agentleFS
KirKruglovKirKruglov/ARMARIUM/.claude/skills/calibration/SKILL.md

calibration

Skill1 starsChanged 5 months agoMITView on GitHub

Maintain the owner's calibration ledger — record dated forecasts, surface bets matured by their due-date for the owner's verdict, log the outcome the owner confirms, and recompute the per-topic accuracy map. Writes only to the agent's service zone; never verdicts outcomes itself and never goes to the web. Invoked by the armarium agent.

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 2

    1. dynos-fit/dynos-work★ 2same name
    2. bcvk/support-lead-kit★ 0same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository6 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read bets-ledger.md. date for the creation date.

    line 68
  2. Distinctive:

    Dedup: if the timeline already holds a bet with the same wording and the same due-date — don't duplicate.

    line 69
  3. Very specific:

    Read bets-ledger.md, select the open bets whose Due-date ≤ current date.

    line 96
  4. Distinctive:

    Do not rule the verdict for the owner and do not go to the web for facts.

    line 106
  5. Very specific:

    Read bets-ledger.md, gather all resolved bets.

    line 139

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

Calibration — bets ledger and calibration

Workflow over the owner's bets ledger: recording dated forecasts, surfacing bets matured by their due-date, logging the outcome from the owner's verdict, and recomputing the per-topic accuracy map of his judgment. One skill, three flows; the flow is chosen by the intent of the request.

The only write target is .claude/armarium/memory/bets-ledger.md. The date comes from Bash (date). No web; the outcome verdict always belongs to the owner.

When to use

  • Flow A — record a bet. The owner states a forecast ("I bet that…", "log this bet…").
  • Flow B — reconcile. "Which bets have matured / let's settle up / let's check the forecasts."
  • Flow C — calibration. "How's my calibration on topic X / show the accuracy map."

The run-mode fork applies only to Flow B (see below): an interactive session with the owner versus a headless run (cron / piggyback on a scheduled run).

Process indication

Status lines are needed only in interactive Flow B — at the gate where matured bets are presented and the owner's verdict is awaited. Format: [calibration] <emoji> <action>: <value> [<context>].

  • ⏸️ — waiting on the owner's verdict for the presented matured bets.
  • ❓ — clarification on ambiguous intent (record a bet vs. reconcile).

Flows A and C, as well as the headless variant of Flow B, have no dialog gate — status lines don't apply there (there's no one to observe/present to). The skill doesn't open a checklist (task plan): it's invoked by the agent without a dialog, and there's no one to show a forward-looking plan to.

Idempotency and write cadence

All write operations (recording a bet in A; marking matured and logging the outcome in B; rewriting the accuracy map in C) run in the carrier-reliability cadence under the global writer lock:

Start:  bin/run-guard.sh check  → (exit 3 → bin/run-guard.sh recover) → bin/run-guard.sh lock calibration
Finish: bin/run-guard.sh commit → bin/run-guard.sh log --mode calibration --outcome ok → bin/run-guard.sh unlock
  • Writer contention surfaces at the lock calibration step: its exit 1 (writer lock held by another writer) — the run is skipped, don't wait.
  • check covers zone cleanliness and orphaned locks: exit 0 — clean; exit 3 — dirty zone (needs recover); exit 4 — orphaned lock (the next lock clears it).
  • A pure read (an accuracy-map query in Flow C) runs without a lock: there's nothing to corrupt.
  • The script lives at .claude/armarium/bin/run-guard.sh; run it from the repo root.

Idempotency is built into each flow:

  • A: repeating the same bet (same wording + due-date) creates no duplicate.
  • B: re-reconciling an already resolved bet doesn't double the outcome; a repeat headless pass doesn't duplicate the matured mark.
  • C: the accuracy map is rewritten from the current set of resolved bets (not appended to) — the recompute is idempotent.

Flow A — record a bet

Light gate: the owner's statement is already explicit intent, so there's no separate "proposed → approved" ceremony (the write goes to the service zone, not to the base).

