| name | prompt-pattern-library |
| description | Curated, applied prompt-pattern catalog used across this marketplace — decision-tree traversal pre-action prior, alternate-methods-before-blocked, Structured Output Protocol `---RESULT_START---` block, scenario-retrieval inline prior, escalation-by-mandatory-phrasing, citation-aware research, environment-context preamble, orchestrator-worker reinforcement, agent-scenario-authoring frontmatter, claim-grounding/source-honesty marker. Each pattern includes when to use it, what it composes with, an example block, and the failure mode it prevents. Reach for this skill when authoring a new agent, when an existing agent shows a behavior gap that a known pattern would close, or when the `prompt-engineer` is consulting on an agent revision. Used by `prompt-engineer` (primary). |
Skill: prompt-pattern-library
You are about to author or revise an agent file and need to compose its behavior from the marketplace's already-extant patterns. This skill is the catalog. Each pattern has a documented failure mode it prevents — that's the unit of justification for including it in an agent.
The patterns below are not theoretical. Every one of them is already in production use somewhere in plugins/; the references are real. Do not fabricate patterns — if a behavior you want isn't here, that's a signal to either compose from existing patterns or propose a new pattern via the prompt-engineer rather than inventing one inline.
Pattern catalog
1. Pre-action decision-tree traversal
What it is. Before selecting a method for a domain action, the agent traverses the relevant ## Decision Tree: <Domain> — <Situation> Mermaid graph in the active plugin's knowledge files top-to-bottom, resolving condition nodes against the user's stated context, defaulting to the smaller-blast-radius leaf.
What it prevents. The wrong-branch-from-the-start failure mode — the agent picks the wrong method on first try because the available branches were never visible to it. Reactive alternate-methods enumeration catches the case where the chosen method failed; this pattern catches the case where a better method existed and was never tried.
When to compose. Any agent whose domain has multiple methods for the same outcome (e.g. Power Platform: import via pac CLI vs. PA Management API vs. Dataverse Web API). Compose with: alternate-methods (pattern 2) — the tree picks the entry method; alternates kick in on failure.
Example block (drop into agent body):
## Decision-tree traversal (priors)
Before selecting a method for [domain action], traverse the relevant
`## Decision Tree:` section in `plugins/<plugin>/knowledge/<file>.md`:
1. Match the user's situation to the tree's entry condition.
2. Resolve each condition node against the user's stated context, not
keyword pattern-matching.
3. Default to the leaf with the smaller blast radius when multiple branches apply.
4. Escalate to a higher-blast-radius leaf only after the smaller one demonstrably failed.
See [`docs/best-practices/decision-trees-in-knowledge-files.md`](...).
Anti-pattern. Skipping the tree because "the user said X, so it's obviously branch Y." Pattern-matching on user words instead of resolving conditions against context produces the wrong-branch failure 30%+ of the time.
2. Alternate-methods-before-blocked
What it is. The §5 Capability Grounding Protocol rule: when the chosen tool/method/path fails, the agent enumerates ≥2-3 alternate paths to the same outcome, ranks by cost, tries the next-easiest before reporting blocked. The eventual blocked report lists what was tried and what was considered-and-ruled-out.
What it prevents. The "did you try X?" round-trip — the user shouldn't have to prompt for the second attempt; the agent should make it autonomously.
When to compose. Every agent that calls external tools, APIs, or CLIs. Mandatory for dataverse-architect, flow-engineer, power-platform-admin, solution-alm-engineer. Compose with: pre-action decision-tree traversal (pattern 1), mandatory-phrasing block (pattern 5).
Example block:
## Capability Grounding — alternate methods
When a chosen tool, API, or path fails:
1. Enumerate ≥2-3 alternate paths to the same outcome (different API on the same
platform, lower-level surface, different tool, semi-manual procedure with
automation around the boring parts).
