WinCreator

Your agent says "done, tested." Can you prove it — six months from now, to someone who wasn't in the conversation? Most agent sessions can't: the same context that wrote the code also graded it, the constraints from message one are long gone by message forty, and a failing check gets re-attacked at the same level, forever.

WinCreator turns any non-trivial engineering task into a hierarchy of verification-gated loops with an auditable Proof Ledger, Builder/Skeptic role separation, and a re-emitted loop state panel that survives long sessions.

ledger-check License: MIT Works with Sponsor

30-second demo

The mechanical gate is scripts/ledger_check.py — stdlib-only, no dependencies. It parses a PROOF_LEDGER.md table and refuses to exit 0 if any claim is unproven. Below is a real run, not a mockup.

Save this exact table as PROOF_LEDGER_dirty.md — one unproven claim, one rubber-stamped one:

| ID | Level | Claim | Gate (what proves it) | Status | Evidence |
|----|-------|-------|------------------------|--------|----------|
| P1 | Meso  | /export endpoint returns 200 for all active cases | Run full dataset, inspect status codes | EVIDENCED | run 2026-07-10: pytest -k export -> 14 passed |
| P2 | Micro | parser rejects malformed input | unit test on truncated file | CLAIMED | |
| P3 | Micro | CSV export matches DB row count | diff row counts after export | EVIDENCED | ok |

$ python3 skill/Skill_WinCreator/scripts/ledger_check.py PROOF_LEDGER_dirty.md
LEDGER NOT CLEAN — 2 violation(s):
  ✗ line 4 [P2]: CLAIMED with no evidence — loop may not report done. Claim: parser rejects malformed input
  ✗ line 5 [P3]: status EVIDENCED but Evidence cell is empty or too vague ('ok')
exit=1

Fix the two claims — real evidence, not "ok":

| ID | Level | Claim | Gate (what proves it) | Status | Evidence |
|----|-------|-------|------------------------|--------|----------|
| P1 | Meso  | /export endpoint returns 200 for all active cases | Run full dataset, inspect status codes | EVIDENCED | run 2026-07-10: pytest -k export -> 14 passed |
| P2 | Micro | parser rejects malformed input | unit test on truncated file | EVIDENCED | pytest -k truncated_input -> 1 passed, AssertionError raised as expected |
| P3 | Micro | CSV export matches DB row count | diff row counts after export | EVIDENCED | export.csv rows=482, SELECT COUNT(*) FROM cases=482, match |

$ python3 skill/Skill_WinCreator/scripts/ledger_check.py PROOF_LEDGER_clean.md
LEDGER CLEAN — 3 row(s) verified in PROOF_LEDGER_clean.md
exit=0

That is the entire mechanism. No LLM call, no network, no config. A single stdlib-only Python script (ledger_check.py, 91 lines) stands between "the agent said it's done" and a merge.

The three failure modes

Every agent-assisted engineering session degrades the same three ways, and knowing them is the key to using this skill well.

  1. Optimism leak — the same context that wrote the code grades its own work, and "it should work" quietly becomes "it works". Countermeasure: the Proof Ledger + Builder/Skeptic separation. Whoever builds never grades their own gate; the Skeptic receives only the claim, the gate, and the raw evidence — never the Builder's reasoning, which contaminates judgment.
  2. Context rot — constraints stated early get lost as the session grows; the agent drifts from the original need without noticing. Countermeasure: the Loop Panel, a four-line state block re-emitted at every gate event (gate announced, gate passed, gate failed, level change), so the current truth survives even when session memory degrades.
  3. Stuck loops — a failing check gets re-attacked at the same level again and again, when the real defect lives one level up. Countermeasure: the Two-Failure Rule — the same gate failing twice forbids a third identical attempt and forces a level audit first ("would this failure disappear if something one level up were different?").

Real worked example

skill/Skill_WinCreator/references/worked-example.md is a full transcript, not a fiction: a semver comparator with five green tests and a docstring promising ValueError on malformed input — a promise no test actually checked. The Skeptic pass caught the gap that "tests pass → done" would have shipped, and a sixth test closed it before the loop was allowed to report up.

