binder-design-campaign

Plan and run a protein design campaign within a user-approved stage and resource boundary. Branches: de novo binder design (epitope choice, generation, co-folding ensemble ranking, ranked order sheet)

SKILL.md

GitHub

Protein design campaign

Three branches, one core. First turn the user's goal into a reviewable campaign plan; after the user chooses how far to execute, the evidence rules and deliverable discipline below hold for all three branches.

Branch Read For
binder design workflows/binder-design.md miniprotein binders against a protein target
structure prediction workflows/structure-prediction.md folds or complexes with calibrated confidence
protein engineering workflows/protein-design-general.md stability, enzyme, interface, scaffold redesign

Provenance and rationale: docs/ and NOTICE.md, relative to this skill.

Non-negotiables

These hold at every depth, including inside sub-agents, and bind hardest on the steps nobody is watching.

  1. State only what you established. Every claim in a report, update or manifest traces to an executed computation or a saved artifact you can point at. Verified means you ran the check and hold its output.
  2. Every external identifier is fetched. DOIs, PMIDs, PDB IDs, UniProt accessions, residue ranges and URLs in any deliverable are the literal output of a lookup executed in this session.
  3. Anomalies are bugs until investigated. A score of exactly zero, a gate that passes everything, a gate that fails everything, a constant metric, a perfect metric, an impossible runtime, zero variance across seeds: halt that stage, diagnose, then proceed.
  4. Cheapest falsifying check first. Run the seconds-scale check that could kill a plan before committing hours of GPU to it.
  5. Lead with the unfavourable reading. Headline the worst defensible interpretation of your own data. Disclose deviations from the stated method, and name analyses tuned after seeing results along with how many configurations you tried. Report inconclusive results as inconclusive.
  6. The instrument is a filter, never an oracle. In silico confidence ranks designs within a target. It does not tell you whether a target will work, and it relates only weakly to affinity — measured, not assumed (references/failure-modes.md §1). Every summary you write says so.
  7. Verify inherited inputs. Check target structures, construct definitions and hotspot numbering against their primary source before building on them.
  8. Scope. This protocol is in silico and it ends at the design sheet. Validation is in vitro, by the user or a CRO; analysing returned assay data is a different task and does not run under these gates. Results from an earlier round enter here as operator input at G0, never as something this protocol goes and fetches. Work on select agents or toxins, and enhancement of pathogen function, is out of scope: stop and say so. See docs/responsible-use.md.

Plan, authorization and checkpoints

Read references/campaign-planning.md before G0. Research what can be found independently, then ask the user for the remaining outcome, constraint and resource boundaries in one concise round. Create state/campaign_plan.md from templates/campaign_plan.md. It states the stage sequence, evidence needed at each checkpoint, expected outputs, estimated compute, unresolved assumptions, and actions needing separate permission. G0 is the one required review: do not start execution until the user approves the plan and selects an authorized_through_stage from 0 to 10. A value of 0 means plan only; 10 authorizes the in-silico campaign through delivery. Record that decision in state/decisions.jsonl and state/authorization.json.

G1–G6 are scientific and artifact checkpoints. At each, read the corresponding section of checklists/gate-checklists.md, save its evidence and decision, and continue when the approved stage limit covers the next stage. The agent makes in-scope scientific choices from sourced evidence; the user may require a review stop at any checkpoint or override a choice. A checkpoint never serves as a request for an unqualified user to certify a protein-engineering claim. Stop and return a reviewable result when the approved stage limit is reached, when a material premise changes, or when an action needs resources or external permission outside the approved plan. Do not silently widen the scope.