  1. Read bets-ledger.md. date for the creation date.
  2. Dedup: if the timeline already holds a bet with the same wording and the same due-date — don't duplicate; name the existing entry to the owner and exit.
  3. Under the lock, append a timeline entry in the entry format (below) with status open.
  4. Assign id: the next in sequence (B-001, B-002, …; take the max existing + 1).
  5. Edge — a bet with no due-date: record it open with no deadline; it won't auto-appear in matured (surfaced only by an explicit "let's settle up on X" command). Say this to the owner plainly, don't lose it silently.

Bet entry format (one timeline entry)

### B-001 · open
- Created: 2026-06-26
- Statement: <the owner's claim — as he phrased it>
- Topic/domain: <topic for the accuracy map>
- Due-date: 2026-12-31      # or "—" if no deadline is set
- Status: open

On a status change the entry is extended (outcome fields added); the id and creation block are not rewritten.


Flow B — reconcile

  1. date → current date.
  2. Read bets-ledger.md, select the open bets whose Due-date ≤ current date.
  3. Under the lock, move them to matured (append matured on <date>). A repeat pass doesn't duplicate an already matured bet.
  4. Fork by run mode:

    Interactive session (owner in dialog):

    • Status: [calibration] ⏸️ Presenting matured bets: <N> awaiting a verdict.
    • Present the matured list to the owner. For each, the owner gives a verdict: hit / miss / partial + reconcile date + a short note.
    • Under the lock, move the bet to resolved, record the outcome (fields below).
    • Do not rule the verdict for the owner and do not go to the web for facts.

    Headless run (cron / piggyback): only mark matured bets matured and defer the reconcile to an interactive session. Don't request a verdict (no one to present to), don't write an outcome. The matured bets will be presented in the next interactive session.

  5. Edge — matured with no verdict: stays matured (not lost, not self-assigned); presented again at the next reconcile.
  6. Edge — ambiguous outcome: a partial verdict or a deadline extension by the owner is allowed; don't second-guess.
  7. Edge — nothing matured / empty ledger: report an honest empty result; don't invent bets.

Outcome fields (extend the entry on resolved)

### B-001 · resolved
- Created: 2026-06-26
- Statement: …
- Topic/domain: …
- Due-date: 2026-12-31
- Status: resolved
- Outcome: hit          # hit | miss | partial
- Reconciled: 2026-12-31   # date of the owner's verdict
- Note: <the owner's short note>

A dated outcome is immutable: once written it's not rewritten or erased (the accuracy map needs the grounding history).


Flow C — calibration (accuracy map)

  1. Read bets-ledger.md, gather all resolved bets.
  2. Group by Topic/domain. For each topic:
    • Hit = count of hit; Partial = count of partial; Miss = count of miss.
    • Accuracy = (Hit + 0.5 × Partial) / (count of resolved for the topic).
  3. Rewrite the compiled top of bets-ledger.md (the accuracy-map table) under the lock. The rewrite from the current set of resolved bets is idempotent.
  4. A map query ("how's my calibration on topic X") is a read without a write, no lock: just return the current top, or the row for the requested topic.

The accuracy map is an aggregate of the owner's judgments, a calibration tool; it is not a measure of the agent's selection quality (a different artifact, not here).


Boundaries (what the skill does not do)

  • Writes only to .claude/armarium/memory/bets-ledger.md. Doesn't write to knowledge-base/; stays read-only over the base. Doesn't write to or read state / taste / sources.
  • The outcome verdict always belongs to the owner. The agent keeps the ledger and surfaces what has matured, but doesn't decide "hit/miss" — neither from the model's parametric knowledge nor from the web.
  • No WebSearch / WebFetch. There's no autonomous web fact-check of outcomes. Bash is only date and run-guard.sh.
  • Doesn't run the base taste diff (base-diff.sh) and doesn't touch recon working state. The calibration mode is isolated from recon.
  • Makes no forecasts of its own. The owner formulates the bet; the agent normalizes and records it.
  • Doesn't erase dated resolved outcomes. Compaction folds them into the accuracy map but leaves them in the archival part of the timeline.

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.