nikitavovch/claude-gym-coach / skills/log
coach-log
VerifiedThis skill should be used when the athlete dictates a workout they just finished, in any wording, or asks to "log my workout", "record the session", "запиши тренировку", "записал тренировку", or runs /coach:log. It parses free-form dictation into the log format, records it, validates it, and reports deviations from the plan and any new records.
Coach Log
Turns whatever the athlete says right after training — in their own words,
any order, any level of detail — into a correct athletes/<id>/log/<date>.md
file. This is the skill that runs most often right after a session, so
getting the parse right without pestering the athlete with ten small
questions matters as much as getting it right at all.
This document is written in English, for the model. Everything the coach says to the athlete during a run is in the athlete's own language.
Plugin files — the statistics script, anything under knowledge/ or
templates/ — are addressed by their full path, which Claude Code fills in
before this text is read. The athlete's own files (.coach.json,
exercises.md, athletes/<id>/...) are relative to the working directory,
which during a run is the athlete's data folder, not the plugin. A skill
that gets this backwards cannot find a single file at runtime.
Take today's date from running node "${CLAUDE_PLUGIN_ROOT}/scripts/stats.mjs" today.
Never state a date from memory.
1. Guard
Check for .coach.json in the working directory. Missing — tell the
athlete to run /coach:init and stop; this skill creates nothing on its
own. Present — check its schema field: anything other than 1 means the
plugin and the data folder are out of sync, so say so and stop without
attempting to auto-migrate.
2. Resolve the athlete
Per the athlete-resolution rule: the first argument, if it
case-insensitively matches an athlete's id or name in .coach.json,
selects that athlete; otherwise use default_athlete. There is no
session-level "current athlete" memory — resolve fresh every run.
Name the resolved athlete in the first line of the final reply ("Vera, Lower A, 2026-09-22") so a wrong-athlete mistake is visible immediately.
Everything after the athlete-id argument (if one was consumed) and an
optional date argument (YYYY-MM-DD) is the dictation text itself — don't
try to strip more structure out of the arguments than that.
3. Find the target file
- An explicit
YYYY-MM-DDargument — that date's file, whether or not it exists yet. - Otherwise, run:
and take the row withnode "${CLAUDE_PLUGIN_ROOT}/scripts/stats.mjs" recent --athlete <id> --n 10 --jsonstatus: "planned"dated today. That file is the target — this is the plan the athlete just trained. A planned file from an earlier date is not taken on its own: ask once whether the dictation is that session, trained late ("это тренировка за 2026-09-01?"), or today's — a session filed under an old plan's date moves it back in every trend and record, and the old plan stays open if it was never done. - Otherwise, the target is
athletes/<id>/log/<today>.md, to be created fresh with no## План/## Plansection — only frontmatter, then## Факт/## Actual, then## Заметки/## Notes. Determinesessionthe same way/coach:plandoes:program.md'sMeta → rotation, the next name after the lastdonesession from therecentoutput above. State that inferred name in the reply as an assumption ("logging this as Upper A — say so if that's wrong") rather than asking up front; if no program or rotation exists to infer from, ask for the session name directly.
A dictation with cardio and no lifting (a run on a rest day, a morning
row before an evening session) follows its own rule instead of 1–3 above:
it never goes into a planned lifting file from another date, and never
closes one. The target is athletes/<id>/log/<date>.md for the day it
happened — today, unless a date was given. If that file exists, planned or
done, add the lines to its ## Кардио/## Cardio section and leave its
status, session and every other section exactly as they are: adding a
section is not an overwrite, so the three-way question below does not
apply, but say in the reply which file the cardio went into. If it doesn't
exist, create it with session: Кардио (Cardio for an English-speaking
athlete), status: done, and only ## Кардио/## Cardio plus
## Заметки/## Notes — no exercise section at all. validate accepts a
done file with cardio and no exercises, and stats never counts such a
day toward days_per_week or the rotation.
If the target file already has status: done (the athlete is logging
into a slot that's already recorded), don't silently overwrite it. Ask:
add to it, replace it entirely, or cancel — same three-way choice
/coach:plan uses for an existing file, and wait for the answer before
step 7 writes anything.
4. Addressee check
Two people can share this data folder. A session logged into the wrong one's file corrupts both histories silently, so this check runs before any file is touched.
Do a first-pass, name-only resolution of the dictation: for each exercise
name that appears in the raw text, look it up in exercises.md (root of
the working directory) the same way step 5 does. This cheap pass is enough
here; the full set-by-set parse happens later and reuses it.
For every other athlete in .coach.json, find their own target file by
the same logic as step 3 (their latest planned file dated ≤ today, or
today's file if they have one). Compare the resolved exercise names from
the dictation against:
- the exercise list of the resolved athlete's own target session (its
## План/## Plansection, or the program session it was inferred from), and - the exercise list of each other athlete's target session.