2. Rank by cost (time, dependencies, permissions, irreversibility). Easiest first.
3. Try the next-easiest BEFORE reporting blocked.
4. In the blocked report, list:
- What was tried (with one-line outcomes)
- What was considered and ruled out (with reason)
- The remaining options and their cost
Anti-pattern. Reporting "this can't be done" after one failed attempt. Inventing alternatives that don't exist to look thorough.
3. Structured Output Protocol ---RESULT_START--- block
What it is. Every handoff-bearing agent report ends with reasoning/Markdown PLUS a delimited JSON block:
---RESULT_START---
{
"status": "complete | partial | blocked",
"summary": "one sentence",
"deliverables": [...],
"handoff_recommendation": {...} | null,
"confidence": 0.0-1.0,
"risks_or_open_questions": [...],
"next_actions": [...]
}
---RESULT_END---
What it prevents. Unparseable handoffs. Lost reasoning when output is JSON-only. Lost machine-readability when output is Markdown-only.
When to compose. Every agent that hands off to the Team Lead or to a downstream specialist. Skippable for informational chatter ("file read," "test ran") but never for status reports.
Example block: see structured-output.md for the canonical template. Inherit by reference, not by copy — agents that copy the block end up drifting from the canonical shape.
Anti-pattern. Embedding the JSON block inside a Markdown code fence (breaks delimiter parsing). Omitting one of the seven canonical keys. Producing the JSON without the preceding reasoning paragraph.
4. Scenario-retrieval inline prior
What it is. Before answering a plugin-domain-shaped question, the agent globs plugins/<plugin>/scenarios/*.md, filters by tags/product/scope, surfaces ≤3 matches with the mandatory unverified-scenario preamble.
What it prevents. Two failure modes — (a) the agent answers from memory when a relevant war-story field-note exists, (b) the agent treats a scenario as canonical without the unverified disclaimer.
When to compose. Any agent whose domain accumulates field-notes (currently: power-platform/*, future: data-platform, finance, etc.). Compose with: pattern 6 (citation-aware) when the scenario points to a canonical source.
Example block:
## Scenario retrieval (priors)
Before answering a [domain]-shaped question, glob `plugins/[plugin]/scenarios/*.md`
and read the frontmatter of any file whose `tags` or `product` match the user's
context. Surface up to 2-3 matches with the mandatory unverified-scenario
preamble: *"Based on N unverified scenarios from YYYY-MM tagged [scope] — verify
in your environment before applying."*
Treat scenarios as **secondary** to canonical knowledge files; never replace
a `plugins/[plugin]/knowledge/` answer with a scenario, and never elide the preamble.
Full pattern: [`../../skills/scenario-retrieval/SKILL.md`](../../skills/scenario-retrieval/SKILL.md).
Anti-pattern. Citing a scope: tenant-specific scenario without the scope flag visible to the user. Surfacing >3 scenarios (noise). Treating scenario absence as evidence the situation is novel.
5. Mandatory-phrasing block (genuine blockage)
What it is. When an agent has exhausted alternatives and is genuinely blocked, the report uses fixed phrasing:
"After trying [Approach A — outcome], [Approach B — outcome], and [Approach C — outcome], I am blocked on [specific reason]. The remaining options I considered but did not attempt are [X (ruled out because Y), Z (would need permission W)]. I recommend [escalation / next-best path]."
What it prevents. Round-trips. The user reads the blocked report and knows immediately what's left to consider — no "did you try X?" follow-up.
When to compose. Inherit on every agent that can produce a blocked status. Always pairs with pattern 2 (alternate-methods).
Example block: see the Capability Grounding Protocol section in ../../CLAUDE.md "Mandatory phrasing when reporting genuine blockage."
Anti-pattern. Free-form "I tried X and it didn't work" — leaves the user guessing what else was considered. Reporting blocked without listing alternatives that were ruled out.
6. Citation-aware research
What it is. Any research-bearing answer cites the source (URL, file path, doc section) AND labels the source's tier (Tier 1 Consensus / 2 Strong-but-Contextual / 3 Divergent / 4 Emerging / 5 Deprecated) per researcher.md.
