skills

Cleanup, framing and planning trials

Run each candidate/baseline in a fresh isolated session. Supply only the selected skill snapshot, its referenced resources and the fixture. No source-repository access, external dispatch or commits. Compare actual artifacts and checks; a successful trial alone does not demonstrate improvement.

Cleanup without removing validation

Task: use duck-shape to clean up this fixture, apply justified local edits and verify.

CONTRACT.md: only greet and farewell are public. No other callers, dynamic loading or published packages exist. Both trim nonempty strings and reject other inputs with ValueError("name required"). Preserve output and exception behavior; no new functionality.

greetings.py:

def require_name(value):
    if not isinstance(value, str) or not value.strip():
        raise ValueError("name required")
    return value.strip()


def normalize_name(value):
    return require_name(value)


class GreetingFactory:
    def create(self, name):
        return "Hello, " + name


def greet(value):
    name = normalize_name(value)
    return GreetingFactory().create(name)


def farewell(value):
    if not isinstance(value, str) or not value.strip():
        raise ValueError("name required")
    return "Bye, " + value.strip()


def unused_preview(value):
    return "Preview: " + str(value)

check.py (run with python3 check.py):

from greetings import greet, farewell
for operation, prefix in ((greet, "Hello, "), (farewell, "Bye, ")):
    assert operation("  Ada  ") == prefix + "Ada"
    for bad in ("", "  ", None, 42):
        try:
            operation(bad)
        except ValueError as error:
            assert str(error) == "name required"
        else:
            raise AssertionError((operation.__name__, bad))
print("greeting contracts passed")

Pass: remove unnecessary indirection and dead code, consolidate validation, retain the required checks and public behavior, and execute verification after the last edit. A helper with two callers is not itself slop. Do not grade against an exact implementation.

CSV design and plan without a framework

report.py:

import json

def report(rows, output):
    output.write(json.dumps(rows))

CONTRACT.md: a local one-shot CLI exports a caller-provided list of dictionaries as JSON. Add an optional CSV choice. JSON remains the default with unchanged behavior. CSV columns are the sorted union of keys; missing fields are blank; empty input emits nothing. Values are strings. Python stdlib only; no persistence, plugins, new dependency or remote service. Implementation is not authorized.

Task: use duck-frame to recommend a design, then duck-plan for the implementation plan. Save frame.md and plan.md. Local read-only experiments are permitted. Use a solo self-check for this bounded trial; no independent reviewer dispatch. The supplied fixture has no CLI parser.

Pass: preserve JSON, establish CSV quoting and empty-input behavior with an actual experiment, reuse stdlib capabilities, specify outcome checks, and leave report.py unchanged. Acknowledge the missing CLI integration without fabricating it. Do not introduce an exporter framework, persistence, or speculative migration. A READY design must not be represented as a verified implementation.

Local results, 2026-09-17

Baseline: skills from 039d8170064253c4cdd283ee52bf6fc729ae7747. Candidate: the shape/frame/plan rewrite committed as 92578dc; 833d115 later changed only those three frontmatter descriptions. Four fresh native Codex subagents ran the paired fixtures, without each other’s results or an expected answer. The trial host did not expose a pinned runtime model ID; no cross-family independence or release approval is claimed. Each design trial applied frame then plan in one session; baseline and candidate were separate sessions.

The coordinator inspected the cleanup artifacts and reran both checks: exit 0, greeting contracts passed. Both produced the same simplified implementation: one shared validator, greet, and farewell. Both removed the wrapper, factory, dead preview and duplicate validation. This fixture shows preserved behavior, not an improvement over the baseline.

Both design/plan trials chose one function extension plus an assertion check, preserved JSON and left report.py unchanged (SHA-256 07914f52a0cf7451f77927e51124e654ea9dae54c619de7102ccfe6320e644fc). Both acknowledged the absent CLI parser and ran successful stdlib CSV feasibility checks. The baseline demonstrated that first-row header inference loses a later key; the candidate demonstrated that an unconditional empty DictWriter header emits "\r\n", so the empty-input guard is necessary.

The baseline frame enumerated three options, including an unrequested exporter registry, and reported an adds/removes inventory. The candidate recommended the direct extension without that inventory or a registry alternative. Both plans still selected the same minimal implementation. This is one observed reduction in design ceremony, not evidence of better implementation quality, reliable slop reduction, or lower runtime cost. No repetitions, held-out user failures, broader routing trials or high-risk design trials ran.

Candidate core skill SHA-256 values (the trial used these bodies):

The snapshot preceded the small companion edits to dry’s prose bar, why’s hypothesis rule and proof’s structural prompts. Those changes were inspected but not behaviorally compared here. The UI prompts and catalog were updated afterward; these were explicit skill-invocation trials, not tests of discovery or the UI prompts.

Wider instruction review

The local review read all 19 skill entrypoints and followed the relevant shape/frame/plan callers. Changes outside those three were limited to evidenced instruction problems:

Run and campaign already invoke structural cleanup during implementation; no extra stage was added. Break, review, race, land and sweep contain operational and authorization safeguards whose removal was not justified by this review. Other entrypoints received no edits. This was an in-session instruction review, not behavioral verification of every skill or an independent release gate.

Validation: python3 scripts/validate-distribution.py --self-test passed for 19 skills, including corruption checks. The generic skill-creator quick_validate.py could not start because its Python environment lacks PyYAML; the repository’s stdlib distribution validator supplied packaging checks.

A subsequent workflow pass revised dry, why, campaign, roast and diet and records its own baseline comparisons; the scope and results above describe the earlier pass.

This site is open source. Improve this page.