Healthy Fitness Coach Skill in Practice: From Manual Logs to a Reviewable Plan
A workout plan often fails between sessions: Tuesday gets crowded, the gym equipment is unavailable, and by the end of the month nobody remembers which exercises were actually completed. The open-source Healthy Fitness Coach Skill connects a personal profile, a flexible session plan, training logs, and periodic review in an AI-agent workflow. This guide shows how to test that workflow with manual records before connecting a fitness app. The Skill helps organize training decisions; it does not replace clinical assessment or an in-person professional.
This overview describes the project’s purpose; it is not evidence that a particular user completed a training cycle. A pilot should test whether profile, plan, sessions, and review remain connected.
Decide whether this is the right kind of tool
The repository describes its audience as ordinary beginners and intermediate exercisers aged 18–55. It covers strength training, home workouts, aerobic work, and recovery-oriented planning. It is a good candidate when the problem is operational: “I have three 40-minute windows each week; how do I build a plan that survives a busy Wednesday?” It is not a route to a diagnosis, injury rehabilitation prescription, pregnancy or postpartum training prescription, or competitive preparation.
The WHO physical activity guidance recommends regular movement for adults, including 150–300 minutes of moderate aerobic activity or an equivalent amount of vigorous activity, and muscle-strengthening activity on at least two days a week. These are population-level benchmarks, not a personalized clearance assessment. The project’s safety section says automated training advice should stop when warning signs such as chest pain, fainting, unusual palpitations, disproportionate breathlessness, or progressive numbness and weakness appear. Seek professional assessment in those situations.
| Situation | Sensible first step |
|---|---|
| Your schedule repeatedly breaks a reasonable plan | Ask for a normal session and a short fallback session |
| You already have four to eight weeks of logs | Review the recorded period and missing data before changing the plan |
| You have one measurement or a single wearable screenshot | Build a baseline; do not treat a “training DNA” label as established |
| You have persistent pain or warning symptoms | Stop automated prescription and seek qualified assessment |
Install it, then build a profile that can change
The repository provides both a complete Skill directory and a distributable dist/healthy-fitness-coach.skill package. Install one of them in an agent that supports Skills. Basic use does not require Node or a connection to Xunji. After confirming the Skill is enabled, start with information relevant to training, rather than uploading unnecessary identity or location details:
Create my training profile before proposing a demanding program.
Goal: build a consistent habit while improving basic strength and aerobic fitness.
Experience: intermittent training in the past six months; limited technique practice.
Time: three sessions a week, 35–45 minutes each; Wednesday may change.
Equipment: light dumbbells, a resistance band, and a mat at home.
Preference: avoid jumping. If a medical judgment is needed, tell me to seek a qualified professional.
First list missing safety and schedule information. Then propose a normal session and a 20-minute fallback.
Create my training profile before proposing a demanding program.
Goal: build a consistent habit while improving basic strength and aerobic fitness.
Experience: intermittent training in the past six months; limited technique practice.
Time: three sessions a week, 35–45 minutes each; Wednesday may change.
Equipment: light dumbbells, a resistance band, and a mat at home.
Preference: avoid jumping. If a medical judgment is needed, tell me to seek a qualified professional.
First list missing safety and schedule information. Then propose a normal session and a 20-minute fallback.
A useful profile separates what is known from what is unconfirmed. “Wednesday is unreliable” is a planning fact; “jumping will cause injury” cannot be inferred from a preference alone. Ask the agent to make ATHLETE_PROFILE.md editable, with goals, available time, equipment, experience, recovery context, reported limitations, data sources, and an update date. This profile is a living input. It should change when your schedule, equipment, or health circumstances change.
Make the plan usable on a bad day
The project’s loop moves through assessment, planning, execution, data review, and adjustment. In the first cycle, the key test is whether a session fits real life. Each training card should state the activity, sets or duration, an understandable effort target, rest, alternatives, and conditions for shortening or stopping. The table below is an output-quality example, not an exercise prescription:
| Card field | Testable output | Vague output to challenge |
|---|---|---|
| Goal | Complete a basic full-body session and record what happened | “Crush today’s workout” |
| Time | Standard 40 minutes; 20-minute fallback | “Train when you can” |
| Equipment conflict | Offer a comparable movement category when equipment is unavailable | Skip everything |
| Effort | Describe a conservative, recordable perceived effort | Every set to the limit |
| Acceptance | Log completion, elapsed time, effort, and unusual symptoms | Mark “done” with no record |
The image shows exercise order, sets and rest. A shorter fallback must be requested separately; pictured movements and loads are not recommendations for an individual reader.