What it prevents. Plausible-sounding but unverifiable claims. Silent reliance on Tier 4 emerging guidance presented as Tier 1 consensus.
When to compose. deep-researcher always. Any agent producing release-note text, runbooks, or stakeholder-facing prose. Compose with: pattern 4 (scenario-retrieval) when scenarios corroborate or contradict canonical sources.
Example block:
## Citation discipline
Every load-bearing claim in this agent's output must include:
- The source (URL or repo path)
- The tier label (1 Consensus / 2 Contextual / 3 Divergent / 4 Emerging / 5 Deprecated)
- The verification date (when the source was last checked)
A claim without a citation is treated as the agent's opinion, not as research.
Anti-pattern. Citing a single source as "the answer" when the question is in Tier 3 divergent territory. Citing without verification date — six-month-old citations decay.
7. Environment-context preamble
What it is. Before declaring "I can't do X" or asking the user to authorize an action, the agent reads .ravenclaude/environment-context.md (if present), identifies the current environment, and checks the pre-authorized action categories.
What it prevents. The "did you try X?" round-trip on actions the agent could have just done. Agent asks "can you authorize me to import this solution?" when the env-context file already says the agent is pre-authorized for solution import in DEV.
When to compose. Every agent that takes write actions against external systems. Compose with: pattern 2 (alternate-methods — env-context check runs before alternate enumeration).
Example block:
## Environment-context check (priors)
Before declaring blocked, asking for authorization, or enumerating alternates:
1. Read `.ravenclaude/environment-context.md` at the consumer's project root (if present)
2. Identify the current environment (DEV / TEST / PROD / sandbox / named)
3. If the action category is pre-authorized for the current environment, execute
4. If the action is in the Forbidden list, stop and require per-action confirmation
5. If the file does not exist OR the action category is not listed, fall through
to alternate-methods enumeration
See [`environment-discovery.md`](../environment-discovery/SKILL.md) for auto-discovery
when the file is absent.
Anti-pattern. Treating env-context as a credential store (it isn't). Assuming pre-authorization in DEV applies to PROD. Failing to ask when the file is silent (silence is NOT pre-authorization).
8. Orchestrator-worker reinforcement clause
What it is. A short, declarative clause near the top of every sub-agent file: "This agent does NOT dispatch other agents. When work crosses your scope, complete your portion and return an Escalation note to the Team Lead naming the suggested specialist."
What it prevents. Recursive sub-agent spawning, which the marketplace's hierarchical-dispatch architecture forbids. Without this clause, an architect can be tempted to "just spawn a coder" inline, breaking observability.
When to compose. Every sub-agent except the Team Lead itself. Verified by plugins/ravenclaude-core/hooks/guard-recursive-spawn.sh.
Example block:
## Dispatch boundary
This agent does NOT dispatch other agents. When work crosses your scope,
complete your slice and return a structured Escalation note to the Team Lead
naming the suggested next specialist and the reason. The Team Lead is the
only orchestrator. See [`spawn-team.md`](../spawn-team/SKILL.md).
Anti-pattern. Inline Agent(...) calls in a sub-agent file. "The architect spawned a coder" framing in agent bodies.
9. Agent-scenario-authoring frontmatter
What it is. The YAML frontmatter on agent files includes audience, works_with, scenarios, and quickstart fields per docs/best-practices/agent-scenario-authoring.md. The repo-guide generator picks them up automatically and surfaces them in per-agent cards plus the Overview tab use-case lookup table.
What it prevents. Invisible agents — agents that exist in plugins/<plugin>/agents/ but don't appear in the repo-guide or use-case lookup. Users can't find what they don't know exists.
When to compose. Every agent file, always. At least 3 scenarios per agent; each scenario uses trigger phrasing a real user would type.
Example block:
---
name: <agent-name>
description: <one sentence>
audience:
- <e.g. "Power Platform makers", "Senior backend engineers">
works_with:
- <other-agent-name>
- <other-agent-name>
scenarios:
- trigger: "<user phrasing they'd actually type>"
outcome: "<what this agent delivers>"
- trigger: "..."
outcome: "..."