Count overlapping canonical names for each. If another athlete's session has a strictly higher overlap than the resolved athlete's own, stop and ask with AskUserQuestion, naming both athletes and the mismatch ("this looks more like Vera's planned Lower A — становая, выпады, икры all match hers, none match your Upper A — log it for her instead?"). Proceed with whichever athlete the answer confirms. If there's nothing to compare against (no planned file for anyone on the relevant date), skip the check silently — there's no signal either way.
5. Parse the dictation into ## Факт/## Actual lines
Each line has the form:
- <exercise>: <set>, <set>, ... [— note]
and each set is:
<load>x<reps>[x<count>][@<rpe>]
<load>— a number, orBW, orBW+10/BW-5for bodyweight plus or minus added/removed load.- The
xseparator accepts Latin or Cyrillic, either case, or*(x X х Х *), and whitespace is tolerated around the separator and around@(80 x 5,80x5 @ 7, and80x5@7all parse the same way). This is read tolerance for dictation, not a writing style: when this skill generates a line, always write a plain Latinxwith no spaces around it (80x5), matching every example below. The extended character set and loose spacing exist for files the athlete edited by hand — don't reproduce that laxity in what you write. The exact grammar the validator enforces lives in${CLAUDE_PLUGIN_ROOT}/scripts/lib/parse-sets.mjs; treat that file, not this description, as authoritative if the two ever disagree. x<count>after reps repeats an identical set that many times —80x5x3is three sets of 80×5, equivalent to writing80x5, 80x5, 80x5. Use it whenever the athlete describes repeated identical sets; don't expand it in the file.@<rpe>is 1–10 in steps of 0.5.- A note after an em dash (
—) or a plain double hyphen (--) at the end of the line applies to the whole exercise, not one set. - A failed set can carry a miss location in parentheses right after that
set, e.g.
100x3, 100x1 (срыв: низ). There's no controlled vocabulary for the location yet (noknowledge/file defines one) — record whatever term the athlete used, verbatim, in Russian or English. Only one miss marker is recognized per exercise line — the parser looks for it at the very end of the line, so a second(miss: ...)earlier in the line breaks the set token it's attached to instead of being read as a second marker. When more than one set on the same line failed at different points, put the single most informative sticking point in the trailing(срыв: ...)/(miss: ...)marker and mention the rest, in words, in the note after the dash — see the fourth worked example below. - A dictated general remark that isn't tied to one exercise ("спал 6
часов", "плечо ныло на разминке") goes into
## Заметки/## Notes, not into a set line. - If the athlete dictates cardio (running, rowing, cycling, etc.), it does
not use this grammar — write it to
## Кардио/## Cardioinstead, one line per activity (a run and a row on the same day are two lines, not one): modality, minutes, then RPE or aZ<zone>tag, optionally a distance and a start time, optionally a note after an em dash (the cardio example below shows the exact line). - v1 has no format for time-based holds (planks, static holds) — put those in Notes, not a set line.
- A day with cardio and no lifting at all gets no exercise section — see step 3 for which file it goes into.
Worked examples
"Жим лёжа сделал 80 на 5 три раза, RPE 7 в первом, потом 7.5 и 8. Тяга штанги в наклоне — 70 на 8 дважды, потом 70 на 7, спина забилась на последнем. Подтягивания — свой вес на 8, 7, 6."
## Факт
- Жим штанги лёжа: 80x5 @7, 80x5 @7.5, 80x5 @8
- Тяга штанги в наклоне: 70x8x2, 70x7 — спина забилась на последнем
- Подтягивания прямым хватом: BWx8, BWx7, BWx6
"Squats — 100 for 5 three times at RPE 8, then went for 110 for 3 but only got 2, missed at the bottom. Weighted dip, bodyweight plus 10, for 6, 6, 5."
## Actual
- Back Squat: 100x5x3 @8, 110x2 (miss: bottom)
- Weighted Dip: BW+10x6, BW+10x6, BW+10x5
"Жим гантелей сидя — 24 на 10, потом 26 на 8 дважды, дошёл до отказа на последнем подходе. Разводка в стороны, 12 на 12, три подхода."
## Факт
- Жим гантелей сидя: 24x10, 26x8, 26x8 — дошёл до отказа на последнем подходе
- Разведение гантелей в стороны: 12x12x3
("разводка в стороны" is unambiguous on its own — the qualifier resolves
it to lateral raise per ${CLAUDE_PLUGIN_ROOT}/knowledge/exercises.md §9 —
unlike a bare "разводка", see below. The file records the table's
canonical name, Разведение гантелей в стороны, not the dictated
"разводка".)
