An evidence-first Agent Skill for surfacing the few unknowns that could change a decision before time and effort are committed.
find-unknown runs a focused 5–10 minute preflight. It inspects available
evidence, identifies at most three direction-changing unknowns, and routes each
one to investigation, a decision question, a prototype, or an experiment.
Plans often fail because a material assumption stayed invisible, not because the team needed a longer checklist. This Skill focuses the agent on unknowns that could invalidate the current direction:
- inspect evidence before asking the user;
- surface zero to three material unknowns, never filler;
- turn confirmed defects and resolved items into findings or plan inputs rather than retaining them as unknowns;
- ask one decision-changing question at a time;
- re-rank the remaining unknowns after every answer;
- reconcile resolved unknowns with the active plan or return an authorized plan delta;
- integrate supporting findings into the requested deliverable instead of appending a duplicate standalone preflight;
- prototype preferences that are easier to recognize than describe;
- turn unknowable future effects into small, falsifiable experiments.
The result is a bounded preflight, not an exhaustive interview.
| Evidence source | Route | Typical example |
|---|---|---|
| Files, code, documents, data, tools, or environment | Investigate | Does the current architecture support the proposed integration? |
| User goal, preference, authority, or trade-off | Ask | Is regional data residency a hard requirement? |
| A preference the user can recognize only after seeing it | Prototype | Which of three visibly different layouts fits? |
| A future effect neither side can know directly | Experiment | Will this onboarding change improve activation? |
If one unknown is a prerequisite gate, downstream branches are deferred until it passes. When a user answer changes the direction, the Skill discards resolved or invalidated unknowns and re-ranks the rest.
After resolving a material unknown, the Skill records whether the plan should add, modify, remove, cover, defer, or block work. It updates an authorized active task plan when available; otherwise it returns a precise delta without creating another project-management authority.
The full runtime contract is in skills/find-unknown/SKILL.md.
npx skills@latest add Odinary-AI/find-unknown --skill find-unknownGive Codex this request:
Install the find-unknown skill from
https://github.com/Odinary-AI/find-unknown/tree/main/skills/find-unknown.
The Skill becomes available on the next turn after installation.
git clone https://github.com/Odinary-AI/find-unknown.git
mkdir -p ~/.codex/skills/find-unknown/agents
cp find-unknown/skills/find-unknown/SKILL.md ~/.codex/skills/find-unknown/SKILL.md
cp find-unknown/skills/find-unknown/agents/openai.yaml ~/.codex/skills/find-unknown/agents/openai.yamlThe repository follows the Agent Skills folder convention and has no runtime dependencies. It is verified in Codex; behavior on other compatible agents has not yet been evaluated.
Invoke it explicitly when a decision deserves a preflight:
Use $find-unknown to scan this rollout plan for assumptions that could change
the direction.
在我们选定供应商之前，用 $find-unknown 做一次 5 分钟的未知预检。
Run a blindspot scan on this architecture decision. Investigate anything the
repository can answer before asking me.
The output identifies the current direction, evidence, direction impact, and smallest next validation. If a user-only decision blocks progress, the Skill asks one question with a provisional recommendation, its basis, and the consequences of materially different answers.
Do not use find-unknown for:
- routine execution with no material decision;
- a straightforward factual lookup;
- speed-first tasks where the user explicitly wants no preflight;
- exhaustive product discovery or a full architecture interview.
If the user wants every branch of a plan explored through a deep interview,
Matt Pocock's grill-me / grilling workflow is a better fit.
Evidence for the current Skill includes:
- 35 public package and behavior contract tests;
- fresh, project-independent, manually observed RED, GREEN, counter-example, and forward samples on the current canonical text;
- focused scenarios for prerequisite gates, answer-driven re-ranking, and prototype-before-asking and conditional output ownership;
- Chinese decision-context triggers, routine omission-check exclusions, unknown lifecycle, and plan reconciliation across available carriers.
The previous release's 36 independent behavior samples remain historical regression context; they were not rerun against the current text.
These checks support the interaction and routing claims above. They do not yet prove reduced rework on long-running real projects, cross-model stability, or that the three-question and 5–10 minute limits are globally optimal. See the evaluation report for the evidence boundary.
.
├── skills/find-unknown/
│ ├── SKILL.md # Runtime instructions
│ └── agents/openai.yaml # Codex UI metadata
├── docs/evaluation.md # Public validation evidence
├── tests/ # Package and behavior contracts
└── .github/workflows/ # Automated contract checks
The adaptive questioning in find-unknown was informed by Matt Pocock's
grill-me and reusable grilling
workflow: ask one consequential question at a time, provide a recommendation,
and let each answer reshape the remaining decision tree.
find-unknown applies those ideas inside a different boundary: an
evidence-first, time-boxed preflight with at most three direction-changing
unknowns. It investigates before asking and routes recognition-based
preferences to prototypes.
MIT © 2026 Odinary-AI