Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
{"version":"2.16","created":"2025-02-10T00:00:00.000Z","updated":"2026-08-18T00:00:00.000Z","author":"Ability.ai","changelog":["2.16: Platform-truth refresh (Trinity v0.9.0, tag 93d7ce7c) — Long-Running-Task Rule: turn-end kills background BASH jobs/monitors, but since trinity#2127 a headless run waits for background subagents/forks (bounded by the execution timeout + 300s idle-finalize, rebuilt base image); Foreground-Fork Rule keeps `background: false` as the contract for the right reason (deterministic on every image, mandatory pre-0.9). Reporting Rule guard now also swallows the `requires an agent-scoped API key` refusal — a user/admin-key session sees the tool but cannot report (mcp-server reports.ts)","2.15: Add the Playbook-Call Rule to Design Constraints + a validation-checklist line — a playbook is the unit of inter-agent work (fleet convention protocols/playbook-call.md, operator direction 2026-08-16): anything another agent, schedule, orchestrator, or pipeline stage may call must run from a single `/name [args]` line, declare its inputs in argument-hint (`--run <id>` when part of a larger run), and when called by another agent run only itself and change state only through its own writes and gates. The SKILL.md is the contract — no I/O schema","2.14: Platform-truth refresh (Trinity dev 88a4e2f7) — corrected the flat falsehood that Trinity never reads template.yaml schedules: at creation; since ent#89 it DOES, but entries land disabled unless they declare a literal YAML true, cap at 20, dedup by name, and are never re-applied on recreate, and firing additionally needs the agent's autonomy gate (OFF on every new agent). Reporting Rule updated to the shipped report lifecycle: payload is an object capped at 5 MiB (not 256 KB), display_hint gains `json` and is no longer cosmetic (the customer-facing Workspace renders through it), and list_reports/get_report are taught as read-before-write so a scheduled series resumes instead of duplicating","2.13: Trinity-first docs refresh (verified vs Claude Code 2.1.220) — ship the five tier templates as bundled supporting files (Step 6 pointed at a nonexistent templates/ dir; now ${CLAUDE_SKILL_DIR}/templates/); reframe the frontmatter reference as engine fields vs Trinity platform fields (add disallowed-tools, background:, ${CLAUDE_PROJECT_DIR}, allowed-tools grant semantics + accepted formats, name≡directory note; requires:/automation:/schedule: documented as platform contract, not caveat); replace the Routines advisory with the Trinity scheduling path (schedule: → template.yaml → create_agent_schedule; message invokes by slash name; timeout ≤ agent cap); add the Scheduled-Invocation Rule (scheduled playbooks keep disable-model-invocation: false, schedule messages use the slash name) and the Foreground-Fork Rule (headless-bound context: fork must set background: false — forks run in background by default since 2.1.218 and are reaped at turn-end) with matching validation-checklist lines; official supporting-file names in Step 4b; description/name coaching in Step 2","2.12: Add the Library-Grade Rule to Design Constraints + Step 5b library-target question — a skill destined for a shared skills library declares a requires: frontmatter contract (env keys, packages, binaries), references credentials only as named env vars (never .env reads, no interactive auth, materialize-from-env when a tool demands a credential file), fails with named missing-key errors, and passes a deterministic env-coherence + secret-scan check before contribution","2.11: New Step 9 registers the created skill in the agent's CLAUDE.md — Core Capabilities row + request-phrased Request Dispatch row when the table exists, and resolves the playbook-gap operator-queue item (playbook-gap-<slug>) the skill was created to close","2.10: Correct the stall-watchdog facts in the Long-Running-Task Rule — since trinity#1369 the no-output watchdog is 1800s (not 300s) and watches mcp__* tools only (Bash is unwatched, piping doesn't 're-arm' anything); add set_reminder as the Trinity-side way to verify a decoupled job's artifact without waiting for the next cron","2.9: Add the Reporting Rule to Design Constraints + a validation-checklist line — a skill that yields a surfaceable result (summary, batch, metrics) ends with a guarded mcp__trinity__report step (namespaced report_type, a display_hint, JSON payload) so scheduled/headless runs leave a visible