track-record - A predictions ledger with forced settlement - log every judgment as a bet, settle it against reality, distill recurring biases into rules.

Mirrored in three places, same version everywhere: Hugging Face (you are here) · GitLab · GitHub

Get the files - GitLab is the most reliable plain-git route:

git clone https://gitlab.com/LucioLiu/track-record.git

# or from this page
hf download LucioLiu/track-record --repo-type dataset --local-dir ./track-record

Then drop the folder into your agent's skills directory - ~/.claude/skills/track-record/ for Claude Code, ~/.agents/skills/track-record/ for Codex, or your client's equivalent.

Note: links to github.com inside the text below point to the GitHub mirror; if one does not resolve, use the GitLab mirror above - it carries the same content.

Licence: PolyForm Noncommercial 1.0.0 - the LICENSE file in this repo is authoritative.

---

English | 简体中文

track-record / log your bets, settle against reality

A junior dev that finally learns from its mistakes — because it keeps a record.

track-record is an installable skill that gives an AI coding agent (Claude Code or any skills-capable harness) a calibration loop: every checkable judgment is logged as a bet — numeric probability, observable settlement trigger; reality settles it; recurring biases get distilled into rules, then mechanisms. The learning signal is reality versus the agent's own predictions — not user corrections.

The problem

A recurring complaint about coding agents — here in a top HN comment — is that you get "a digital junior dev" that "doesn't really learn from its mistakes, and lies all the time". Users describe being told "I've fixed it" when nothing was fixed (example), and a feature request for a self-improving learning loop was closed as not planned — if the harness won't ship this, it has to live in a skill.

Existing tools attack this by harvesting user corrections. That works — but it only learns what the user bothered to say. The larger signal is free and nobody collects it: the agent constantly makes checkable predictions ("this cache will cut P95 below 200ms", "this refactor won't break the exporter"), reality eventually answers every one of them, and by then nobody remembers what was predicted. track-record makes the agent keep the receipt.

The loop

flowchart TD
    J([Checkable judgment made]) --> T{"Triage<br/>Goldilocks zone?"}
    T -- "too easy or unknowable" --> X([not logged])
    T -- "genuinely uncertain,<br/>reality will answer" --> L["Log<br/>7-field ledger entry: the bet, its basis,<br/>a numeric probability, an observable<br/>settlement trigger"]
    L --> R([reality unfolds])
    R --> S{"Settle<br/>trigger fired, or<br/>window expired?"}
    S -- "window expired" --> F["forced choice, no fourth option:<br/>settle on current evidence /<br/>mark unsettleable with reason /<br/>one explicit extension with a new trigger"]
    F --> B
    S -- "trigger fired" --> B["Backfill<br/>record actual outcome vs. prediction<br/>incremental, self-caught and<br/>overturned entries all welcome"]
    B --> D["Distill<br/>one-off note → recurring bias →<br/>strongly-worded rule → enforced mechanism"]
    D -- "rules feed back into<br/>the next judgment" --> J

What a settled ledger actually looks like:

#	date	bet	p	settlement trigger	outcome
12	03-02	cache layer will bring P95 under 200ms	70%	perf report after 2 weeks live	WRONG — P95 at 210ms; gain eaten by cold starts. Bias: overweighted local benchmark
13	03-04	migration script is idempotent	90%	re-run against staging snapshot	CORRECT
14	03-09	flaky test is a teardown race	55%	fix lands, then 50 CI runs	【partial】 flake rate halved; second cause suspected

Two design points worth calling out:

• Forced settlement. When a bet's observation window expires, "wait a bit longer" is not on the menu — the agent must settle on current evidence, mark the entry unsettleable with a reason, or take one explicit extension anchored to a new observable event. As far as we could find, no existing agent-skill ships this; it exists because the soft version demonstrably fails (see below).
• Independent settlement. Entries are settled by a fresh context (new session or separate agent) that sees only the ledger row and the real-world evidence — no self-grading, no explaining your way out.

How it's different

	learns from	loop closes when	typical output
claude-improve	user corrections & feedback signals	the user tells you something (again)	updated instructions/rules
claude-reflect	user corrections + session signals, confidence-tiered	a session ends and is mined	updated rules / best practices
track-record	reality vs. the agent's own logged predictions	the real-world outcome arrives, however much later	a settled ledger + calibration-driven rules & mechanisms

The three are complementary: the first two capture what the user said; track-record captures what actually happened — including the mistakes no one bothered to point out.

What's in the box

track-record/
  SKILL.md                    # the discipline itself (operative text)
  references/                 # (written in Chinese — the English SKILL.md is self-contained without them)
    ledger-format.md          # 7-field ledger spec, empty template, worked examples
    settlement.md             # hard triggers, timeout triage, independent settlement, optional hook config
    distill.md                # three-tier promotion ladder, bias fingerprints, feed-back chain
    rationalizations.md       # excuse/rebuttal table, red-flag phrases, violation symptoms
  README.md                   # this file
  plugin.json                 # skill-package manifest (package metadata)
  LICENSE                     # PolyForm Noncommercial 1.0.0
  CREATION-LOG.md             # design history — itself written as a mini ledger

No executable code, no daemons, no state outside the ledger file you keep. An optional host hook (session-start / stop reminders) is documented with an example config in references/settlement.md — it is an optional attachment, not a bundled capability.

Why this design

An internal prototype of this ledger ran for about 3 weeks and accumulated 30+ logged judgments. Almost none were ever settled and the archive section stayed empty; the few settlements that did happen were nearly all driven by external events — a user catching an error on the spot, an acceptance review, an independent audit — essentially none by the logger's own initiative. "Log-only, never settle" is the empirically dominant failure mode, which is why forced settlement, session-start scans, and the excuse-rebuttal table are the core of this skill rather than nice-to-haves.

One more scar tissue decision: every numeric threshold in this skill (promotion counts, file-bloat lines, extension limits) is defined in exactly one table in SKILL.md and referenced everywhere else — sibling tools in this space have shipped contradictory thresholds across their own docs.

Install (2 minutes)

Three ways, pick one:

1. GitHub CLI (Claude Code):

   gh skill install LucioLiu/track-record track-record --agent claude-code --scope user
   

2. Manual clone:

   git clone https://github.com/LucioLiu/track-record ~/.claude/skills/track-record
   

   (Codex-style harnesses: clone into ~/.agents/skills/track-record instead.)
3. Single-file copy: copy SKILL.md into your skills directory (Chinese-speaking users should copy SKILL.zh-CN.md instead and rename it to SKILL.md locally). The references/ files add depth but the discipline stands without them.

Then just log your first bet — the skill creates the ledger from the template in references/ledger-format.md on first use.

Uninstall = delete the directory. Your ledger file is yours and stays wherever you put it.

Version & self-check

• Current version: 1.0.0 (see plugin.json).
• Repository: https://github.com/LucioLiu/track-record
• Self-check: diff your installed SKILL.md and plugin.json against the repo's main branch. The skill ships no auto-update mechanism — updating is a deliberate act, like everything else in it.

Verification boundary

The unified structural entrypoint, the ledger schema negative test, and the independent behavior bundle are not published in this repository yet — don't expect a tests or schemas directory in what you installed. A structural PASS would not settle the pressure tests anyway: SK-TR-P-001 and SK-TR-N-001 remain not-run until a fresh subject and independent judge produce the required receipts.

License

PolyForm Noncommercial 1.0.0 — free for personal and non-commercial use.

---

Both language editions are maintained in sync. If they ever diverge, the English edition is authoritative — please open an issue if you spot one.