wayfinder
CLAUDE.md0 starsChanged 4 months agoMITView on GitHub
Hackathon build (Built with Claude: Life Sciences, Researcher track — a researcher who also builds). A hypothesis referee: Claude agents adjudicate a mechanistic T-cell claim and return a verdict with a receipt for every hop. The framing: a researcher who builds their own instruments — the referee, the LBD question-engine, and the Claude-Science automation are tools built to ask and answer the research questions. The deliverable is a reproducible T-cell finding + how Claude Science reached it; the tooling is…
In this repository5 other files it gives its agents.
- AGENTS.mdAGENTS.md
- atomic-plannerSkill
- freshenSkill
- session-closerSkill
- session-startSkill
The file
We split this CLAUDE.md into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
NEW WORK ONLY: Every submitted file is authored during the event (started July 7, 2026).
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Wayfinder — project map
Hackathon build (Built with Claude: Life Sciences, Researcher track — a researcher who also builds). A hypothesis referee: Claude agents adjudicate a mechanistic T-cell claim and return a verdict with a receipt for every hop. The framing: a researcher who builds their own instruments — the referee, the LBD question-engine, and the Claude-Science automation are tools built to ask and answer the research questions. The deliverable is a reproducible T-cell finding + how Claude Science reached it; the tooling is the method/vehicle, and the builder-craft is evidence of a scientist who can build to get the science done. Full plan: docs/plan.md (v7). This file is the lean map.
Hackathon + what we're doing now
Full canonical fact sheet: docs/HACKATHON.md (event, track, deadline, prizes, rules, judging, submission runbook). Quick version:
- Event: Built with Claude: Life Sciences (Anthropic × Gladstone, run by Cerebral Valley), Researcher track ("a researcher who also builds").
- Timeline: started July 7, 2026; deadline July 13, 2026, 9:00 PM ET (see Hard constraints). Solo builder; async/interruptible.
- Deliverable: a reproducible T-cell finding (NAB2 → Th1/Th2 → atopic eczema, receipt-backed) + how Claude Science reached it — the referee, the LBD question-engine, and the headless Claude-Science automation are instruments built to get the science done.
- Submission status: BUILT + FIRE-READY. Repo
dayanjan/wayfinder(private); say "scrub and flip" to runSUBMIT_CHECKLIST.mdand flip it public. Demo video + 3-screen Streamlit app = the MVP. - Current post-hackathon thread: a manuscript for FRMA (
docs/manuscript/latex/main.pdf) hardened by a 13-agent contribution/novelty audit and a claims↔experiments reconciliation (below). Live state:WORK_PROGRESS.md+memory/NEXT_SESSION.md.
Claims ↔ evidence traceability (READ-FIRST for any claim work)
docs/CLAIMS_EVIDENCE_LEDGER.md is the canonical living index: every manuscript claim (M1–M10, B1–B7, R1–R5) → status → the experiment artifact that supports/refutes it → primary source (S1–S11) → review-panel critique + resolution.
- Anti-amnesia protocol (mandatory): any claim audit / review / referee-response / revision reconciles against BOTH the literature AND the experiment corpus the ledger indexes — literature tells you if a claim is novel, the experiments tell you if it is supported; never one alone. (This exists because a 2026-07-12 literature-only audit forgot the Claude-Science confounder experiments and over-generalized the NAB2/STAT6 critique — 2 of 3 confounder channels were already experimentally closed.)
- Dated records under
docs/reviews/contribution-novelty-audit_2026-07-12/anddocs/claims-vs-experiments_2026-07-12/are immutable evidence (cite, don't edit); the ledger is the living index you update. Open gaps →docs/claims-vs-experiments_2026-07-12/NEW_EXPERIMENTS.md(G1 held-out eval = the publish gate).
The thesis (don't lose this)
The edge is the confident, receipt-backed NO — refuting a plausible claim with a real experimental receipt, and catching artifacts (failed knockdown -> untested, not negative). Falsification, not confirmation, is the moat. The knockdown-QC gate is a hero feature, not a footnote.
Hard constraints
- NEW WORK ONLY. Every submitted file is authored during the event (started July 7, 2026). No reuse of prior projects (PyZoBot POC is reference only, zero file copy). Git history is the compliance proof.
- Open source, MIT. Entire repo is public.
- Deadline: July 13, 2026, 9:00 PM ET. Solo builder; async, interruptible workflow — keep
memory/NEXT_SESSION.mdcurrent so any session resumes cleanly. - Claude for reasoning/generation; Voyage for embeddings (Anthropic has no embedder). Never OpenAI.
Conventions
- Every causal edge shown traces to a data receipt (OR / p-value / effect size). Claude interprets receipts; it never asserts biology.
- Calibrated language only: "consistent with / re-derived / refuted / untested / flagged." Never "discovered/proven."
- Data lookups are deterministic tools, not "agents." Visible agency is reserved for judgment (the Skeptic weighing literature, the Adjudicator calibrating confidence).
- Compute is CPU-only (pandas over the aggregated CSVs). No GPU, no 22M-cell raw matrices.
Structure
docs/plan.md # the v6 build plan (thesis, agent cast, 3-hop substrate, judging map)
data/ # public Perturb-seq supplementary tables (fetch via fetch_data.sh; gitignored)
src/arbiter/ # the tool (fresh code)
tests/
memory/NEXT_SESSION.md # async handoff — canonical (read/written by session-start & session-closer)
MEMORY.md # lean session index (status, decisions, recent sessions)
WORK_PROGRESS.md # live dashboard (phase, milestones, blockers, progress log)
memory/ # sessions/, decisions/ (ADRs), lessons-learned.md, patterns.mdSession workflow (PM skills)
Async, interruptible, solo. Use the session-lifecycle skills at the boundaries:
/session-start— at the top of a session: reads the handoff, cross-checks state, briefs you./session-closer— at the end: writes the session log, syncsMEMORY.md/WORK_PROGRESS.md, refreshesmemory/NEXT_SESSION.md, commits housekeeping (or checkpoints)./freshen— drift check across the PM docs./atomic-planner— decompose a workstream into 1–3h testable task cards.
Codex delegation + review skills (codex-rescue, codex-review-diff, codex-spike, …) are user-level and already available.
Judging weights (aim every hour here)
Demo 30% - Claude Use 25% - Impact 25% - Depth 20%. See docs/plan.md section 12.
Delegation
Solo + agent fleet. Delegate mechanically-specifiable chunks (data-lookup tools, provider adapters, deploy) to Codex; keep judgment (anchor vetting, verdict UX) local. Cross-agent brief: AGENTS.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