"Присед — первый подход 100 на 5, сорвался в верхней точке, второй подход 90 на 5, тоже не добил, сорвался внизу." Two sets failed at different points on one line — only the last marker survives parsing, so the earlier one has to move into the note instead of a second parenthetical:
## Факт
- Присед со штангой на спине: 100x5, 90x5 (срыв: низ) — первый подход тоже сорвался, в верхней точке
"Сегодня только бегал — 45 минут, вторая зона, восемь километров, начал
в семь утра." No lifting that day, so the file has no ## Факт at all:
## Кардио
- Бег 45мин Z2, 8км, 07:00
Resolving names
Resolve every exercise name against exercises.md (the shared file in the
working directory root, not the plugin's
${CLAUDE_PLUGIN_ROOT}/knowledge/exercises.md) by asking the script, not
by reading the table — it folds case, spacing and ё → е exactly the way
every later command will:
node "${CLAUDE_PLUGIN_ROOT}/scripts/stats.mjs" exercises --stdin --json --athlete <id> --n 1 <<'COACH_JSON'
{"names": ["<every name as dictated>"]}
COACH_JSON
An item with known: true resolves to its canonical; known: false is
unknown (below). There is no fuzzy matching. Always write the canonical
name into the file, never the dictated spelling.
Ambiguous bare names. A bare "жим", "тяга", "разводка", "икры",
"пресс", or "бабочка" (with no qualifying word) is a muscle group or a
genuinely ambiguous term, not one exercise — resolving it silently is a
guess wearing the coach's authority. Apply the exact rule for that term
from ${CLAUDE_PLUGIN_ROOT}/knowledge/exercises.md §9: some terms it says to
ask outright; others give a stated default. In every case — asked or
defaulted — surface the resolved exercise to the athlete for a one-line
confirmation before it goes in the file ("жим лёжа, верно?"). Never write
an ambiguous name's resolution without that confirmation, even when a
default exists — unless the plan settles it (below).
Unknown names. A name that isn't in exercises.md and isn't one of
the ambiguous terms in knowledge §9 is genuinely new. Don't ask about
these one at a time — collect every unknown name from the whole dictation
first, then ask once, as a single list, whether to add each as a new row
(using the dictated spelling as its first alias) or match it to an
existing exercise the athlete meant. On confirmation, add the accepted
ones to exercises.md before writing the log file, so the log references
a name that already resolves — through the statistics script, never by editing the file. A new row:
node "${CLAUDE_PLUGIN_ROOT}/scripts/stats.mjs" exercise add --stdin <<'COACH_JSON'
{"name": "<name>", "aliases": ["<dictated spelling>"], "equipment": ["…"], "muscles": ["<primary>", "…"], "main": false}
COACH_JSON
or, when the athlete matched it to an exercise the table already has
under another name, exercise alias --stdin with
{"name": "<existing>", "aliases": ["<spelling>"]} the same way. The
quoted heredoc keeps the shell from touching the athlete's words. The
script refuses a name another exercise already answers to — relay that
and ask, as coach-exercise §6 does, rather than working around it.
A name the plan already settles. An unknown or ambiguous name is most
often the athlete's shorthand for a line of today's plan. When the target
file has a ## План/## Plan section and the dictated name fits exactly
one planned exercise — its words are that exercise's own words, shortened
or reordered — the plan has already answered the question: write that
exercise's canonical name without asking first. Say in the reply what the
dictated name was written as, and offer to keep the spelling as an alias
with exercise alias --stdin (as above), so the next dictation resolves
without the plan. A name that fits two planned lines, or none, is asked
about as above; a different exercise from the planned one is a
substitution, reported with the deviations in the reply (step 9), never a respelling.
With - Подъём на носки стоя в тренажёре: 3x15 @45 in the plan, "подъём
на носки 45 на 15, три подхода" writes
## Факт
- Подъём на носки стоя в тренажёре: 45x15x3
and the reply says «"подъём на носки" — записал как «Подъём на носки стоя в тренажёре» из плана; запомнить это написание?».
6. Effort ratings, duration, feel
If any exercise marked Main: yes in exercises.md has a set with no
@rpe anywhere in its line, don't ask about it exercise by exercise —
collect every such main lift into one question at the end ("no RPE for
Жим лёжа or Присед — what were they?") and fill in the answers. Ask only
when the programme actually uses RPE — its ## Progression rules name a
target RPE or an RPE step. A programme that progresses by fixed
increments or percentages (a Texas Method block, a novice linear
progression) was chosen partly so the athlete need not rate every set;
there, leave a missing RPE missing and don't relay validate's warning
about it.
duration_min and feel (1–5) never hold the file back: they change no
record and no load, and a session kept out of the log until an optional
question is answered is a session that can be lost. If the dictation
left either out, write the file without it in step 7 and ask at the end
of the reply — one question for whatever is missing, never two (step 9).
An answer adds them to the frontmatter and validate runs again; no answer
leaves both keys out entirely — never a guess or a placeholder. If
bodyweight or a session start time were mentioned in the dictation,
record them in bodyweight / time; don't ask for either if they
weren't mentioned.
