Prompts Reference

Prompts are multi-step, guided workflows — unlike a single tool call, invoking a prompt hands the agent a full script of steps to follow, chaining several tool calls together with decision logic in between.

​

create_project

Arguments: none — discovers your organization automatically via context_get. Trigger phrases: “create a new project”, “start a new engineering initiative”, “set up a project for X” Creates a new project and immediately hands off into scoping it. Steps:

1. Calls context_get to resolve the organization ID. If the caller has no organization, stops and tells them to contact an admin.
2. Calls projects_phases_list to get available phases.
3. Asks the user for a project name, a 1–3 sentence description, and which phase it’s currently in.
4. Calls projects_create with those answers.
5. Confirms creation, then automatically invokes scope_planning_items for the new project.

​

scope_planning_items

Arguments: projectId (required). Trigger phrases: “what framework items apply to this project”, “help me scope this project”, “expand scope”, “add more checklist items” Runs a discovery conversation to select which framework checklist items are relevant to a project — not every item is meant to apply to every project. Steps:

1. Calls projects_get and planning_list_items to see current scope.
2. Asks 3–4 discovery questions: primary technology/architecture, team size and new-vs-existing system, reliability/security/compliance requirements, deadlines or external dependencies.
3. Calls framework_list_categories and reasons about relevance per category given the answers, skipping items already in scope.
4. Presents recommendations grouped by category, each marked [RECOMMENDED] or [OPTIONAL] with a one-sentence rationale, and lists any categories intentionally skipped.
5. Lets the user add, remove, or approve items, then calls planning_create_item for each approved item (silently skips ones already in scope rather than surfacing a conflict error).
6. Offers to hand off into start_planning_session or generate_artifact for one of the newly-scoped items.

​

start_planning_session

Arguments: projectId (required). Trigger phrases: “start planning”, “review planning health”, “what’s missing on this project” Analyzes a project’s current planning health and surfaces prioritized gaps. Steps:

1. Calls projects_get and planning_get_status.
2. If the project has zero scoped items (totalItems === 0), stops and hands off to scope_planning_items instead — there’s nothing to report gaps on yet.
3. Otherwise presents a summary (phase, completion %, status breakdown, integration config status), then calls planning_get_gaps with the project’s current phaseId.
4. If the org has Phase Mapping configured (hasPhaseMapping: true) and there are phase-scoped gaps, presents those first. If phase-scoped gaps are empty, falls back to all gaps. If no phase mapping exists, presents all gaps and prioritizes by phase relevance manually.
5. Presents the top 5 gaps (prioritizing items with a template available), then asks the user to pick one — which hands off into generate_artifact — see the full list, or say “expand scope” to hand off into scope_planning_items.

Does not generate any artifacts unless the user explicitly picks an item.

​

generate_artifact

Arguments: projectId (required), checklistItemSlug (required — use framework_list_items to find the slug if you only have the item’s name). Trigger phrases: “generate / write / create / help me with [artifact name]” — e.g. “write the rollback plan for this project” Produces a planning artifact as a markdown document, using the item’s template and the project’s context. Steps:

1. Calls planning_get_item_context to get everything needed in one call.
2. If the item is already complete with an artifact URL, confirms with the user before regenerating.
3. Summarizes what it’s about to generate (item, purpose, template sections).
4. Generates the artifact: follows the template structure exactly if one exists; fills sections from real project context (name, description, phase, team, integration config); inserts [TODO: ...] placeholders for anything it doesn’t know rather than fabricating details like SLA numbers or capacity estimates; if no template exists, infers structure from the definition and says so at the top.
5. Presents the draft and asks for approval, revising in a loop until the user is satisfied.
6. Once approved, gives the user the suggested file path ([docsPath]/[checklistItemSlug].md) and the exact markdown to commit, and asks them to commit it and share the resulting URL.
7. When the user provides the URL, calls planning_update_item with status: "complete" and that artifactUrl.

Never marks an item complete without an artifact URL the user actually provided, and never generates content outside this workflow.

​

review_artifact

Arguments: projectId (required), checklistItemSlug (required). Trigger phrases: “review / check / evaluate [artifact]” — e.g. “review our rollback plan” Reviews an existing (or in-progress) artifact against its checklist item’s definition and template, and produces structured feedback. Steps:

1. Calls planning_get_item_context for the item’s definition, template, status, and artifact URL.
2. If there’s no artifact yet (not_started, no URL), offers to switch to generate_artifact instead and stops.
3. If an artifact URL exists, asks the user to paste its current content (the server doesn’t fetch external URLs itself). If the item is in_progress with no URL, asks for the draft content directly.
4. Reviews the content against the definition, template structure, and whether it’s specific to the project or generic boilerplate.
5. Produces a structured review: Overall Assessment, What’s Working, Gaps (each tied to a specific missing definition/template requirement), Recommendations (specific, not vague — e.g. names the exact section and what it’s missing), and a Suggested Status (complete, in_progress, or not_started).
6. Offers to either help revise the artifact (looping back into the generate_artifact finalize-and-record steps) or just update the planning record status via planning_update_item.