record on the Trinity Reports tab; guarded to skip silently off-Trinity (reporting is an upgrade, never a gate)","2.8: Add the Long-Running-Task Rule to Design Constraints + a validation-checklist line — a headless/scheduled run is one agent turn and CANNOT host a >~10-min job (the harness auto-backgrounds it past the ~10-min sync Bash ceiling, active waiting is blocked, and ending the turn reaps every background task/monitor). Such work must be decoupled to an OS-level cron/systemd/sidecar + done-marker; the run only triggers and verifies the artifact moved. In-turn oversight is an interactive-only affordance, not a headless one","2.7: Generated skills now include the what's-new banner + a seed changelog; documented the required changelog + banner convention for every tier","2.6: Add the Composition Rule — playbooks invoke child skills by name (compose, don't copy); reuse-check step, Composes section, transitive autonomous check","2.5: Add when_to_use/arguments/shell/effort/substitution-vars to frontmatter; fix hot-reload advice; add supporting-files step; add Routines note","2.4: Add Single-Task Rule — scheduled skills must be scoped to one task type per invocation","2.3: Note project-specific vs official frontmatter; list newer official fields (model, context, paths, hooks) for Tier 3","2.2: Add No-Gates Rule — autonomous playbooks cannot have approval gates (breaks execution)","2.1: Add 45-minute rule to Design Constraints — autonomous playbooks must complete within this limit"]}
Create Playbook
ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. create-playbook vX.Y — recent: <summary>. Then proceed.
Create a new skill. Determines the right complexity tier and generates from the appropriate template.
For concepts and patterns, see the README.
Tier templates are bundled with this skill in templates/ — reference them at runtime via ${CLAUDE_SKILL_DIR}/templates/.
Workflow
Step 1: Check for Existing Skills
ls -d .claude/skills/*/ 2>/dev/null | head -10
ls -d ~/.claude/skills/*/ 2>/dev/null | head -10
If similar skill exists, ask: update it, create variant, or proceed with new?
Step 2: Gather Core Requirements
Ask or extract from context:
Name: lowercase-with-hyphens, descriptive. For project/personal skills the directory name is the command (name: is display-only); for plugin skills name: sets the command segment — keep name and directory identical so they never drift.
Purpose → the description: field: third person, what it does + when to use it, key use case first — the skill listing truncates description + when_to_use at 1,536 characters.
Tools needed: Which tools will it use?
Step 3: Determine Complexity Tier
Ask these questions to classify:
Q1: Does this skill read or write any files, APIs, or external state?
NO → Tier 1: Simple Skill (skip to Step 4)
YES → Continue to Q2
Q2: Will this run unattended, on a schedule, or need reliability guarantees?
NO → Tier 2: Stateful Skill (skip to Step 4)
YES → Continue to Q3
Q3: What automation level?
Safe to run completely unattended, no approval needed at any point → autonomous
Needs human approval at checkpoints → gated
Human monitors entire execution → manual
⚠️ Critical: If user says "autonomous" but also mentions approval/review steps, clarify:
"Autonomous playbooks cannot have approval gates — they run unattended and would hang waiting for approval that never comes. Should this be gated instead?"
→ Tier 3: Full Playbook
Step 4: Gather Tier-Specific Requirements
For Tier 1 (Simple Skill):
Process steps (what does it do?)
Outputs (what does it produce?)
For Tier 2 (Stateful Skill):
State dependencies (what files/APIs does it read/write?)
Process steps
Outputs
For Tier 3 (Full Playbook):
State dependencies
Automation level (autonomous/gated/manual)
Schedule (if autonomous or gated): cron expression
Process steps
Approval gates (if gated): where? ⚠️ Not allowed for autonomous — see No-Gates Rule
Prerequisites
Step 4b: Supporting Files
Ask: Does this skill need supporting files (templates, example outputs, helper scripts)?
If YES: plan supporting files alongside SKILL.md using the official names — reference.md (detailed docs), examples.md (usage examples), scripts/ (executables), templates/ (files to fill in) — each referenced from SKILL.md so Claude knows when to load them (they load only when referenced, never automatically). Point at bundled scripts via ${CLAUDE_SKILL_DIR}. Keep SKILL.md under 500 lines — move large reference material to separate files.
If NO: proceed.
Step 4c: Self-Improvement Option
Ask the user:
Should this skill be self-improving?
Self-improving skills include a checklist at the end to consider tactical improvements after each run—things like clearer steps, better error handling, or more efficient flow. The skill's core purpose stays the same; only execution can improve.
If in a git repo, improvements are committed for version control.
If user confirms YES, include the Self-Improvement Checklist (see below) at the end of the generated skill.
Step 4d: Deep Reasoning
If this skill involves complex multi-step logic or architectural decisions, add ultrathink anywhere in the skill body to request deeper reasoning when it runs.
Step 4e: Reuse Check (Composition)
Does any planned step duplicate work an existing skill already does? If so, invoke that skill instead of reimplementing it — Invoke `/child-skill` (namespace cross-plugin calls: /plugin:child-skill ). Then:
Add Skill to allowed-tools.
List each child under a ## Composes section so the dependency is greppable.
Call the unversioned name to ride the latest version (so child fixes propagate automatically); pin /child-vN only to freeze against a child's breaking changes.
See The Composition Rule in Design Constraints. Never call another skill's scripts//reference.md/templates directly — go through the skill entry point.
Step 5: Determine Location
Scope
Location
Use When
Personal
~/.claude/skills/[name]/
Only you use it
Project
.claude/skills/[name]/
Team shares it
Plugin
plugins/[plugin]/skills/[name]/
Distribute widely
Default to project scope unless specified.
Step 5b: Library Target?
Ask: Is this skill destined for a shared skills library — a catalog repo distributed to many agents — rather than just this agent?
If YES: apply The Library-Grade Rule (Design Constraints) during generation — the requires: frontmatter block, env-var-only credentials, named missing-key errors, no host-specific assumptions — and run its deterministic check before finishing.
If NO (default): proceed as agent-local.
Step 6: Generate Skill
Read the appropriate template (bundled with this skill) and fill in the gathered requirements:
Replace every [bracketed] placeholder; drop optional sections (e.g. ## Composes) that don't apply.
Step 7: Confirm and Create
Present summary before creating:
## New Skill: [name]
**Tier**: [1/2/3] ([Simple/Stateful/Full Playbook])
**Automation**: [autonomous/gated/manual/n/a]
**Location**: [path]
**Self-Improving**: [yes/no]
**Library-Grade**: [yes/no]
**State Dependencies**: [list or "none"]
**Process**: [N] steps
**Approval Gates**: [count or "none"]
Create this skill?
After confirmation:
Create directory: mkdir -p [path]
Write SKILL.md
Verify creation
Step 8: Verify
cat [path]/SKILL.md | head -20
Edits to existing skill files hot-reload without restart. A restart is only needed if the top-level skills directory (~/.claude/skills/ or .claude/skills/) didn't exist before this session. Plugin-hosted skills refresh with /reload-plugins.
Step 9: Register in the Agent's CLAUDE.md
If the agent's CLAUDE.md lists capabilities, register the new skill so it's discoverable and dispatchable:
Add a row to ## Core Capabilities (skill + purpose).
If a ## Request Dispatch table exists, add a row phrased as the incoming request that should route to this skill — what gets asked, not the skill name restated.
If this skill was created to close a flagged playbook gap, resolve the flag: on Trinity, find the playbook-gap-* operator-queue item (mcp__trinity__list_operator_queue) and resolve it via mcp__trinity__respond_to_operator_queue, noting the new skill's name.
Skip silently when CLAUDE.md has no such sections, or for internal helper skills not meant for direct dispatch.
---name:skill-namedescription:Whatitdoesallowed-tools: [tools]
user-invocable:true---# Skill Name## Purpose## State Dependencies|Source|Location|Read|Write|## Process## Outputs
Tier 3 - Full Playbook:
---name:playbook-namedescription:Whatitdoesautomation:gated# Trinity platform field (see note below)schedule:"0 9 * * 1"# Trinity platform field, optionaldisable-model-invocation:false# keep false for scheduled playbooks (Scheduled-Invocation Rule)allowed-tools: [tools] # include `Skill` if it invokes other skillseffort:high# optional: low/medium/high/xhigh/maxuser-invocable:true---# Playbook Name## Purpose## State Dependencies## Prerequisites## Composes # optional: child skills this playbook invokes## Process### Step 1: Read Current State### Step N: [Work]### Final Step: Write Updated State## Completion Checklist## Error Recovery
Every tier (required): include the metadata: block (with a newest-first changelog) and the what's-new banner shown in the Tier 1 example, placed right after the H1. The banner surfaces the top (newest) changelog entry on launch, so keep the changelog newest-first. See the repo CLAUDE.md → "Skill Changelog & What's-New Banner".
Frontmatter: Engine Fields vs Trinity Platform Fields
Skills here run on the Claude Code engine and deploy to the Trinity platform, and their frontmatter mixes both layers — both are load-bearing. The engine ignores unknown frontmatter by design, so platform fields are safe extensions, not hacks.
allowed-tools — a permission grant: listed tools run without prompts while the skill's turn is active, and the grant clears on the next user message. It does not restrict Claude to those tools. Accepts space/comma-separated strings or YAML lists; supports Bash rule syntax (Bash(git add *)) and ${CLAUDE_SKILL_DIR} substitution
disallowed-tools — the actual restriction: tools removed while the skill is active
user-invocable, disable-model-invocation — ⚠️ scheduled playbooks must keep disable-model-invocation: false (see the Scheduled-Invocation Rule)
when_to_use — additional trigger context; combined with description, capped at 1,536 chars in skill listing
arguments — named positional args: arguments: [issue, branch] → $issue, $branch in content
model, effort — override model/effort while the skill is active (model: inherit keeps the session model; effort low/medium/high/xhigh/max)
shell — bash (default) or powershell for ! command blocks
context: fork + agent: + background: — run the skill in a subagent; forks run in the background by default — set background: false for anything bound for scheduled/headless use (see the Foreground-Fork Rule)
${CLAUDE_SKILL_DIR} — absolute path to the skill's directory (use for bundled scripts and templates)
${CLAUDE_PROJECT_DIR} — project root (the same path hooks receive)
Trinity platform fields (this plugin's playbook model — read by the platform and fleet tooling, not by the engine):
metadata: block with version, changelog (newest-first), author — required on every skill: bump version and prepend a changelog entry on each edit, and pair it with the what's-new banner after the H1 (see the repo CLAUDE.md → "Skill Changelog & What's-New Banner")
automation: autonomous | gated | manual
schedule: "<cron>" — the durable schedule declaration; see below
requires: — the library-grade dependency contract (env keys, packages, binaries) — see the Library-Grade Rule
How a schedule: becomes a live schedule (Trinity): the per-skill schedule: feeds the agent's template.yamlschedules: block. Since ent#89 Trinity does materialize that block at agent creation — but every entry lands disabled unless it declares a literal YAML enabled: true (a non-boolean is treated as false), at most 20 entries are materialized, names are deduped against the agent's existing schedules, timezone: defaults to UTC, and the block is never re-applied on recreate. Firing also requires the agent's autonomy gate, which is OFF on every new agent. /trinity:onboard / /trinity:sync are still what reconcile the block onto an already-live instance. The schedule's message is the prompt the agent receives — it must invoke the skill by its slash name — and its timeout_seconds must fit the agent's execution cap (default 3600s). The skill must also work when invoked manually, without Trinity — Trinity is the upgrade, never the gate.
When generating Tier 3 playbooks, keep the platform fields (the rest of the plugin depends on them) and add engine fields like model:, context: fork, or paths: when they fit.
Self-Improvement Checklist (Appendix)
When user opts into self-improving skills, append this section to the generated skill:
## Self-Improvement
After completing this skill's primary task, consider tactical improvements:
- [ ] **Review execution**: Were there friction points, unclear steps, or inefficiencies?
- [ ] **Identify improvements**: Could error handling, step ordering, or instructions be clearer?