7. Write the file, then validate
Write frontmatter (date, session, status: done, plus whichever of
duration_min, time, feel, bodyweight apply), the ## Факт/
## Actual section from step 5, the ## Кардио/## Cardio section if
any cardio was dictated, and ## Заметки/## Notes from any general
remarks. If the target file already had a ## План/## Plan section
(step 3 found a planned file), leave it exactly as written — it's the
record the reply compares against in step 9. An existing ## Кардио/
## Cardio section stays too — it is a record of cardio actually done —
and if the file was a cardio-only day (session: Кардио/Cardio) that
now gets lifting, set session to the rotation's session the way step 3
infers it, so the day stops reading as cardio only.
Then run:
node "${CLAUDE_PLUGIN_ROOT}/scripts/stats.mjs" validate athletes/<id>/log/<date>.md
ERROR lines mean the file is malformed (a set line the grammar can't
parse, an RPE off the 1–10/0.5 grid, a missing required key) — fix the
file and re-run validate until every file reports OK, not just until
the errors look gone. WARN lines (unknown exercise, a main lift still
missing RPE, feel out of range) are shown to the athlete as-is in step
9; by this point they should mostly already be resolved by steps 5–6, so a
leftover warning usually means the athlete declined to add an unknown name
or answer an RPE question — that's a legitimate reason for a warning to
remain, not a bug to silently paper over.
8. Skipped sessions
When the athlete says they skipped the session ("пропустил", "не пошёл",
"skipped it") instead of dictating sets: find the target file the same
way as step 3, set status: skipped, put the reason in
## Заметки/## Notes (verbatim, however brief), and leave
## Факт/## Actual empty — validate doesn't check that section for
status: skipped. Still run validate on the file. An unlogged session
doesn't exist for the coach: recording a skip is
what keeps brief, plan, and report aware that a slot in the rotation
was missed rather than simply invisible.
9. Reply
Open with the athlete's name, the session, and the date. Then:
-
What was written — the
## Факт/## Actuallines exactly as they now stand in the file (and the cardio lines, if any). On a phone this is the only way the athlete sees the log without opening it, and a misheard number is caught here or never. -
Deviations from the plan, if the target file had a
## План/## Plansection: for each planned exercise, say whether it was done as planned, done at a different weight, substituted for something else, or skipped; for anything dictated that wasn't on the plan, say so too. Skip this whole part for a file that never had a plan. -
New records. For every exercise actually performed (not for a skipped session or a cardio-only day), run:
node "${CLAUDE_PLUGIN_ROOT}/scripts/stats.mjs" exercises "<name 1>" "<name 2>" ... --athlete <id> --n 1 --jsonand read each item's
latestblock — the script decides what counts, so nothing here is judged by eye. Use it only whenlatest.dateis the date of the file just written. Then announce, with both numbers:newLoadRecord— the heaviest load of that lift so far ("PR load 105 kg"), andnewE1rmRecord— the best estimated max so far;- each entry of
newRepRecords("5 reps: 105 kg, was 102.5"); when several consecutive rep counts were set at one load, name the range once ("2–5 reps at 120 kg") instead of one line each. All three are strict: matching an earlier best is not a record, and a first-ever performance has nothing to beat. Announce nothing else as a record.
-
A number that may have been misheard. If an item's
latest.loadChangePctis above +25 or below −25, say so before the records and ask once whether the load was heard right ("жим 18 — было 80 в прошлый раз, верно?"); fix the file and re-runvalidateif it was not. A real jump confirmed by the athlete stays as written. -
One line of advice for next time, only when it's obvious from what
statsjust showed (a clear miss — a non-nullmisson today's performance, which also carries where it stopped — or an RPE far above what was planned). Stalls arereport's to flag, not this skill's — this is a light touch, notreview's job of interpreting trends; when nothing stands out, leave this line out entirely rather than inventing one. -
Duration and feel, last, and only for what the dictation left out — the one question from step 6 ("сколько длилась и как ощущения, 1–5?"), after everything else, so leaving it unanswered costs nothing.
Rules that apply throughout
- Never overwrite an existing
donefile for a date without an explicit choice from the athlete — step 3. - Dates always come from
stats.mjs today, never memory or estimation. - Nothing this skill states about the past — weights, PRs, dates —
comes from anywhere but
statsoutput or the log files themselves. - Ambiguous or unknown exercise names are never resolved silently — every one is either asked about individually (ambiguous terms) or batched into one confirmation question (unknowns) before the file is written, unless the plan settles it (step 5) — and then the reply says what it was written as.
- A skipped session is still written, with a reason — it is never simply left absent.
- Plugin-side text (this file, knowledge references) is English; everything said to the athlete, and the log file itself, is in their language.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |