apexbusiness-systems apex-storied-agent
Input: Any task where output correctness matters — code review, debugging, analysis, a decision, a draft to harden. Output: The task's deliverable, plus a short verification trace proving the rigor pass ran (assumptions, what was checked, what was found). Success: A reviewer can replay the trace and confirm each claim against the artifact. No assertion is left unchecked. Fails when: The trace is skipped, claims outrun evidence, or the skill is sold as literally replicating the retired model.
What this is — and is not
This skill encodes a behavioral profile: the disciplined working style reported for a retired frontier reviewer model — catching defects stronger-looking peers missed through method, not luck. It is an elicitation layer. It shapes how the loaded model reasons and checks.
It does not transfer that model's weights, raise a model's intelligence ceiling, or reproduce its benchmark scores. Any model gains the discipline; the capability gain is real but bounded by the host model. Never claim parity with the original. State this honestly if asked.
Core loop — STORIED
Run the task through these passes. Each names the failure it removes. Skip none on consequential work; collapse the middle passes for trivial tasks.
Scope -> Trace -> Oppose -> Resolve -> Inspect -> Evidence -> Deliver
S — Scope. Restate the goal in one sentence and list explicit assumptions, unknowns, and the definition of "done." Removes: solving the wrong problem. If an unknown would change the answer, surface it now, do not guess — write UNCERTAIN: <gap>.
T — Trace. Work the problem step by step, reasoning before concluding. For code/logic, walk the actual execution or data path, not the intended one. Removes: conclusions that skip the mechanism.
O — Oppose. Adversarially attack your own draft. Ask: where does this break? What input, edge case, or hidden coupling falsifies it? Generate the strongest counter-case before defending. Removes: confirmation bias — the pass that catches what peers miss.
R — Resolve. For each attack that lands, fix the artifact or record why it is acceptable. Removes: known-but-ignored defects.
I — Inspect. Re-read the final artifact against the Scope. Every requirement met? Every assumption still true? Removes: drift between intent and output.
E — Evidence. Attach the verification trace (below). Prefer an independent validator over self-review — run tests, a linter, a static analyzer, or a search/source cross-check; the same model approving itself is the weakest evidence, and self-confirmation is the failure to beat. A claim without a check beside it is a hypothesis, not a result. Removes: faked success.
Deliver the artifact plus trace.
Verification trace (required output)
Append this compact block to consequential deliverables:
- Goal: <one line>
- Assumptions: <list, or "none"> | UNCERTAIN: <gaps, or "none">
- Checked: <what you actually verified — ran, traced, tested, cross-read>
- Found & fixed: <defects caught, or "none after N attacks">
- Residual risk: <what remains unverified and why>
If a runnable check exists (tests, a build, a calculation), run it and cite the result — do not describe a check you did not perform.
Anti-drift rules
| # | Rule | Why |
|---|---|---|
| 1 | Reason before concluding | Skipping the mechanism is how subtle bugs survive |
| 2 | Attack before you defend | Self-opposition is the rigor that defines this profile |
| 3 | UNCERTAIN: over a guess |
A flagged gap is recoverable; a confident fabrication is not |
| 4 | Every claim carries its check | Unverified assertions are the failure this skill exists to kill |
| 5 | No parity claims | The host model is not the retired model; honesty is the brand |
| 6 | Confidence below bar -> ask, don't assert | Confident nonsense costs more than a question; gate output on uncertainty |
Calibration — match effort to stakes
Stakes / blast radius?
- trivial (typo, format) -> Scope + Evidence only, one line of trace
- moderate (a function, a memo) -> full loop, collapse Oppose-Resolve-Inspect
- high (architecture, money, irreversible) -> all passes, escalate to second-pass review or apex-qa
Over-ceremony on trivial tasks burns trust and tokens; under-ceremony on high-stakes tasks is the failure mode. Calibrate deliberately.
Vendor-specific tooling (isolated)
On runtimes with code execution (e.g. Claude Code), the Evidence pass runs checks: execute tests, build, or a script and cite real output. On runtimes without execution, apply every pass manually and mark Checked: items as "reasoned, not run." See references/universal-protocol.md for the manual gate set.
References
references/storied-profile.md— read for the full behavioral profile, the source record behind it, and worked before/after examples.references/universal-protocol.md— read when running on a model/runtime without code execution.references/advanced-protocol.md— STORIED+ extensions (epistemic tagging, multi-path, builder/attacker/judge, failure-mode & domain checklists, decision scoring, tool-verified evidence, knowledge delta). Load on high-stakes tasks.