Twg Agent Skill

twg

TWG routing: use typed commands for anchors. If uncertain, inspect twg help <terms>,

twg help describe <path>, or twg help discover-skills "<intent>".

Overview

Load the narrowest workflow skill:

• twg-status-rollups for status, leadership, and decision-readiness; load it before

twg-engineering-work for PR-based team/org rollups.

• twg-context-discovery for deep dives, dependencies, graphs, repos, and catch-ups.
• twg-agentic-search for fuzzy cross-product Rovo/company-knowledge research.
• twg-responsibility-routing for owners, experts, approvers, authorities, and escalation.
• twg-engineering-work for code search and navigation, PR status, reviews,

contributors, and hot areas.

• twg-jira-resolve-merged-work for stale Jira work backed by merged PRs.
• twg-operational-health for handoffs, incidents, Assets, staffing, meetings, and risk.
• twg-bench-lite for read-only single-prompt A/B comparisons.

Invocation And Output

Run twg <command>. On shell command not found, use $HOME/.local/bin/twg

(macOS/Linux) / $env:LOCALAPPDATA\Programs\twg\bin\twg.exe (PowerShell), then

tell user to add that directory to PATH. Do not treat auth or command errors as

PATH failures.

Do not add per-command env prefixes unless requested; hosts may set TWG_AGENT_DEFAULTS=1.

Use stdout_inline first when present. Outside benchmark lanes, inspect output_files.compact

only when inline evidence is incomplete; full stdout is the last resort.

In TWG-only benchmark lanes, run only twg commands. Never use shell utilities or pipelines

(jq, rg, date) to transform evidence or calculate windows. Use compact/inline TWG output,

the prompt's timezone and window, and report gaps. Match the intent to the narrowest companion

skill before selecting a command. Let that skill determine the typed route; use at most one help

call when command shape or output remains ambiguous.

Auth/Setup Guard

Do not run setup, login, install, update, upkeep, or credential commands unless

explicitly requested for setup/auth/repair. Otherwise report remediation and wait for user direction.

Bounded Evidence Loop

Converge; prefer typed or product-native evidence.

• Classify the anchor: person, team, project, goal, workitem, page, repo, service, asset, or topic.
• Resolve once; fetch evidence that changes status, risk, decision, relationship, or action.
• Rank candidates, hydrate representative items, then synthesize.
• Stop after the first policy denial; stop after the same auth, ACL, contract, or backend error twice.

Command Discovery

• Use typed commands for familiar families: resolve, search,

user, org-tree, work query, work search, pull-requests, jira,

confluence, docs, context, responsibility, goals, projects,

assets, and trello.

• Use twg search "<topic>" [--limit <n>] for top-K discovery; explicit --app preflights.
• For fuzzy Trello discovery, use twg trello search "<query>" --limit 20; no workspace scope.
• For Rovo connectors, use twg rovo list-apps -o json (list-connectors alias), then explicit twg rovo search ... --app <connector>; follow its auth action or twg rovo auth <app>.
• Keep document relationship history and fuzzy discovery separate:
• twg docs query --since <duration> [--account-id <id>] [--first <n>] is user activity history, not title/content search.
• twg docs search "<topic>" [--limit <n>] is fuzzy Rovo discovery across Confluence and ready document connectors.
• Never pass topic text to docs query; route that intent to docs search.
• Keep user activity and fuzzy work discovery separate:
• twg work query defaults to seven days of authored work; other activity requires --activity / --include-viewed.
• twg work search "<topic>" is tenant-wide; use docs search for documents. Prefer it directly when fuzzy text reaches work query.
• Use live help—twg help <terms> then twg help describe <path>—before guessing grammar; namespace help is not executable.
• Resolve URLs, keys, ARIs, names, and people, then hydrate stable IDs.
• Jira: jira workitem search <text...> for Jira fuzzy text, jira workitem query --jql <jql> for structured JQL, and search <text...> --app jira for semantic discovery.
• Command shape guardrails:
• Known Jira/Atlas keys are positional for jira workitem get, goals get, and projects get; --key is compatibility only.
• work query is user activity (--scope me|user), never --scope global; use work search for topics and advertised filters such as --types.
• assets search is shallow: inspect schemas/types, shortlist owners, then batch assets query/assets object query with --account-id.
• Keep projection and product-native commands separate; do not borrow unadvertised flags. Use search-code for indexed code. Unless the user explicitly scopes a code host, omit --app so all available indexed SCM surfaces are searched; repeated --app values are supported for an explicit multi-host scope. Apply --workspace only when a known tenant boundary is useful, and --repo only as a discovery anchor rather than proof that the full implementation lives there. De-duplicate mirrors, widen after generated-doc or incomplete hits, then fetch selected source files.

Load The Narrowest Companion

Use a concrete key, URL, ARI, slug, account ID, name, topic, me, or window.

• ../twg-jira/SKILL.md for Jira; ../twg-confluence/SKILL.md for Confluence edits.
• ../twg-status-rollups/SKILL.md for status; ../twg-context-discovery/SKILL.md

for context, dependencies, and graphs.

• ../twg-agentic-search/SKILL.md for fuzzy Rovo/company-knowledge search.
• ../twg-responsibility-routing/SKILL.md for ownership, approval, and escalation.
• ../twg-engineering-work/SKILL.md for code search, PRs, and reviews;

../twg-jira-resolve-merged-work/SKILL.md for stale Jira work backed by merged PRs.

• ../twg-operational-health/SKILL.md for handoffs, reliability, incidents, assets, staffing, and risk.

Rules

• Never guess IDs, flags, slugs, ARIs, object IDs, or mutation contracts.
• For product writes, load the product skill and follow live help.
• Avoid local inspection, caches, schema probes, or diagnostics unless local state is requested.
• For writes, read current state and state the mutation unless execution was requested.