Checkpoint Records Stage reached
G0 user-approved plan, execution limit, resources and permissions 0
G1 assay construct, reference structure, cofactors 2
G2 epitope and hotspot residues 3
G3 scoring evidence level and frozen instrument 4
G4 compute allocation per target 5
G5 the ranked set proposed for testing 9
G6 acceptance checks against delivered artifacts 10

Two carry most of the risk:

  • G2 decides most of the outcome. Epitope choice dominates model choice.
  • G3 is mechanical. Production jobs require a validated gate file. A fold-only gate is visibly weaker than positive-control separation and requires explicit inclusion in the approved plan; scripts/preflight.py refuses it otherwise. A stated belief that the instrument works is not evidence.

Checkpoints also re-open mid-run. Changing the assay construct, a catalytic or otherwise load-bearing residue, the off-target definition, or the endpoint the campaign is aimed at changes the scientific question — so return to the checkpoint that owns it and obtain approval for the revised plan when the change affects the user's goal, resource ceiling, external actions or authorized stage limit.

Completion bar for an authorized segment: every stage through its approved limit has saved artifacts and checks; G6 completion requires the full campaign artifacts. The user's G0 authorization and any later amendments are recorded in state/decisions.jsonl. Agent scientific decisions are recorded separately as agent decisions with their evidence.

Between gates

You work on your own here, subject to five standing obligations.

  • Ledger before narrative. Every job appends one row to the JSONL ledger (scripts/ledger.py). Every count you state anywhere is an aggregation over that ledger computed at report time.
  • Nothing is discarded. If it was computed, it is saved: the exact config that produced it (model, version, weights checksum, environment id, command line, contig or epitope spec, seeds, input paths), every structure file, every metric at the level it was computed, the run manifest, stdout and stderr.
  • Canary before fan-out. Any fan-out past roughly 10 jobs follows one canary from the same spec, confirmed past startup with sane output.
  • Progress, not liveness. A job alive and producing nothing is a failure in progress. Gate completion on output counts and explicit markers.
  • Surface a blocker when it blocks, not at the next gate.

Campaign spine

G0 approved plan -> setup + tool bring-up -> G1 dossier -> G2 epitope
   -> G3 instrument validation -> G4 scale -> generate & screen
   -> optimisation -> final-tier scoring -> G5 selection -> G6 delivery

Bring-up runs before G1 because it is slow and its outcome constrains what you can honestly promise at G2 and G4. A tool joins the roster only after running end to end on a real campaign target, in the environment and flags production will use, with its output consumed by the next stage — the handoff, not the tool alone (references/tool-catalog.md §1).

Sub-agents

Where the runtime offers parallel sub-tasks, split by (target x stage) or (target x model); each brief inherits the non-negotiables and owns only its scope. Where it does not, run the same decomposition serially and say so — the protocol depends on the decomposition, not on the parallelism.

Adversarial review is part of the protocol. Before G6, one reviewer re-derives a sample of scores from raw inputs, and one forms its verdict from raw data without reading your conclusions. Every finding ships with the query or script that produced its numbers.

Reference index

File Read it
references/campaign-planning.md before G0, and when the authorized stage or scope changes
checklists/gate-checklists.md at every gate — G0 also sets the scale preset
templates/campaign_plan.md write the staged plan for user review at G0
references/target-dossier-guide.md building the dossier, before G1
references/failure-modes.md before G2, before G5, and whenever a result looks too good
references/validation-gates.md at G3, or when a control fails
references/scoring-instrument.md defining, running or changing the score
templates/scoring_policy.yaml the ranking formula itself — frozen at G3, and the only place it is written
references/thresholds.md choosing absolute cutoffs, or diagnosing a low pass rate
references/tool-catalog.md choosing, sizing, installing or dropping a tool
references/compute-backends.md wiring the scheduler, at G0
references/deliverables-spec.md building any deliverable
templates/ starting a dossier, config, sheet schema or report
scripts/ before G5, and against the as-shipped artifacts before G6

Tool names, flags, weight URLs and licence facts here were correct at their recorded date and are re-verified against the upstream README, model card or paper before use; upstream wins, and the discrepancy goes in the campaign log.