agentleFS
fortunto2fortunto2/solo-factory/skills/mirror/SKILL.md

solo-mirror

Skill18 starsChanged 4 months agoMITView on GitHub

Weekly reflection that compares what a person says they focus on with what the data shows they did, names at most three repeating patterns with counts and dates, sets one experiment, and shrinks the load (cut list, three items for this week). Read-only; returns a report and ledger entries, the caller applies them. Use when user says "weekly mirror", "reflect on my week", "what do I keep doing without noticing", "am I working on what I said", "lower my load", "зеркало", "разбор недели". Do NOT use for a pipeline retrospective (use /retro), for checking memory notes (use the dream workflow), or for anything personal that belongs in a private journal.

The file

We split this skill into the rules it gives an agent

11 rules. About 0 are common, 7 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Write the report from the numbers, nothing else.

    line 44
  2. Very specific:

    Check the form: mirror_collect.py lint <report.md> (length, em dashes, at most 3 numbered observations).

    line 52
  3. Very specific:

    Goals disagree: If goal_conflicts is non-empty (a goals file says a decision is pending) or two goals files contradict each other, the …

    line 60
  4. Very specific:

    Effort is stated, never guessed: Every "this week only" item and every parked protected item carries effort (the item's own time field, or …

    line 62
  5. Distinctive:

    No adjectives about the person, no diagnosis, no psychology, no motives.

    line 102
  6. Distinctive:

    Never print identifiers (passport, tax, account numbers), dates of birth or health data.

    line 104
  7. Very specific:

    Never read paths the config does not list.

    line 105
  8. Distinctive:

    Read-only: no sends, no edits, no task changes.

    line 106
  9. Distinctive:

    Don't turn the report into advice about how to live.

    line 117
  10. Distinctive:

    Don't add a fourth observation because the third was weak.

    line 118
  11. Distinctive:

    Don't count assistant-written commits as attention without saying that agents commit too.

    line 119

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

mirror: a weekly reflection with numbers

Reads files, git and past prompts, returns a short report. It lowers mental load: every open item either has a reason, is on a cut list, or is explicitly "not this week". It never writes, sends or edits anything; the caller saves the report and runs ledger-append.

Inputs (all paths in one config file)

InputConfig keyWhat it is
Stated goalsgoals_filesmarkdown the person wrote: the plan, the focus rule (for example one product plus one keystone goal). The model reads these; the script only lists them
Task registryregistry_diropen/*.json topics with parents, updated, items
Todo liststodo_listsyaml lists with status, due, optional purpose: "<goal slug or ?>: <reason>", dated due_reason* keys
Projectsprojects_registryyaml projects: [{name, path}]; git logs are read per project
Past promptssessions.dirClaude Code transcripts; user turns only, never assistant text; exclude_dir_substrings are never opened; automation_regex drops headless jobs
Draftsdrafts_cmdoptional command printing JSON [{ref, title, drafted_at, reschedule_dates}]
Focusfocus, project_goal, goal_prioritywhich projects/goals are the declared focus, which goal each project serves
Ledgerledgerjsonl of patterns and experiments

Template: references/mirror.example.yaml.

Workflow

  1. Collect. This is the only step that touches data, and it is deterministic:
    uv run ${CLAUDE_PLUGIN_ROOT}/skills/mirror/scripts/mirror_collect.py collect --config <mirror.yaml> --days 14 --json
    (--end YYYY-MM-DD to reproduce a past week.) Read the stated goals files once, for the words of the focus rule; do not search anywhere else.
  2. Write the report from the numbers, nothing else. Three questions, same every week:
    • Said vs did. said_vs_did: share of sessions and commits per project and per goal against the stated focus. One sentence with the two numbers.
    • What repeats. Only entries in patterns (the script already applied the rule: at least 3 dated occurrences and a number). At most 3, the largest count or longest weeks_in_a_row first. Each: what, the count, the dates, where the evidence is (refs). If weeks_in_a_row is 2 or more, say "N weeks in a row". If patterns is empty, write "повторов с тремя датами нет" and stop; do not invent one from below_threshold.
    • One experiment. Verdict on last week's from experiment_last_week.result (met / missed / unknown, with the actual number), then one new experiment for next week: one behaviour, checkable from the same data, as {metric, op, value} using a key in metrics (for example share_focus_sessions >= 0.4).
  3. Load reduction, part of the report, not advice:
    • Cut list from cut_list: open items with no goal behind them or silent for the configured days. One line each: ref, why, proposal (drop, park until a date, or delegate to an agent). Keep the script's proposal unless the title shows it is wrong.
    • This week only: at most 3 items from this_week_candidates, each with its reason. Everything else is named in one line as "не на этой неделе".
  4. End with ONE question the person may take to a private journal or therapy context. A question about the pattern, not a diagnosis. This skill never reads such a place and never asks for it.
  5. Check the form: mirror_collect.py lint <report.md> (length, em dashes, at most 3 numbered observations).
  6. Return: the report, and a JSON list of ledger entries for the caller to append: {"type":"pattern","id","week","count","refs"} per reported pattern and {"type":"experiment","week","text","metric","op","value"} for the new experiment. The caller runs ledger-append.

