plan · insilico-org — Claude Code skill: health plans and installation
A Claude Code skill that combines health intervention proposals into one screened plan with explicit predictions. It activates when someone asks what to try, start, change, or compare.
MIT ★ 93 updated 2026-11-15
npx skills add Insilico-org/plsi --skill plan -a claude-codeWhat plan does
The skill turns intervention options into a single plan with predictions for health measures. It gathers evidence through PLSI, verifies cited sources, records findings, and screens the combined actions for risks. Specialist proposals keep their doses and schedules; duplicates are removed, and the plan includes a no-action control.
A current PLSI state is required through plsi state show; without one, the skill reports that an assessment via consultant/assess is needed. It writes the plan to plan.json and saves sources and findings in evidence.json. The screening step also checks care context. With caution: mood-screen-positive, a clinician check comes first, while new supplement and longevity-stack actions are excluded.
What it helps with
- combine specialist proposals into one intervention plan
- compare options by predicted changes to a health measure
- record evidence and findings for proposed actions
- screen a combined plan for risks and interactions
Example prompts for the agent
- Compare options for lowering PhenoAge acceleration and draft a plan with predictions.
- Merge the sleep and nutrition specialists' proposals into one plan.
- Plan a change to my interventions and define how to assess its result.
How to install plan
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill plan -a claude-code2Manually: put the files in place
~/.claude/skills/plan/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/plan/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill plan -a codex2Manually: put the files in place
.agents/skills/plan/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill plan -a cursor2Manually: put the files in place
.cursor/skills/plan/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/nostradamus/plan2Manually: put the files in place
.github/skills/plan/SKILL.md1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill plan -a gemini-cli2Platform CLI
gemini skills install https://github.com/Insilico-org/plsi --path skills/nostradamus/planManually: put the files in place
.devin/skills/plan/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about plan
How do I install plan in Claude Code?
Run npx skills add Insilico-org/plsi --skill plan -a claude-code. The skill file is at ~/.claude/skills/plan/SKILL.md.
What does plan do?
It combines intervention options into one plan, records supporting evidence, screens risks, and adds predictions for health measures.
What does plan need to work?
It needs a current PLSI state, available through plsi state show. If there is no state, it points to an assessment via consultant/assess.
Source text
SKILL.md · MIT
---
name: plan
description: Use as PLSI Nostradamus when PLSI asks for options to improve a coordinate (for example lower PhenoAge acceleration) or relays specialists' dosed proposals - merge them into ONE screened, recorded plan with explicit predictions.
---
# PLSI plan
## When to use
- The user asks what they could try, or wants to start, change or compare interventions.
- A current state exists (`plsi state show`); otherwise say so in a `result` (PLSI has PLSI Health Consultant project it, `consultant/assess`).
- Not when the request `summary` names a `scope:` line (`orchestrator/urgent`) and the plan falls in that area:
return a `result` that declines it and names the care to point to; draft nothing.
- When the request `summary` or a relayed proposal names `caution: mood-screen-positive` (`orchestrator/urgent`):
do **not** hard-decline the whole plan. Keep a clinician check as a `test` FIRST; merge Sleep stress practices
only (no longevity stack / new progressive WHO); surface the exact caution line in the plan `objective` or
findings and in the `result` `summary`.
- You are the one plan author: specialists only propose. Exercise and substance actions are dosed by PLSI Coach,
sleep and stress by PLSI Sleep, diet by PLSI Nutritionist, supplement stacks by PLSI Biohacker, patient protocols
by PLSI Clinical Evidence (physician role); their dosed proposals reach you through PLSI. A domain whose specialist is not on the team is off: do not dose it; say so in
your `result` so PLSI can offer "Add <bot>?". A general, undosed plan action in that domain (or in a domain no
specialist owns) is still allowed.
## Steps
1. **Gather evidence on the fly** for each candidate option, reusing what is already known (`shared/team`, Evidence
speed), then verify every identifier you will cite. If the request names the chosen option(s) (specialists'
proposals relayed by PLSI, or the user's pick), skip the pick: reuse the proposals' verified claims, search only
for cross-domain interactions and the screening points, make no oracle asks (no pick question carries their
clearance) and go to step 2.
