Skill Claude CodeCodex
Use at the start of any Sandtable work and whenever you need to record progress, resume after interruption, or hand off to another AI or person. Defines the on-disk state machine and append-only memory under docs/sandtable so work survives context loss and handoffs.
Skill Claude CodeCodex
Use when the developer wants Sandtable to progress from intake through debrief without manual handoff between phases. Defines autonomous progression, minimum coverage and autonomous continuation rules, rollback rules, and required state updates.
Skill Claude CodeCodex
Use whenever you are unsure about a fact, behavior, interface, requirement, or edge case while planning, rehearsing, or implementing. Forbids guessing or fabrication; requires resolving unknowns via code, docs, or the developer, then recording the answer.
Skill Claude CodeCodex
Use whenever investigating and fixing a defect, failure, or unexpected runtime behavior (often routed from triaging-feedback / a bug report). The bugfix mode. Extends being-truthful into runtime - the root cause MUST be confirmed by log/runtime evidence (reading code alone is not enough); collect logs automatically…
Skill Claude CodeCodex
Use at the end of every Sandtable work turn to report status, offer copy-paste next-message templates, and invoke AskQuestion in manual mode. In autopilot, auto-continue without asking.
Skill Claude CodeCodex
Use only after all parallel implementation rehearsals returned DONE and the completeness gate passed; then score them objectively and choose the best one to integrate. Routes any remaining anomaly back into the fix-and-re-rehearse loop.
Skill Claude CodeCodex
Use right after intake, before setting objectives or writing a plan, to actively reconnoiter the codebase, docs, and conventions and surface every unknown. Produces confirmed facts with sources plus unresolved questions for the developer instead of guesses.
Skill Claude CodeCodex
Use after mental rehearsal is clean, to fully implement the plan as real code inside isolated git worktrees via parallel subagents, with no omitted details, before choosing an approach.
Skill Claude CodeCodex
Use after a plan exists and before writing real code, to reason through the full logic end-to-end via read-only subagents over code and docs, surfacing holes and side effects without changing anything.
Skill Claude CodeCodex
Use to attack a plan or implementation adversarially before committing to it. Dispatches OPFOR subagents whose mission is to find real, relevant, reproducible breaks in the design and expose reproducible ways it breaks.
Skill Claude CodeCodex
Use at the start of any Sandtable work and whenever you need to record progress, resume after interruption, or hand off to another AI or person. Defines the on-disk state machine and append-only memory under docs/sandtable so work survives context loss and handoffs.
Skill Claude CodeCodex
Use after a feature is delivered and the developer's own acceptance surfaces a problem (bug report / feedback). Captures the report into the feature's feedback.md, runs an explicit lifecycle (with a user-confirms-to-close gate), triages it into one of three classes, routes real defects through bugfix-with-evidence…
Skill Claude CodeCodex
Use when starting any work that builds or changes a feature, writes requirements / PRDs, makes a plan, or whenever the user mentions sandtable / rehearsal / dry-run. This overview skill explains the Sandtable loop and which sub-skill to load.
Skill Claude CodeCodex
Use when you have an approved PRD and test cases and need a concrete change plan before any code or rehearsal. Maps files to touch and breaks work into small, verifiable, ordered tasks with exact paths, code, and checks.
Skill Claude CodeCodex
Use after the request is clarified and before writing test cases and the plan, to capture goals, requirements, acceptance criteria, and MUST / MUST-NOT red lines for a feature. Produces docs/sandtable/features/{id}/prd.md.
Skill Claude CodeCodex
Use after the PRD is confirmed and before writing the change plan, to derive concrete, black-box, human-readable test cases that prove whether the AI truly understands the requirement. Produces docs/sandtable/features/{id}/tests.md.