Rules the collector supports

  • Goals disagree. If goal_conflicts is non-empty (a goals file says a decision is pending) or two goals files contradict each other, the FIRST line of the report asks which goal is current. Do not measure against the older one.
  • Protected items are never cut. An item whose text matches protected_regex (compliance, verification, legal, tax, penalty) is in protected_not_cut, not in cut_list. It may be parked with a date or sized; the report never proposes dropping it.
  • Effort is stated, never guessed. Every "this week only" item and every parked protected item carries effort (the item's own time field, or minutes quoted from its notes). When it says "объём неизвестен", sizing the item is the first step.
  • Noise. noise_waiting_already_answered lists topics waiting on a person while one of their own items records that the person answered. Report them under one line "Шум: ждём ответа, который уже записан".
  • Buckets. A pattern may carry bucket: finish (a complete artefact that was not sent) or middle (work stopped partway). State the bucket the data puts it in; do not decide for the person.
  • Done log. done_log is what was closed, sent or published in the window. A voice block in the config may ask to open the report with it.
  • Portfolio goal. said_vs_did carries two numbers when focus.top is set: the share of attention on the whole earning portfolio (focus.projects) and the concentration inside it (focus.top over the portfolio), plus share_defocus_* for projects listed in focus.defocus (neither earning nor required by a deadline). Report them as plain numbers.
  • Releases first. releases (from releases_cmd, an adapter that prints [{ref,title,version,date}]; the script never calls a store API) opens the done log: versions shipped, then sends, then the rest. known: false means the source is unavailable (say «источник недоступен»), never «релизов нет»; a known source carries measured, the date to print beside the line. Commits and sessions are never presented as progress.
  • Journal. collect first brings an append-only journal of typed prompts up to date (journal in the config; mirror_collect.py journal). One line per prompt: time, local day, project, session, message uuid, character count, label, confidence, who labelled it, label-set version. Never text. A cursor file keeps the byte offset per transcript, so a run reads only new bytes and a run with nothing new makes no model call. A truncated or replaced transcript is re-read and the uuid key prevents duplicates. Labels (one per prompt): task sets new work, steer corrects or adds to work in progress, decide answers or approves, kick nudges without new information, watch asks to look at something running, think thinks aloud, other a bare paste. A bare "continue" is labelled by rule; the rest by a headless model call in batches (prompt text is truncated and goes only to that call; excluded directories are never read). Labels change only when the label-set version changes, and then only on demand (journal --relabel).
  • Second profile as time only. work_profiles may name a separate profile whose transcripts are counted for time: only timestamp, message uuid and character count are read; its rows are project: work, mode: work, by: none, and no text, directory name or session id reaches the journal, the cursor, a model call or a log.
  • Local time follows the person. energy.timezones is a dated timeline (from, tz, optional place). Build it from evidence such as the committer offsets in git history and trip records; list approximate boundaries as uncertain. Zone-change days and the day after are marked as travel days, not judged.
  • Energy is reported in blocks, not hours. Hours stay as the raw layer (own_h, work_h, overlap_h, total_h by 15-minute bucket, overlap counted once). Three derived numbers lead: (1) creative blocks, runs of at least 45 minutes on one own project where the top label is task, think or steer and nothing of another project, of the work profile or of a routine label interrupts (one silent bucket is tolerated, config block_gap_buckets); the budget is blocks_per_day (default 3), and a day over capacity lowers the next days' suggestion (debt_blocks, suggested_blocks_today); (2) leak, the share of own active time whose top label is decide, kick or watch, plus project switches per active hour; (3) open loops, open items whose ball is on the person and that have no dated next step. A plan closes a loop, not the completion, so the cut list and the this-week list are presented as loop-closing. Background, not a claim about the user: demanding work is limited per day to about four hours in sessions (Ericsson et al. 1993); fatigue after a day of hard cognitive control has a metabolic account (Wiehler et al. 2022, Current Biology); switching leaves attention residue (Leroy 2009); making a concrete plan removes the intrusion of an unfinished goal (Masicampo and Baumeister 2011). Do not cite ego depletion or decision-fatigue studies. energy.json carries the same numbers, a per-day table and a per-week history (hours, hours per label, blocks, place). The energy subcommand recomputes from the journal in about a second; --ingest also reads new transcript bytes.
  • Fit to capacity. Each "this week only" candidate has cost_h (from its stated effort) and an optional kind (deep or light). The listed items must fit the week's remaining blocks; an item that does not fit is named as not fitting instead of being listed. Unknown cost stays "volume unknown".
  • Week template. Optional week_template: {blocks_by_weekday, autopilot_weekdays, work_expected_h, sunday_flag_h, balance} lets the person state their rhythm. Capacity per weekday replaces the flat blocks-per-day in the budget and the debt. Own kick/watch time on an autopilot weekday (plans set, an agent kicked) is counted separately as autopilot_h and excluded from the leak share. rhythm reports the balance of work and own hours, the last Saturday's work hours and the last Sunday's own blocks and kicks, flagging only above work_expected_h / sunday_flag_h; protected days, rest mode (exempt_days, from audit_cmd) and travel days never flag.
  • Work band. Optional work_week: {target_h, band_h, uncounted, midweek_template, work_extra_h}: hours of a second profile per ISO week by sessions only (a conditional measure; calls, reading and thinking are not in it), the last four weeks, state в полосе / ниже / выше with the gap to the band edge, нет данных for a week with no rows (never zero), manual work_extra_h added and shown separately. On Thursday and Friday the energy subcommand sets midweek_line only when the projection to Friday evening leaves the band; inside the band it is null.
  • Do the tools serve you? tools in the config reads three generic, read-only sources: items_cmd (JSON list of cards an assistant or bot put in front of the person: kind, state, first_seen, closed_at, shown_at, optional closed_by and close_reason), a checkins JSONL (energy taps {date, level}, drop reasons {date, reason}, effort answers {date, effort_min}), and audit_cmd (JSON counters: work_origin_leaks, mutes_added, rest_days, protected_days). Per week it reports cards created, done, dropped, expired, reaped (closed because the source died), median minutes from shown to closed, the busiest day against cards_per_day_cap, cards that arrived on protected days, and one noise ratio: the share of closed cards that were dropped with a noise reason (noise_reasons) or expired unseen. A non-zero work_origin_leaks is the first line of the section. A missing source is "unknown", never zero. directions is an optional checklist (id, title, sign, state: done|live-unverified|not-started, since); the report gives state counts and asks at most ONE live check per week, the oldest first. A proposal never covers a direction whose state is not not-started (it is already built or queued). At most three proposed changes, each one line with the number behind it and its effort; no proposal without a number. Week-over-week arrows (arrows) link focus and energy to the tools number, for example leak share against cards per day. Counts and dates only; anything deeper is a question the person takes elsewhere.
  • Optional voice block in the config: a list of extra rules the caller wants applied to the wording and layout (for example: done-log first, counts and dates only with no blame wording, finish-vs-middle question on one live item, a one-line definition of done and a first step for every this-week item, no "not now" state). The skill applies them literally and never invents them.