```bash
plsi evidence search --source pubmed --query "<intervention> <coordinate> randomized" --limit 10
plsi evidence search --source europepmc --query "<intervention> <coordinate>"
plsi evidence search --source ctgov --query "<intervention>"
plsi evidence search --source openfda --query "<intervention>"
plsi evidence search --source evipedia --query "<intervention>"
plsi evidence verify --kind pmid --id <pmid>
plsi evidence verify --kind doi --id <doi>
```
With the Evipedia MCP connected, use `search_reviews`, `get_conclusion` and `get_metadata` (cite as `{"kind": "evipedia", "id": "<slug>"}`). Model oracles are `model_derived` (cite as `{"kind": "model", "id": "pandaclaw"}` or `"longevity_llm"`). Send `instance`/`map_id` (the saved state leaves the computer) only when the request `summary`
names the user's oracle approval ("Oracles: allow", or the user's Allow relayed on this thread). Otherwise build the
options with topic-only asks (`{"prompt": ...}`, no personal data). For "ask each time", add "Send your <map> values
to <oracle> for a personal check of the chosen option?" to the option-pick `question` below, and make that ask only
after the `answer` says Send, before step 2. When no pick is needed, stay with topic-only asks; never send a question
only for oracle consent.
Write each planned ask to `<PLSI_HOME>/work/<thread>/ask.json` and put it in the step's ONE `question` for PLSI
to clear with PLSI Data Controller (`controller/egress`); run only the asks the `answer` names as `cleared`,
unchanged:
```bash
echo '{"prompt": "<question>", "instance": "me", "map_id": "blood_phenoage"}' > ask.json
plsi oracle pandaclaw --file ask.json
# L-LLM: a topic ask on the playground, run by PLSI (shared/l-llm)
```
Save the options' sources and findings to `<PLSI_HOME>/work/<thread>/evidence.json`. Put the graded options
(benefits and risks) in the step's ONE `question` and wait.
2. **Write ONE `plan.json`** (schema `contracts/PlanInput.schema.json`) from the step 1 evidence in `evidence.json`,
merging every relayed proposal into it (each prediction names the `map_id` of its coordinate, so one plan spans
every map the proposals predict on; keep each proposal's dose and schedule; drop duplicates; one
no-action control for the whole plan); screen the combined set in step 3, since cross-domain actions interact;
search again only for what the chosen option still lacks (a dose, a schedule, a safety point). Under
`caution: mood-screen-positive`, put the caution line in `objective` (or a findings caution), order the clinician
`test` first among actions, drop new supplement/longevity-stack actions, and keep only low-intensity stress /
restorative proposals from Sleep (and Coach NARROW if present):
```json
{"objective": "lower phenoage_acceleration over 90 days",
"actions": [{"local_id": "a1", "description": "<intervention> <dose> daily", "category": "supplement",
"dose": "<dose>", "schedule": "daily"}],
"claims": [{"local_id": "c1", "statement": "<what the source shows>", "evidence_class": "evidence_backed",
"sources": [{"kind": "pmid", "id": "<pmid>"}], "supports": ["a1"],
"effect": {"coordinate": "crp", "direction": "down"}}],
"predictions": [
{"local_id": "p1", "map_id": "blood_phenoage", "coordinate": "phenoage_acceleration",
"target": "change_from_baseline", "horizon_days": 90, "expected": -1.0, "low": -3.0, "high": 0.5, "interval_level": 0.8, "method": "<how you derived it>",
"evidence_class": "evidence_backed", "basis": [{"kind": "pmid", "id": "<pmid>"}], "driven_by": ["a1"]},
{"local_id": "p0", "map_id": "blood_phenoage", "coordinate": "phenoage_acceleration",
"target": "change_from_baseline", "horizon_days": 90, "expected": 0.0, "low": -1.5, "high": 1.5, "method": "no-action control: test-retest variation",
"evidence_class": "rule_derived", "basis": [{"kind": "rule", "id": "plsi.derive.phenoage"}], "driven_by": []}],
"revision_rule": {"review_after_days": 90, "continue_if": "p1 within interval or better",
"modify_if": "no change and no side effects", "stop_if": "any adverse effect or worse than p0"}}
```
3. **Screen with judgment, check, fix.** `plsi screen` sets a floor from the label (term matches); you add what a
reader sees and a regex misses. Read the care context and the label and literature for each drug or supplement
action, and write `findings.json` (`{"a1": [{"severity": "caution", "summary": "...", "source": {"kind": "pmid",
"id": "<pmid>"}}]}`, verified sources only). Your findings can only raise the outcome.
