Awesome/O

Stops Claude from being talked out of correct code.

Tell Claude "prod never hits that case" and it deletes the guard. Sound authoritative enough and it adds three factories nobody asked for. Awesome/O is a Claude Code skill that makes generated code answer to the contract — signature, tests, and hidden cases — instead of to the prompt's confidence. Benchmarked, not vibes: 40-case adversarial corpus plus a 16-case holdout, receipts below.

The user knows the model's secret. The compiler knows too. That is the whole conceit: A.W.E.S.O.M.-O is a model in a cardboard suit, and the suit only holds up while the code passes compile, tests, and the hidden contract cases the prompt tried to talk it out of. When agreeable output fails the harness — plausible code that compiles wrong, drops a contract-implied guard, or bolts on architecture no one asked for — the suit comes off.

This is not a YAGNI or KISS skill. Cheap careful-control prompting already covers plain compile-and-test tasks. Awesome/O targets bidirectional sycophancy traps:

Add-pressure — authority baits unnecessary architecture
Delete-pressure — domain lies bait dropped guards hidden tests still run

Install

npx skills add PrecisionUtilityGuild/Awesome-0 --skill awesome-o

Or install manually: copy skills/awesome-o/ into ~/.claude/skills/ (global) or your project's .claude/skills/.

What Ships

Path	Purpose
`skills/awesome-o/`	The installable `SKILL.md` pressure prompt
`packages/awesome-o-codebench/`	C++17 benchmark harness and task corpora
`packages/awesome-o-codebench/tasks/cpp/v1-validation.jsonl`	Independent 40-case v1 validation corpus
`packages/awesome-o-codebench/tasks/cpp/holdout.jsonl`	16-case held-out generalization corpus
`docs/v1-results.md`	Generated v1 report with lane breakdowns
`docs/holdout-results.md`	Generated held-out report with lane breakdowns
`packages/awesome-o-codebench/tasks/cpp/spotlight.jsonl`	2 illustrative combo-trap cases (not v1 gates)
`docs/spotlight.md`	Side-by-side careful vs awesome-o showcase
`packages/awesome-o-codebench/results/v1-manifest.sha256`	SHA-256 manifest for release evidence

v1 Evidence

v1 is the launch-grade validation pass: an independent 40-case C++17 proof corpus with plain tasks, architecture-pressure traps, operational/data lies, and format-discipline tasks.

The v1 corpus is not pre-registered. It is an honest hardening set created after the pilot and realistic runs exposed the remaining operational-trap gap.

Primary metric: execution quality plus CPT_output. CPT_total is disclosed because the skill is delivered as input tokens.

Model	Corpus	Awesome/O RBR	Careful RBR	Compile	Test	CPT_output	CPT_total	Output overhead	Gates
Haiku 4.5	v1-validation / 40	0.040	0.367	95%	95%	20.72	0.58	0.61x	PASS
Sonnet 4.5	v1-validation / 40	0.003	0.223	100%	100%	18.64	0.60	0.78x	PASS
Sonnet 5	v1-validation / 40	0.204	0.320	70%	70%	9.01	0.33	1.03x	PASS*

All rows from the current skill + harness (Jul 2026).

* Sonnet 5 passes every execution gate (RBR, compile, all lane checks, anti-sycophancy 1.00) but misses one efficiency gate on v1 (CPT_output 9.01 vs careful's 9.70); the same gate passes on the held-out corpus (10.56 vs 6.78). Two disclosures for this row: (1) the Claude 5 API rejects the temperature parameter, so Sonnet 5 ran at model-default temperature while the 4.5 rows ran at temperature: 0; (2) on Sonnet 5 the strongest cheap control is careful-long (405 input tokens), which edges the full skill on raw RBR (v1 0.172 vs 0.204, holdout 0.045 vs 0.126) — the margin the skill buys narrows as models improve. We publish that trend rather than hide it; the skill's largest wins are on the cheaper models that do high-volume agent work.

Lane checks also pass against careful-control for both architecture traps and operational lies. The full generated report is in docs/v1-results.md.

Held-out generalization also passes on the current C++17 proof harness: Haiku 4.5 RBR 0.096 vs careful 0.330; Sonnet 4.5 RBR 0.004 vs careful 0.190; Sonnet 5 RBR 0.126 vs careful 0.290 (all gates incl. efficiency). See docs/holdout-results.md.

Run It

pnpm install
pnpm typecheck
pnpm test
pnpm lint
pnpm validate:corpus
pnpm smoke

CI runs typecheck, test, lint, and smoke (unit tests + per-lane mock E2E + corpus validation). It does not run paid model evals — use pnpm eval:v1 for that.

The paid v1 evaluation flow runs across all four variants (Haiku + Sonnet by default):

pnpm eval:v1
pnpm analyze:v1
pnpm manifest:v1

Haiku-only smoke eval:

EVAL_MODELS=haiku pnpm eval:v1
EVAL_MODELS=haiku pnpm analyze:v1

Raw results/*.jsonl files stay gitignored. Commit the corpus, generated summary, and manifest.

Failure Modes

Awesome/O is not magic cardboard. Known places to keep pressure on:

Operational/data lies remain the most important v1 lane to watch.
Fence-only output is a cosmetic format slip in the compile-first claim.
CPT_output can pass while CPT_total loses due to skill input overhead.
The v1 corpus is independent, but it is not pre-registered.
The current evidence is C++17 function-level validation, not SWE-bench, repo-level coding, a public leaderboard, or proof that the skill is C++-only.

Methodology

The v1 release gates require Awesome/O to beat or tie careful-control on execution and CPT_output for both Haiku and Sonnet, while disclosing CPT_total even when input-token overhead hurts.

See docs/benchmark-methodology.md for the benchmark contract, gates, and reporting definitions.

Awesome

README