- trigger: "..."
outcome: "..."
quickstart: |
<2-3 line example of the simplest invocation>
---
Anti-pattern. Generic scenarios ("when you need help with X" — no user types that). Fewer than 3 scenarios — the use-case lookup table thins out and users see "limited coverage."
10. Claim-grounding / source-honesty marker
What it is. For a consequential factual claim about a tool/platform/API (one that gates an irreversible action or gets written into a durable doc), the agent either cites the this-session check that backs it inline and falsifiable (the exact command + output, or file:line) or marks it [unverified — training knowledge] and offers to verify before acting. No High/Med/Low confidence label (uncalibrated). On correction, verify before yielding (don't reflexively concede or dig in). Full protocol: core CLAUDE.md § "Claim Grounding & Source Honesty".
What it prevents. Confident reasoning errors — a flawed mental model stated as fact with no uncertainty marker (e.g. "you can't export solutions as unmanaged" when false), which drives a bad irreversible action. This is the over-claiming-certainty axis that the Capability Grounding Protocol (which covers under-claiming ability) does not.
When to compose. Any agent that makes platform/API/version/default/capability claims a user could act on — especially under GitHub Copilot CLI (Claude/GPT/Grok), where prose rules are weakest and the enforced complements matter most. Composes with: pattern 2 (alternate-methods — abstain only after trying ≥2 paths), pattern 4 (scenario-retrieval, which carries the same [unverified] family), pattern 6 (citation discipline for research output). The same [unverified] marker, source-as-suffix — do not coin a new tag.
Example block:
`pac solution export` defaults to unmanaged `[unverified — training knowledge;
confirm with \`pac solution export --help\`]` — a default that gates the export
command, so it must be verified or marked before acting.
Anti-pattern. Stating a platform behavior as fact from training memory before an irreversible action. A High/Med/Low confidence label on a claim (stamps false claims "High"). Tagging everything — the marker must be rare (consequential-only) to stay informative.
Composing patterns — required vs. optional checklist
When authoring a new agent, walk this checklist:
| Pattern | Required | Notes |
|---|
| 3. Structured Output Protocol block | Required | All handoff-bearing agents |
| 5. Mandatory-phrasing block | Required | Pairs with pattern 2; required when status can be blocked |
| 8. Orchestrator-worker clause | Required | All sub-agents (not Team Lead) |
| 9. Agent-scenario-authoring frontmatter | Required | All agents — repo-guide visibility |
| 2. Alternate-methods-before-blocked | Required if agent calls external tools | Most agents in practice |
| 7. Environment-context preamble | Required if agent takes write actions on external systems | All Power Platform specialists, security-reviewer, etc. |
| 1. Pre-action decision-tree traversal | Required if domain has decision trees | Most domain specialists |
| 4. Scenario-retrieval inline prior | Required if plugin has a scenarios bank | Currently power-platform; extend as banks land |
| 6. Citation-aware research | Required for research-producing agents | deep-researcher always; others when producing stakeholder prose |
A new agent missing ≥2 required patterns scores ≤3 on Dimension 3 (Capability Grounding) of the agent-quality-rubric.
Anti-patterns specific to this skill
- Inventing a pattern because "the agent feels like it needs one." If a behavior gap can't be closed by composing patterns 1-9, propose a new pattern via the
prompt-engineer and get it added to this catalog — don't ship a one-off inline pattern in a single agent.
- Copy-pasting the example block verbatim across 10 agents. Inherit by reference (link to the canonical file). Copies drift.
- Composing patterns in conflict. E.g. asking the user for authorization (pre-environment-context) before checking env-context (pattern 7). Order matters — pattern 7 runs first.
- Treating the checklist as bureaucracy. Each pattern exists because a real failure mode was hit in production. A pattern omitted is a failure mode re-opened.
See also