- [ ] **Scope check**: Only tactical/execution changes—NOT changes to core purpose or goals
- [ ] **Apply improvement** (if identified):
- [ ] Edit this SKILL.md with the specific improvement
- [ ] Keep changes minimal and focused
- [ ] **Version control** (if in a git repository):
- [ ] Stage: `git add <skill-path>/SKILL.md` - [ ] Commit: `git commit -m "refactor(<skill-name>): <brief improvement description>"`
Design Constraints
The 45-Minute Rule: Agent reliability degrades exponentially after ~45 minutes of continuous execution. Design playbooks accordingly:
Autonomous playbooks must complete in under 45 minutes
If a task is larger, break it into multiple scheduled runs (e.g., "process 50 items" not "process all items")
Build checkpoints where state is saved — if interrupted, the next run can resume
Long processes → multiple scheduled tasks with handoff via state files
On Trinity this is a hard ceiling, not just a heuristic: a schedule's timeout_seconds must fit the agent's execution cap (default 3600s / 60 min) — 45 minutes is the headroom that survives it
When gathering requirements for Tier 3 playbooks, ask: "Can this complete in under 45 minutes? If not, how should we chunk it?"
The No-Gates Rule for Autonomous Playbooks: Autonomous playbooks run unattended on a schedule — there is no human to approve gates. An [APPROVAL GATE] in an autonomous playbook will cause execution to hang indefinitely, breaking the scheduled run.
Autonomous playbooks MUST NOT contain any [APPROVAL GATE] markers
If the workflow needs human approval at any point, it MUST be gated or manual, not autonomous
When user requests autonomous + approval gates, explain the incompatibility and ask them to choose
The Scheduled-Invocation Rule: A scheduled playbook must stay invocable by the scheduler:
Keep disable-model-invocation: false (the default). A natural-language schedule message reaches the skill through model invocation, which disable-model-invocation: true blocks — it also blocks preloading into subagents.
Write the schedule message to invoke the skill by its slash name (/skill-name args), never as a paraphrase — on Trinity a message naming a missing skill fails loudly as SKILL_NOT_FOUND; a paraphrase that half-matches just drifts silently.
The Single-Task Rule for Scheduled Skills: Autonomous playbooks execute in a single context window. Iterating over multiple different tasks (e.g., "process all backlog items") fills that window with context from each prior item, adding noise to every subsequent step.
Each scheduled invocation must be scoped to one task or one task type
Process one item per invocation; let the scheduler re-invoke for the next
Exception: batch tasks where every item has identical context needs (same files, same pattern) are fine — e.g., running the same quality gate on N wizard files all read the same kind of data
When a user asks for a scheduled loop, design it as single-item-per-invocation and explain that the cron handles repetition
The Long-Running-Task Rule (>~10 min): A headless/scheduled run is a single agent turn, and it cannot host a job longer than the synchronous Bash window (~10 min max tool timeout) — a hard platform ceiling, not something a bigger budget fixes. Design around it:
Don't try to babysit it in-turn. Past ~10 min the harness auto-backgrounds the job (not your choice), active waiting is blocked (sleep/poll loops rejected — "use monitor with an until-loop"), and the moment the turn ends, every background Bash job and monitor spawned in it is killed (the completion event fires as killed, not completed; the promised re-invoke never comes). Since Trinity v0.9.0 (trinity#2127, rebuilt base image) a headless run does wait for background subagents/forks — up to the execution timeout, finalizing after 300s of stream idleness (AGENT_IDLE_FINALIZE_S) — but that changes nothing for shell jobs, and the ~10-min Bash ceiling stands. Streaming heartbeat output changes nothing — the no-output stall watchdog (AGENT_TOOL_STALL_LIMIT_S, default 1800s) watches mcp__* tools only since trinity#1369, Bash isn't watched, and neither the ~10-min ceiling nor the turn-end reaping cares about output. The async monitor / re-invoke model works interactively (the session persists) but NOT in a headless run.
≤ ~10 min: run it as one foreground, un-piped, streaming Bash call, in-turn. Don't pipe through tail/grep (buffers output and hides live progress from the transcript).
> ~10 min (a FAISS/index rebuild, full bootstrap, bulk embedding, a big migration): it must run outside the agent turn — an OS-level job (container cron / systemd unit / small non-LLM sidecar) that builds the artifact and writes a done-marker. The scheduled skill only triggers it and does the fast parts: check the marker / artifact freshness and, if fresh, run the quick follow-ups. On Trinity, the triggering run can also set_reminder (one-shot deferred self-trigger, trinity#1296) to come back and verify the artifact landed instead of waiting for the next cron tick.
Always verify the artifact moved (mtime advanced and a stats count > 0) before declaring success — never trust the exit code or business_status. A run that ends without the artifact changing is a failure, not a skipped.
The Foreground-Fork Rule (headless forks): context: fork skills run in the background by default. A headless/scheduled run is one agent turn; on agents whose base image predates Trinity v0.9.0 (trinity#2127) turn-end reaped every background fork, so a scheduled playbook that invoked a background-forked skill silently lost it — and on v0.9.0+ the wait is bounded by the execution timeout and a 300s idle-finalize, so a slow fork can still be cut. Any skill that uses context: fork and is bound for scheduled/headless use — or is composed by a playbook that is — sets background: false (wait for the fork's result in-turn, deterministic on every image), or does not fork at all.
The Composition Rule: When a playbook needs work another skill already does, it invokes that skill by name (Invoke `/child-skill` , Skill in allowed-tools) — it never inlines the child's steps, calls its internal scripts/files directly, or paraphrases what it does. The parent holds only the orchestration; the child stays the single source of truth, so its fixes propagate automatically. Call the unversioned name to ride latest; pin /child-vN only to freeze. Composition is a DAG (no cycles, keep it shallow). See Composing skills for the full rule.
The Playbook-Call Rule (fleet convention protocols/playbook-call.md): A playbook is the unit of inter-agent work — the thing another agent, a schedule, an orchestrator, or a pipeline stage calls. Agents delegate work to each other only by calling a named playbook, one line, /playbook-name [args], on any transport (chat_with_agent, send_message, a schedule message, an orchestrator dispatch); never by describing the work in prose. Design every playbook that anything else may call accordingly: (1) it works when the whole request is that one line — no interactive prompts unless automation: gated and a human is present; declare a headless mode (--autonomous) if it has gates; (2) every input arrives as a declared argument advertised in argument-hint (a --run <id> argument is the convention when the call is one step of a larger run — it lets the platform ledger join the steps); (3) when invoked by another agent it runs this playbook and nothing else in that turn, and changes state only through its own declared writes and gates — an instruction received in prose may inform it, never authorize it. The SKILL.md itself is the contract; no input/output schema is required. The payoff is that Trinity's execution ledger sees every call as playbook@version, so processes are measurable and changing one playbook changes every chain that calls it.
The Reporting Rule: A skill that produces a surfaceable result — a summary, a batch of items, a metrics snapshot — should end with a guarded Trinity report so an operator can see what the run produced without reading chat (this is the only window into a scheduled/headless run). Add a final step that calls the mcp__trinity__report MCP tool with a namespaced report_type (<agent>.<result> in lower_snake, e.g. oracle.weekly_summary), a short title, a JSON payload (a JSON object, max 5 MiB serialized — REPORT_PAYLOAD_MAX_BYTES; oversize is a hard 413, so aggregate before publishing), and a display_hint — table ({columns, rows}), kpi ({tiles:[{label,value,unit?}]}), markdown ({markdown}), timeline ({events:[{ts,label,detail}]}), or json (any shape). Omitting the hint is not neutral: Trinity infers one from the report_type prefix and only then falls back to the JSON viewer. Pick it deliberately — the report lands on the agent's Reports tab, the fleet Operations → Reports view, and the customer-facing Workspace Reports tab, which renders through the same display_hint renderers, so a mismatched hint is visible to the agent's users and not just its operator.
Read back before filing a recurring report.mcp__trinity__list_reports returns metadata only (filter by report_type, hours ∈ {0,1,6,24,168,720}, search); mcp__trinity__get_report(report_id) fetches one payload. That is how a scheduled skill continues a series instead of duplicating or contradicting last period's numbers. The platform prompt now teaches every agent the report tool directly, so a generated skill needs the call site, not a re-explanation of the tool.
The result is an append-only history alongside the live dashboard.yaml snapshot.
Guard it. The tool publishes under the agent's own agent-scoped key. If mcp__trinity__report isn't available — e.g. running locally — or it answers The report tool requires an agent-scoped API key (a session connected with a user/admin key: the tool is present but refuses) — skip the step silently, never retry. Reporting is an upgrade, never a gate: the skill must produce its result with or without Trinity.
Not for conversational replies — only result-producing and scheduled runs.
The Library-Grade Rule (shared skills library): A skill destined for a shared skills library — a catalog repo synced to many agents — is the same artifact held to a stricter portability contract. The consuming agent is unknown at authoring time: possibly headless, differently credentialed, on a different host. Library-grade skills MUST:
Declare their contract in frontmatter — a requires: block listing every env key the skill reads plus its runtime deps, so a platform or reviewer can check fulfillment before the skill ever runs:
requires:env: [SLACK_BOT_TOKEN] # every credential/config key the skill readspackages: [playwright] # runtime packages (pip/npm), if anybinaries: [ffmpeg] # required executables, if any
Keep automation: and user-invocable accurate — they are part of the same contract.
Reference credentials only as named env vars — never read .env or other credential files directly. The environment is the interface; how keys get there is the consuming platform's business, and coupling to the delivery mechanism breaks the skill the moment it changes.
Fail with a named, actionable error when a key is missing — e.g. "SLACK_BOT_TOKEN not set — add it to this agent's credentials" — never a silent skip or a generic failure.
Never assume interactive credential acquisition (gcloud auth login, browser OAuth) — the consuming agent may be headless.
Materialize-from-env when a tool demands a credential file: write the env var's content to a temp file at runtime and point the tool at it. The env var stays the contract; the file is the skill's own runtime artifact.
Carry no host-specific assumptions — no absolute paths, no workspace files not shipped inside the skill directory; reference bundled scripts via ${CLAUDE_SKILL_DIR}.
Every name in that list must appear in requires.env — and every declared key must actually be referenced (no decorative declarations). Then scan the skill directory for literal secrets (tokens, keys, passwords): there must be none. A failed check is a blocker for contribution, never an advisory.
Autonomous Playbook Validation Checklist
Before generating any autonomous playbook, verify:
No approval gates — grep the content for [APPROVAL GATE] — must return zero matches
No human decision points — no "ask user", "wait for confirmation", "present options"
Complete error handling — all failure paths handled without human intervention
Notifications on failure — errors must alert via Slack, email, or logging
Under 45 minutes — execution time within agent reliability window
No in-turn job over ~10 min — a headless run can't host a job past the ~10-min sync Bash ceiling (auto-backgrounded, then reaped at turn-end); anything longer (index rebuild, bulk embedding, big migration) is decoupled to an OS-level cron/systemd/sidecar + done-marker, and the run only triggers + verifies the artifact moved (never off an exit code / business_status)
Idempotent or safe to retry — can re-run without causing duplicate effects
Single-task scope — processes one task type per invocation; iteration over varied items happens across invocations, not within one
Composed children are autonomous-safe — autonomy is transitive: recurse into every /invoked skill; none of them may contain [APPROVAL GATE] or human decision points, and the whole tree must fit the 45-minute / single-task budget
Invocable when scheduled — disable-model-invocation is false/absent, and the schedule message invokes the skill by slash name (the Scheduled-Invocation Rule)
Callable as one line — the playbook runs correctly when the entire request is /name [args] from another agent or a schedule; args are declared in argument-hint; no undeclared prompt blocks a headless caller (the Playbook-Call Rule)
No background forks — the skill and every composed child using context: fork sets background: false (the Foreground-Fork Rule) — a background fork is reaped at turn-end in a headless run
Result-producing runs report — a skill that yields a surfaceable result ends with a guarded mcp__trinity__report step (the Reporting Rule), skipped silently when the tool is absent — so a scheduled/headless run leaves a visible record on the Reports tab
If any check fails, the playbook cannot be autonomous. Recommend gated instead.
