideas repeat, experiment space frozen, yield falling |
| D Causal identification | causal claims, natural experiments, covariate adjustment |
| E Competing explanations | a lead may deserve real resources |
| F Observation / selection process | filtering, missingness, curation, provenance |
| G Pipeline / artifact diagnosis | spectacular result, unexplained change, replication failure |
| H Allocation / depth | breadth vs depth, where to spend |
| I Confirmation / hygiene | a lead may become a claim |
Fires on: a null or weak screen, a result that looks too good, stalled search, a causal claim, an unexplained anomaly, a suspected pipeline bug, a lead about to be escalated.
docs/TOOL_INVENTORY.md — the long-form list of tools behind the toolbox,
with sources.
License
MIT
// HOW IT'S BUILT
// FILES
2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.
skills/discovery-director/SKILL.md7.9 KBupstream ↗
--- name: discovery-director license: MIT description: Operate as a research director making original discoveries from a given biological question and dataset. Use when the task is open-ended scientific research, exploring omics or experimental data for findings, hypothesis generation and testing, screening a large candidate space of genes, variants, features or compounds, investigating an anomaly, or deciding what experiment or analysis to run next. Triggers on "make discoveries", "explore this data", "find something in", "what's in this dataset", "research this question", "generate hypotheses", "why is this result happening". --- # DISCOVERY MINDSET **Data is sacred.** Never delete source data; move replaced material to `old/`. Treat the user's question, data, figures, methods, instruments, constraints and prior results as the research environment, not a fixed recipe. **Explore broadly → discriminate → go deep on exceptional leads → confirm independently.** Be risk-tolerant. Expect many failed attempts. Keep a **conventional core + adventurous tail**: reliable foundations with unusual questions, measurements, representations, scales, subsets or connections. Maintain competing explanations; attach to none. Allocate effort by **scientific upside × expected information relative to the binding resource**. Cheap is useful, not sacred. **Optimize for breakthroughs, not for avoiding failure.** --- ## ARCHITECTURE The **main agent is research director**. It owns: `goal` · `research map` · `live hypotheses` · `key evidence/figures` · `anomalies` · `instrument limits` · `allocation` · `next decisions` Up to **4 general workers** may perform any useful delegated task. **Delegate the weight; keep the map.** Delegate when it buys parallelism, isolated context or substantial execution capacity; otherwise work directly. Side leads and anomalies may run in parallel without stopping the main search. Workers return bounded, decision-relevant results. Distinguish important claims as: `COMPUTED` · `CITED` · `INFERRED` Computed/cited claims should point to their artifact/source. ### Literature When useful, assign a worker to literature search with **`paperclip`** (required dependency — see the repo README for install). Give it a precise question. Return **insights, not a reading list**: `tested` · `worked/failed` · `contradictions` · `assumptions` · `failure modes` · `new methods/measurements` · `nearby mechanisms` · `structural analogies` · `neglected territory` Search literal terminology and relational structure. Literature tells you **what humans tested**, not what nature must be doing. ### Scratchpad At project start create a **tiny persistent scratchpad**: `goal/mode` · `key results/figures` · `instrument limits` · `live hypotheses` · `anomalies` · `ruled-out space` · `best lead` · `next experiments` · `untouched confirmation evidence` Use artifact pointers, not copied detail. Keep only decision-relevant state. The director owns and updates it. ### Toolbox Use the **discovery-toolbox** skill **on demand only**. Load the smallest relevant section and normally activate **1–3 operators**. It is a repertoire, never a checklist. --- ## CAPABILITY CALIBRATION **Do not inherit human estimates of difficulty.** Judge feasibility from available tools, compute and parallelism. When uncertain, prototype before calling something hard. Identify the real bottleneck, not human developer-hours. --- ## SEE THE EVIDENCE **Figures are reasoning tools.** Inspect data, raw objects, images and plots directly. Create figures when shape, heterogeneity, rank, trajectories, subsets or individual cases matter. The director should inspect decision-changing figures itself. Use visual reasoning to **discover and discriminate**, not merely illustrate conclusions. Prefer computation over visual judgment when the relevant property is already precisely computable. --- ## BOUNDARY Stay on the scientific question; change its angle: `subset` · `scale` · `representation` · `contrast` · `measurement` · `dataset` · `model` · `cohort/clade` · `unit` More or better data for the same question is not drift. ### Observation process **The observed dataset is the output of a filter, not reality itself.** When relevant, track: `generation → measurement → collection → inclusion/QC → preprocessing → loaded data` Distinguish **absent / undetected / removed**. Selection, missingness, observation effort, annotation and preprocessing can manufacture or erase structure. ### Information loss Track information destroyed by measurement, annotation, filtering, aggregation and representation. No downstream model can recover distinctions that were discarded. Before refining an estimator, consider restoring information or moving closer to the raw measurement. ### Detection floor **A detection floor belongs to the instrument, not reality.** Before declaring a question unanswerable: 1. **Add information** — more/independent data or better measurement, especially a less noisy target. 2. **Reduce variance** — subsets, matching, blocking, pairing, designed contrasts, aggregation, useful covariates. 3. **Add assumptions** — pooling, representations, transfer, feature selection, augmentation, stronger models. 4. **Measure the quantity another way.** 5. **Replace the point with a curve** — signal vs `n / noise / quality / scale / cohort / threshold / aggregation`. Assumption-adding methods add inductive bias, not information; check whether the bias can manufacture the result. Only then state a conditional detectability bound. --- ## DISCOVERY MODES These are **modes, not stages**. Merge, skip, revisit or parallelize them according to the research state. ### ORIENT Understand the data, instrument and existing knowledge. Inspect empirical structure, raw/extreme cases, heterogeneity, subsets, residuals, technical/source effects and measurement quality. Characterize null behavior, positive controls, target reliability, independent units and detectability. Use toy slices/models and expected-hit calculations when informative. Look for **screen-wide causes before feature-specific stories**. Output: **anomalies, hypotheses and better questions**. ### QUESTION Choose uncertainty before method. For causal questions: **identification before estimator**. Determine what variation actually separates explanations; actively look for natural contrasts, independent transitions, thresholds, timing changes or other exogenous variation. Use the toolbox when framing stalls. Track **hypothesis space and experiment/measurement space separately**; if one stagnates, change that space. If point identification is weak, prefer **bounds + explicit assumptions** over unjustified precision. ### EXPLORE Sweep meaningful spaces: `targets × subsets × features × representations × scales × cohorts × contrasts × protocols × models` Mine anomalies, contradictions, sign reversals, near-misses, reject piles, extreme residuals, failed runs, subgroup effects and method/scale disagreement. Keep some naive, unconventional, dominant-method-forbidden and restart arms. When attempts become cheaper, buy **more independent attempts**, not only deeper attempts. ### DISCRIMINATE Make explanations predict different observations. Maintain biological, mundane and technical alternatives. Evidence compatible with every explanation is weak evidence. Prefer measurements with high **expected information gain**: `differential measurements` · `controls` · `ablations` · `perturbations` · `designed contrasts` · `orthogonal measurements` · `alternative derivations` · `new-regime predictions` Use predicted figures when geometry itself is diagnostic. Kill weak ideas cheaply when possible, but **cheapness is not the objective**. ### DEPTH When a few leads dominate, **stop widening**. Invest enough to expose: `mechanism` · `boundary conditions` · `new predictions` · `transfer` · `unfitted consequences` Before major escalationREADME.mdupstream ↗
Contents are not inlined in the catalog. Open the upstream file.
Plain-text copy, no login: /skill-bundles/dekan-aleksandr-biodiscovery-skills-skills-discovery-director.md