Ask the agent what happens if only 20 minutes remain. It should preserve the most relevant movement categories and explain what was removed. An unusual pain or warning symptom should stop the automated recommendation; completing the checklist is not more important than a safety signal. Nor should a missed Wednesday automatically trigger a double session on Thursday. Review the whole week’s completion and recovery first.
Use a small manual log before adding integrations
Without records, an agent can produce a plausible plan but cannot check whether it worked. A minimum log needs the date, planned session, actual work, duration, perceived effort, and any unusual feedback. The numbers here show format only:
Date: 2026-10-15
Plan: basic full-body session / 40 minutes
Actual: 32 minutes; substituted a resistance-band variation for exercise 3
Main movements: name | sets x reps | load or resistance | perceived effort
Recovery context: about six hours of sleep; felt tired
Unusual symptoms: none / describe location, timing, and whether worsening
Next constraint: only 25 minutes available on Friday
Date: 2026-10-15
Plan: basic full-body session / 40 minutes
Actual: 32 minutes; substituted a resistance-band variation for exercise 3
Main movements: name | sets x reps | load or resistance | perceived effort
Recovery context: about six hours of sleep; felt tired
Unusual symptoms: none / describe location, timing, and whether worsening
Next constraint: only 25 minutes available on Friday
After each session, ask the agent to summarize recorded facts, identify missing fields, and write one question to test next time. One log should never become a confident claim about your ideal training volume. The repository describes “training DNA” as an evolving, evidence-based hypothesis. It says four to eight weeks of structured data are generally needed for a useful individual pattern; sparse records warrant only preliminary observations.
The Skill can keep four connected artifacts: ATHLETE_PROFILE.md for goals and constraints, CURRENT_PROGRAM.md for the current cycle, TRAINING_DNA.md for patterns supported by repeated records, and DECISION_LOG.md for each change and its rationale. Their value is in preserving why a plan changed, not merely saving a schedule that will soon be stale.
Review four weeks without overreading a single data point
The project can summarize manual records or Xunji data into frequency, movement distribution, performance, and consistency views. It also describes a local HTML dashboard. The picture below illustrates how to read a period view; it is not evidence of a reader’s results.
The image shows an approximately three-month trend. Recalculate a four-week pilot from your own four weeks of logs. Check coverage before trends; a chart alone cannot explain a lower week.
Use this review sequence:
- Coverage: How many sessions were planned, completed, and recorded? Missing records are not proof that a session did not happen.
- Adherence: Did the shorter version keep a busy week viable? A regularly skipped day may indicate a scheduling problem rather than poor motivation.
- Like-for-like performance: Compare similar movements, equipment, and perceived effort. A heavier number from an unrelated movement is not progress evidence.
- Recovery and warning signs: Repeated fatigue or worsening discomfort should reduce confidence in automatic progression and may require professional assessment.
- Controlled adjustment: Change one or two variables, such as the training day or session duration, before rewriting frequency, exercises, and intensity together.
Work through an illustrative four-week log
Suppose a person planned three sessions each week. The numbers below are invented to show the calculation; they do not imply a recommended training frequency or outcome.
| Week | Planned | Completed | Question raised by the log |
|---|---|---|---|
| 1 | 3 | 2 | Wednesday was missed because of a schedule conflict |
| 2 | 3 | 3 | One session used the 20-minute fallback |
| 3 | 3 | 2 | Another Wednesday was missed because of a schedule conflict |
| 4 | 3 | 3 | All three sessions have complete logs |
Across four weeks, 10 of 12 planned sessions were completed: 10 ÷ 12 ≈ 83%. This is an adherence measure, not evidence that the workouts were too hard or recovery was poor. Since both misses occurred on Wednesday, the next cycle might move that session to a more reliable day, or use the short version. Keep recording with the same definitions before judging the change. If the shorter session lacks exercise and duration details, mark it “completed, details missing” rather than treating it as a fully comparable training volume.
An agent could write a traceable entry in DECISION_LOG.md:
Observed: 10 of 12 planned sessions completed; two Wednesday sessions missed due to scheduling.
Unknown: whether fatigue, exercise difficulty, or movement choice played a role.
Change for next cycle: move Wednesday to Saturday; do not increase other variables at the same time.
Check: record planned sessions, actual completion, and recovery feedback for another four weeks.
Observed: 10 of 12 planned sessions completed; two Wednesday sessions missed due to scheduling.
Unknown: whether fatigue, exercise difficulty, or movement choice played a role.
