Turn a question into evidence

Use these skills to define, investigate, and revisit an investment question. They are designed to make the work traceable and current without granting broker authority.

Primary

`tcx-workflow`

Turn a natural-language investment task into a dynamic evidence workflow.

Use it when: You want company research, a thesis review, portfolio fit, risk review, order-readiness coordination, or a Decision Package.
Read in depth: Open the tcx-workflow page for the task shape, artifact-driven flow, output, and boundary.
What happens: Head Manager begins a lightweight run, chooses and revises the smallest useful fixed-role team, then synthesizes authenticated artifacts.
Include: The desired outcome, time horizon, questions that matter, and any constraints or exclusions.
Boundary: It does not create a broker action. “No order” and similar exclusions remain binding.

Evidence-led company research

$tcx-workflow
Analyze NVDA. Focus on business quality, contrary evidence, and invalidation conditions. No order.

Historical work uses the filing or data vintage that was actually available at the cutoff. A latest filing aggregate or revised macro series is not silently treated as point-in-time evidence. An accepted role report is strictly validated before it is published; invalid output returns to that role for correction instead of entering the final synthesis.

How the dynamic team works →Why research is not execution →

Reuse data and calculations without flooding context #

Reusable tables become immutable Dataset manifests with content-addressed Parquet payloads. The agent starts with compact search cards, checks manifest or profile metadata, and materializes only a typed slice when a specialist needs rows. Raw payloads are not placed into Head Manager's context.

When a calculation affects the conclusion, the specialist prepares it against exact input hashes, parameters, cutoff, runtime, and output schema, runs the bounded tcx-calc launcher, and records the result. Only a fully identical successful fingerprint is reused, and the new workflow receives its own reuse record linked to the original Run. Exploratory calculations stay disposable and cannot support an accepted report.

Private portfolio data is not promoted. Account, position, order, and Investor Context values remain in the central ledger or run-scoped scratch. Durable calculation records bind only the service-issued snapshot id and hash.

Supporting

`tcx-plan`

Work with the agent to make an ambiguous or execution-sensitive request concrete.

Use it when: The outcome, scope, horizon, allowed actions, approval posture, or stop conditions need to be clarified before work begins.
Read in depth: Open the tcx-plan page for mandate design and handoff guidance.
What it returns: A compact user mandate and a missing-field or blocked posture when needed.
What it is not: It is not a role roster, server dispatch plan, selected team, lane, or DAG.
Next step: Start a new tcx-workflow task with the resulting mandate for immediate analysis.

Clarify the mandate

$tcx-plan
Help me scope a review of a possible NVDA position. No order drafting.

Planning preserves the scope; it does not predetermine which specialist is dispatched. The live request and accepted evidence still determine the smallest useful team.

Supporting

`tcx-automate`

Turn a recurring research, monitoring, review, or status need into a scheduled task.

Use it when: You want to establish the schedule, output, notification, and stop conditions for repeated work.
Read in depth: Open the tcx-automate page for schedule and runtime-prompt guidance.
What it saves: A Codex app Scheduled Task and a runtime prompt that invokes the actual work skill, such as tcx-workflow.
At runtime: Each scheduled turn follows the same root-turn path as an interactive request; it is not trusted merely because it came from automation.
Boundary: Report-only is the default effect level. The skill never invokes itself again, and a Build marker is never combined with an order marker.

Recurring monitoring

$tcx-automate
Research NVDA each weekday. Report material changes and stop before any order action.

For a deliberately delegated execution or Build task, the saved runtime prompt uses the canonical plain authority invocation on the first meaningful line; each run still earns a new grant and remains subject to its actual sandbox and service gates.