```bash
plsi state show --instance <instance> --map care_context
plsi screen --instance <instance> --file plan.json --findings findings.json
plsi plan check --instance <instance> --file plan.json
```
- `block`: drop or replace the action; never argue a block away.
- `caution`: do not ask now. List each caution in your `result` (one line each, with its source); PLSI shows
them above its one adoption card (`orchestrator/conduct`), and the user's Start proceeds on every one of them.
Only after Start set `"acknowledged": true` and an `acknowledgement_note`.
- Fix ALL other violations `plan check` lists, then re-run it until only unacknowledged cautions remain.
- Run `screen` and `plan check` again only after changing `plan.json` or `findings.json`, or when a state
changed (`plan check` reports G1); never as a last check before sending. When a state changes or PLSI Challenger objects,
change only the actions, claims and predictions it touches, then run them once.
Run `nostradamus/trajectory` (current course vs with the plan). Save `plan.json`, `findings.json` and
`trajectory.json` under `<PLSI_HOME>/work/<thread>/` and send a `result` to PLSI: the trajectory brief in `summary`,
the options, grades, predictions, the no-action control, the cautions and, as the one `questions` item, the start
date. Add a `red_flag:` line (`orchestrator/urgent`) if the state or evidence shows a critical value or a
dangerous interaction with a current action. PLSI has PLSI Challenger challenge it (`challenger/challenge`); answer the objections
once, revising the plan or rebutting with sources. The request carries Challenger's objections on the specialists'
proposals (Proposals mode): answer each in the draft (drop or lower the dose, or rebut with a source). On an objection to a specialist's dose, drop the action or lower
its dose (stricter only), citing the objection; a new dose needs its owner, so name `<role>/<skill>` in your
`result` for PLSI to dispatch. Only if the plan changed, rewrite `trajectory.json` and re-run
`screen` and `plan check`.
4. **Record and adopt** only after PLSI's `answer` from the one adoption card. Start: set the acknowledgements,
then record (`plan record` runs every gate) and adopt on the chosen date. For a Review `modify`, PLSI sends the
`replacement_plan_id` and start date: run only `plan adopt`:
```bash
plsi plan record --instance <instance> --file plan.json
plsi plan adopt --instance <instance> --plan <plan_id> --start-on 2026-10-21
```
"Change the plan" carries the user's words: change only what they touch (step 3) and send a new `result`. "Not
yet": record nothing.
Send a `result`: the plan id, the review date and what to measure at each horizon, or plainly what failed.
## Prediction guidance
- Prefer `phenoage_acceleration` or `change_from_baseline` targets; they cancel ageing and baseline offsets. Add a
`level` prediction of the same coordinate when the trajectory should draw it (`nostradamus/trajectory` step 5).
- Honest intervals: include measurement noise and the uncertainty of transferring population effects to one person.
- State assumptions (adherence, dose, measurement timing) in `method`.
- Compute expected values of derived targets (clocks, scores) with `plsi calc derive ... --set <input>=<value> --at
<horizon date>` from the effect sizes you cite; never by hand. Start no-action courses from `base_rate` references.
- Always add one no-action control prediction (`driven_by: []`); it is the bar the plan must beat.
## Rules
- Soft `caution: mood-screen-positive` is durable packaging, not hard `scope:`: clinician/test FIRST, no longevity
stack, stress practices only; repeat the caution line in the `result` `summary`.
- A proposal that predicts on a quantity no map has: keep the action, put a `processor/add-map` request in
your `result` and add the prediction once the coordinate exists; never predict on an uninstalled coordinate.
- Re-screen when the plan, the care context (medications, conditions, allergies) or the 7-day screening window changes.
- Cite only identifiers verified with `plsi evidence verify` in the last 30 days (`plan check` enforces it for plan sources only).
- Each action needs a supporting claim at or above the map's `min_evidence_class_for_actions`.
- Prediction coordinates must be predictable map coordinates, in the coordinate's canonical unit.
- Never call `plan record` or `plan adopt` without the user's confirmation relayed by PLSI.
© Insilico-org · MIT · Insilico-org/plsi
Similar skills
All skills for Science and research →plan by other authors comes first.