Change for next cycle: move Wednesday to Saturday; do not increase other variables at the same time.
Check: record planned sessions, actual completion, and recovery feedback for another four weeks.
If the log never recorded why a session was missed, do not invent a scheduling cause. Ask for the missing context first. A decision log is useful precisely because it makes that uncertainty visible.
A shareable report can communicate a period of training. Its claims should still trace back to the covered dates and underlying logs.
This prompt tests whether the agent distinguishes observation from speculation:
Review only the four weeks of logs I provide.
List the date range, missing records, and actual completion rate first.
Separate observed facts from hypotheses. Suggest no more than two changes for the next cycle.
For each change, explain why and how it will be evaluated in four weeks.
If evidence is sparse, do not claim my “training DNA” is confirmed.
Update CURRENT_PROGRAM.md and DECISION_LOG.md, preserving their prior versions.
Review only the four weeks of logs I provide.
List the date range, missing records, and actual completion rate first.
Separate observed facts from hypotheses. Suggest no more than two changes for the next cycle.
For each change, explain why and how it will be evaluated in four weeks.
If evidence is sparse, do not claim my “training DNA” is confirmed.
Update CURRENT_PROGRAM.md and DECISION_LOG.md, preserving their prior versions.
Connect Xunji only when the data flow is clear
Manual logs are enough for a pilot. Connecting Xunji is optional. The project documents reading data, analyzing it, previewing changes, and writing back after user confirmation. Before connecting, check the app’s current API documentation and the installed version of the Skill. Do not place an Open API Key in a public chat, shared document, or screenshot. The repository says credentials are stored in protected local storage, but your actual host environment and installed version still need checking.
Begin with read-only access to a short date range. Compare the dates, number of sessions, and movement names with your own records. For any write-back, inspect a preview that includes the correct training IDs and dates, then confirm explicitly. Training records can contain health-related personal information. Before sharing a report, remove names, contact details, location, face photographs, and unneeded raw records. A team pilot should document who can read or edit data, where credentials are kept, and how caches can be removed.
Movement photos illustrate settings, not personal suitability or technique clearance.
Apito: keep the model endpoint separate from training-data rights
Apito can be relevant as a Claude API service endpoint, but the ability to answer a training question does not grant Xunji read access, write-back permission, or local credential protection. If the host agent accepts a custom endpoint, first test a redacted sample log, failure handling, and visible billing. Assess the Skill connector’s authorization and write-back confirmation independently. Native integration with Apito is not claimed here; verify API compatibility before deployment.
A practical acceptance checklist
| Check | Passing result | Reason to redo the workflow |
|---|---|---|
| Profile | Goals, constraints, and unknowns are explicit | Demanding plan before relevant safety questions |
| Day-of adjustment | Shorter version and alternatives have reasons | Repeats the original plan unchanged |
| Logs | Dates, completion, effort, and unusual feedback are preserved | Treats missing data as failure |
| Review | States date range and separates facts from inference | Declares a precise profile from one record |
| Safety | Stops automatic advice for warning symptoms | Treats pain as a target to push through |
| Data rights | Read-only pilot and preview before write-back | Edits history without confirmation |
If the basics fail, organize the logs and four artifacts before adding more devices or generating elaborate reports. The useful outcome is a plan you can adapt, a record you can trust, and a reason for each change.
Sources and scope
- Healthy Fitness Coach Skill repository: installation, described features, data flow, and safety limits.
- WHO physical activity guidance: population-level adult activity benchmarks.
The sample card and log demonstrate a workflow. They are not individualized exercise or medical advice.
FAQ
Can I review training without connecting Xunji?
Yes. Consecutive manual logs with dates, actual sessions, effort, and unusual feedback support basic frequency and adherence review. A connector mainly reduces the work of importing and organizing records.
Is one week enough to establish a “training DNA”?
No. One week can support preliminary questions and a list of missing information. The project says useful personal patterns generally require four to eight weeks of structured records and continued validation.
Should I finish a session if an agent’s suggested movement causes pain?
Stop the activity that produces unusual pain. Chest pain, fainting, unusual palpitations, marked breathlessness, or progressive numbness or weakness call for professional medical assessment, not an agent’s remote go-ahead.
Can a generated report be posted publicly?
Check that its numbers trace to your logs, then remove unnecessary personal details such as name, location, contact information, and face images. A report documents records; it is not proof of a medical outcome.
Will connecting data automatically rewrite my training history?
It should not by default. Follow the documented read-and-preview flow, check dates and training IDs, and confirm any write-back explicitly. Verify the actual permissions of your installed version and host application.