Installation

Claude Code

git clone https://github.com/winterbim/wincreator.git
cp -r wincreator/skill/Skill_WinCreator ~/.claude/skills/Skill_WinCreator

npx skills

npx skills add winterbim/wincreator

Manual download

Download Skill_WinCreator.skill from the latest release and unzip it into your skills directory.

This repo eats its own dogfood

PROOF_LEDGER.md at the repository root is not a sample — it is the real ledger for the act of publishing this repository: every claim made while shipping WinCreator (structure preserved unmodified, demo output real, no placeholder credentials committed, attribution consistent) is a row in that file, and .github/workflows/ledger.yml runs ledger_check.py against it on every push. The green badge at the top of this page is that check, live, not a static image — its most recent run verified all rows in PROOF_LEDGER.md on GitHub's own runner.

Bootstrap the evolution loop in your own project

The skill itself (doctrine + gate + SKEPTIC_CATCHES.md) travels in the .skill package. The evolution circuit that lets the skill improve itself — the .claude/agents/, the CLAUDE.md wiring, the proposals queue — is repository scaffolding, so it travels via git clone, not the single-file .skill. To run that circuit inside one of your own projects:

# sources: your clone of this repo, and wherever you installed the skill
WINCREATOR=/path/to/wincreator                 # git clone of this repo
SKILL=~/.claude/skills/Skill_WinCreator        # where you installed the skill

cd /path/to/your/project

# 1. the four evolution agents
mkdir -p .claude/agents
cp "$WINCREATOR"/.claude/agents/wincreator-*.md .claude/agents/

# 2. APPEND the loop section to your CLAUDE.md — never overwrite; your project
#    may already have one (bimwin does), and cat >> preserves it
cat "$WINCREATOR"/templates/CLAUDE.evolution.md >> CLAUDE.md

# 3. create the project-level queue and catches log (only if absent)
[ -f EVOLUTION_QUEUE.md ] || printf '# Evolution Queue\n\n_(empty — awaiting the first retro-analyst pass)_\n' > EVOLUTION_QUEUE.md
[ -f SKEPTIC_CATCHES.md ] || printf '# Skeptic Catches — project log\n\n| Date | Class | Why missed | Question that would have caught it |\n|------|-------|------------|-------------------------------------|\n' > SKEPTIC_CATCHES.md

# 4. the ONLY install smoke test:
python3 "$SKILL"/scripts/ledger_check.py --self-test     # -> self-test: 20/20 passed

Every command above was executed in a virgin directory before being published here (PROOF_LEDGER-evolution.md, cycle 4). Two things a fresh setup must know:

  • --catches is STALE at day zero, and that is correct. A new project has no catches yet, so ledger_check.py --catches ./SKEPTIC_CATCHES.md returns CATCHES STALE, exit 1 — the expected state, not a broken install. That gate only becomes meaningful after real loops feed the log. Use --self-test (not --catches) as the install check.
  • Two SKEPTIC_CATCHES.md can exist; the authority is decided. The one in ~/.claude/skills/Skill_WinCreator/ is read-only inherited heritage; your project's ./SKEPTIC_CATCHES.md is the live log the retro-analyst writes. A Meso+ loop re-reads both at its start. The full rule is in templates/CLAUDE.evolution.md, which step 2 appends to your CLAUDE.md.

Support this project

WinCreator is free and MIT-licensed. If it saved you from shipping an unverified "done," consider supporting its development: buymeacoffee.com/wintfernanh · bimcheck-consulting.com

License

MIT — see LICENSE. Copyright Winter Fernandes.

If you use, fork, or build upon this skill, credit Winter Fernandes as the original creator with a link back to this repository. The MIT license already requires the copyright notice to be preserved; this line makes that expectation explicit and visible.

Created by Winter Fernandes — BIMCheck Consulting bimcheck-consulting.com

Attribution required in derivatives — see NOTICE.