exercise — Claude Code skill for workout plans and setup
This Claude Code skill builds exercise plans from health-state data and research. For a single upcoming workout, it uses recent Hevy sessions and submits the session for approval.
MIT ★ 93 updated 2026-11-15
npx skills add Insilico-org/plsi --skill exercise -a claude-codeWhat exercise does
exercise gathers state data, searches for research, and prepares an exercise plan for PLSI Nostradamus. The proposal includes exercise dosing and predictions for available measures, alongside a no-action comparison. Nostradamus checks and adopts the plan; the skill does not do that itself.
For one workout, it considers recent Hevy sessions, sleep, recovery, and recorded limits. New exercises get no assigned weight; familiar ones start from the user’s latest sets. PLSI and the account owner’s connected Hevy account are needed. PLSI Wearables must be on the team to connect Hevy. A positive mood screen narrows the session to easy or restorative work.
What it helps with
- build an exercise plan from current measures and research
- compare a plan’s forecast with a no-action scenario
- choose a next workout from Hevy history
- adjust exercise intensity to state and recovery
Example prompts for the agent
- Build an exercise plan from my current measures and relevant research.
- Plan my next workout using my recent Hevy sessions.
- Create an easy recovery session based on my current state.
How to install exercise
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill exercise -a claude-code2Manually: put the files in place
~/.claude/skills/exercise/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/exercise/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill exercise -a codex2Manually: put the files in place
.agents/skills/exercise/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill exercise -a cursor2Manually: put the files in place
.cursor/skills/exercise/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/coach/exercise2Manually: put the files in place
.github/skills/exercise/SKILL.md1Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill exercise -a gemini-cli2Platform CLI
gemini skills install https://github.com/Insilico-org/plsi --path skills/coach/exerciseManually: put the files in place
.devin/skills/exercise/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 exercise
How do I install exercise in Claude Code?
Run npx skills add Insilico-org/plsi --skill exercise -a claude-code. The skill file will be at ~/.claude/skills/exercise/SKILL.md.
What does exercise do?
It prepares an exercise plan using state measures and research, then passes the proposal to PLSI Nostradamus. It can also build one upcoming workout from Hevy history.
What does exercise need to work?
Plans require PLSI. A single workout requires the owner’s Hevy account; PLSI Wearables handles the connection and must be on the team.
Source text
SKILL.md · MIT
---
name: exercise
description: Use as PLSI Coach when the user wants an exercise plan - dose exercise from their state and run-time evidence, predict it against a no-action control and propose it to PLSI for nostradamus/plan - or one next workout session, built from their recent Hevy workouts and sent to Hevy on approval.
---
# PLSI exercise plan
**Status: placeholder**.
## When to use
- "Give me a workout plan", "I want to get fitter".
- "Plan my next workout", "what should I do today?": **Next session** below, not the steps.
- An activity picked in `navigator/activities`, or a goal in `personal_context` about fitness or exercise.
- Not when the request `summary` names a `scope:` line (`orchestrator/urgent`): return a `result` that declines
and names the care to point to.
- When the request `summary` names `caution: mood-screen-positive` (`orchestrator/urgent`): do **not** hard-decline;
NARROW to easy/restorative work only (RPE ≤ 3 or talk-test easy), no progressive overload or progressive WHO until
clinician ack / revision_rule; keep a clinician check as a `test` when the screen is still open.
## Steps
1. **Read the state** and the coordinates that exist now (BMI, waist, blood pressure, glucose, inflammation, clocks).
Resting heart rate, HRV, sleep duration, daily steps, weekly exercise minutes, active days per week and VO2max
estimates are on `sleep_activity` (wearables, `wearables/import`; without PLSI Wearables on the
team, only what was recorded in chat). Never invent a coordinate: a marker with no map coordinate goes in your `result` as a request for
`processor/add-map` (a measured quantity) or `processor/add-score` (a formula; PLSI routes it).
```bash
plsi maps list
plsi state show --instance <instance> --map cardiometabolic
plsi state show --instance <instance> --map sleep_activity
plsi state show --instance <instance> --map care_context
plsi state show --instance <instance> --map personal_context
```
A resistance or recomposition plan also reads `body_composition` (`body_fat_pct`, `grip_strength`) and predicts
on them, not on weight alone, which can stay flat while fat falls and muscle grows.
2. **Fill gaps in ONE `question` handoff** (then wait), only for what neither the state nor the request's `Known:` block holds: current exercise routine, limitations, time and budget band.
List the answers in your `result` as `goal`, `preference` or `condition` observations for `processor/intake`.
3. **Gather evidence at run time** and verify every identifier:
```bash
plsi evidence search --source pubmed --query "<exercise type> <coordinate> randomized" --limit 10
plsi evidence search --source europepmc --query "<exercise type> <condition> safety"
plsi evidence verify --kind pmid --id <pmid>
```
Reuse the evidence library and verify only what you cite (`shared/team`, Evidence speed).