Report form (short, in the person's language)

Total at most 2400 characters (the card is one message; details live in the collector output and energy.json). Lead with the finding. Layout:

Зеркало <week>
Сделано: <releases, then sends, then the rest; "no data" when unknown>
Блоков за неделю: <X> из <Y> возможных, самый длинный <N> ч. На рутину: <P>% времени, <K> переключений в час. Открытых петель: <L> (было <M>).
Всего <X> ч: свои <Y>, работа <Z>, параллельно <W>.
Сказал / сделал: <focus share of sessions> и <of commits> против плана. <one line>
Повторы:
1. <what>: <N> раз, <dates>. <where to look>
Эксперимент прошлой недели: <met|missed|unknown>, <actual> при цели <target>.
Эксперимент на эту неделю: <one behaviour>, проверим по <metric>.
Срезать: <ref> (<proposal>), ...
Инструменты: <leaks line if non-zero>; карточек <N> (максимум <M> в день при лимите <cap>), закрыто <done>/<dropped>/<expired>, шум <P>%; предложения (до 3, с числом и объёмом)
Только на этой неделе (закрывает петли): 1) <deep|light>, <cost>: <item>, зачем: <reason>  2) ... 3) ...
Не на этой неделе: все остальное.
Вопрос для дневника: <one question>

Hard limits

  • At most 3 observations; each has a count, dates and a place to verify.
  • No adjectives about the person, no diagnosis, no psychology, no motives. Behaviour and numbers only.
  • A pattern with fewer than 3 dated occurrences is not reported, however vivid.
  • Never print identifiers (passport, tax, account numbers), dates of birth or health data; the collector redacts long digit runs in prompt excerpts, but check anyway.
  • Never read paths the config does not list. Sessions from private areas are excluded by exclude_dir_substrings; do not work around it.
  • Read-only: no sends, no edits, no task changes. Proposals only.

Gotchas

  • Share of sessions counts sessions with a typed user turn in the window, so a long session and a one-line session weigh the same; commits are the corrective. Report both.
  • A project outside the projects registry shows as (none); say so instead of guessing its goal.
  • Reschedules are read from dated due_reason* keys; an item moved without a dated note is invisible. Say "по датированным заметкам".
  • The first week has no ledger: weeks_in_a_row is 1 and there is no experiment verdict. Say that.

Don't

  • Don't turn the report into advice about how to live. The experiment is the only suggestion.
  • Don't add a fourth observation because the third was weak. Fewer is the right answer.
  • Don't count assistant-written commits as attention without saying that agents commit too.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.