4. **Dose exercise** (`category: exercise`), one plan action per type (aerobic, resistance, balance): sessions per
week, minutes, intensity (talk test or perceived exertion; heart-rate zones only from the user's own data), progression
and recovery weeks as the cited guideline sets them. Start from the user's own `exercise_minutes` and `active_days` when the
state has them; the weekly minutes target comes from the cited guideline, not from memory. Limits from
`care_context` conditions (cardiac, joint, glucose-lowering medication): lower start, warning signs that stop a
session, and a clinician check as a `test` action where the evidence says so. Under
`caution: mood-screen-positive`: keep intensity easy/restorative (RPE ≤ 3), drop progressive overload, and say
NARROW in the `result` `summary` while repeating the caution line.
5. **Predict** on predictable coordinates only (e.g. `sbp`, `waist`, `bmi`, `hba1c`, `hs_crp`,
`phenoage_acceleration`, `exercise_minutes`, `active_days` and `vo2_max`, the last as a trend of
device estimates only): `change_from_baseline` at 90 days with an honest interval that includes adherence and
measurement noise, effect sizes from step 3, derived values with `plsi calc derive ... --set <input>=<value>`,
plus one no-action control (`driven_by: []`) started from `base_rate` references (`plsi refs list`). Adherence
targets (sessions per week) go in `method` and the `revision_rule`, not as predictions.
6. **Propose to PLSI** for `nostradamus/plan` (PLSI Nostradamus) steps 2-4: write the actions, claims and
predictions to `work/<thread>/proposal-coach.json` (`shared/team` rule 3) and return a `result` with its path; Nostradamus merges them into the one plan, screens against `care_context`, runs `plan check`,
has PLSI hand it to `challenger/challenge` and records and adopts only after PLSI's adoption card. Never check,
record or adopt a plan yourself.
## Next session
One workout, not a plan: no predictions and no Nostradamus. A request for a program goes to the steps above.
Only for the Hevy account's owner, the user's own instance (relation `self` in `plsi instances list`): one key is one person's log, so a session for anyone else
(family, a patient) is a plan in chat, never a Hevy routine. Cues name the exercise and its form only (tempo,
rest, range), a ladder written with commas ("50, 60, 70 kg"): never a reason drawn from the record (a
condition, an injury, a medicine, a value).
1. **Read** the ask (focus, minutes, what to leave out), an adopted exercise plan (`plsi status --instance
<instance>`, so the session fits it), the latest sleep, resting heart rate and HRV on `sleep_activity`,
limits in `care_context`, and the last two weeks of training:
```bash
plsi hevy check
plsi hevy workouts --since <14 days ago> > <PLSI_HOME>/work/<thread>/hevy-workouts.json
```
Hevy not connected: ask in ONE `question` what was trained in the last few days, and offer "Connect Hevy?"
(PLSI Wearables, `wearables/hevy`) when Wearables is on the team (`plsi team show`), else "Add PLSI Wearables to
connect Hevy?"; only Wearables connects it.
2. **Compose** within the minutes, warm-up included: the muscle groups asked for, a muscle group worked hard in
the last 48 hours only lighter, and what the user excluded never. Exercises the user has done before come
first; their load starts from the user's own last sets of that exercise and moves up only after the last
session's sets were all done. A new exercise gets no weight: its note says to pick one that leaves two reps.
Sleep or recovery well below the user's own usual: a lighter session, and say why. Pick ids with:
```bash
plsi hevy templates --muscle <muscle group>
plsi hevy templates --search <exercise name>
```
3. **Write** `<PLSI_HOME>/work/<thread>/session.json` (`title`; `notes`; `exercises` with
`exercise_template_id`, `rest_seconds`, short `notes` with cues and `sets` with `type`, `reps`, `weight_kg`,
`duration_seconds`) and check it; its notes are about the exercises only, never health details:
```bash
plsi hevy routine --file <PLSI_HOME>/work/<thread>/session.json
```
4. **Return one `result`**: the session as the user reads it (exercise, sets × reps, weight, rest, about how many
minutes, and why it looks like this), `session.json` and its `session_hash` in `files`, and the card "Put it in
Hevy?" (Yes / Change something / No). Without Hevy: the session only.
## Rules
- Crisis signs (self-harm or suicide thoughts, chest pain on exertion, fainting) in any input: stop planning and
return a `result` with a `red_flag:` line (`orchestrator/urgent`) at once. Eating-disorder signs: stop and return a
`result` with a `scope:` line (`orchestrator/urgent`) instead. Soft `caution: mood-screen-positive` is neither:
NARROW, do not hard-decline.
- Start low for anyone inactive, older or with conditions; progress only when the previous step is tolerated.
- Sample sessions go in the `result` summary for PLSI to show; only plan actions and predictions go into `plan.json`.
- No product, brand or drug names; placeholders such as `<device>`.
© Insilico-org · MIT · Insilico-org/plsi