Create, edit, search, transition, close, and triage Jira tickets via the Atlassian MCP. Use when asked to file a bug, request a feature, create a task, log a defect, search the backlog, triage findings into the tracker, edit ticket fields, transition status, or manage Jira work items. Also use when the user says 'create a Jira issue', 'file a bug', 'open a ticket', 'add to backlog', 'search Jira', 'close ticket', 'move to Done', or names any Jira issue key (e.g. 'PROJ-123'). Handles type discovery, parent linking, label assignment, duplicate detection via JQL, status transitions, and issue-link creation. Defers all field-content formatting to the `jira-syntax` skill. Do NOT use for pull requests, changelog entries, non-Jira trackers (GitHub Issues, Linear, GitLab), or managing local TODO.md.
Instalación
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Create, edit, search, transition, close, and triage Jira tickets via the Atlassian MCP. Use when asked to file a bug, request a feature, create a task, log a defect, search the backlog, triage findings into the tracker, edit ticket fields, transition status, or manage Jira work items. Also use when the user says 'create a Jira issue', 'file a bug', 'open a ticket', 'add to backlog', 'search Jira', 'close ticket', 'move to Done', or names any Jira issue key (e.g. 'PROJ-123'). Handles type discovery, parent linking, label assignment, duplicate detection via JQL, status transitions, and issue-link creation. Defers all field-content formatting to the `jira-syntax` skill. Do NOT use for pull requests, changelog entries, non-Jira trackers (GitHub Issues, Linear, GitLab), or managing local TODO.md.
Manage Jira tickets via the Atlassian MCP. Every ticket must be self-contained, professional, and actionable without prior context. The skill applies only to Jira sites the Atlassian MCP is connected to; if the MCP is unavailable, stop and tell the user.
Prerequisites
Atlassian MCP server is available. MCP is the only path to Jira - never curl, never bare HTTP, never gh.
jira-syntax skill is loaded before composing any field content. Jira description and comment fields are wiki markup, not Markdown.
Discover project metadata (run first)
Before any create, edit, search, or transition operation, fetch the live metadata for the target site and project:
Need
Tool
Jira site cloudId
getAccessibleAtlassianResources
Visible projects and their keys
getVisibleJiraProjects
Types accepted by chosen project
getJiraProjectIssueTypesMetadata
Required fields for a chosen type
getJiraIssueTypeMetaWithFields
Site-wide issue link types
getIssueLinkTypes
Valid transitions for an issue
getTransitionsForJiraIssue
Resolve display name to account id
lookupJiraAccountId
Use the values verbatim. Never memorize project keys, type names, transition IDs, label names, or account IDs across sessions.
Branching on what discovery reports:
Discovery reports
Behavior
getAccessibleAtlassianResources returns empty
Stop. The MCP is not connected to any Jira site for this account.
getVisibleJiraProjects returns multiple
The user must name a project key. List keys and short names; ask.
Chosen type missing from getJiraProjectIssueTypesMetadata
List accepted types; ask which fits. Never silently substitute a different type.
Ask one question per required field before drafting.
Body shape
Tickets share three common sections; the middle is type-specific.
Section
Position
Purpose
Summary
Top
One paragraph stating what the ticket addresses. Not a restatement of the title.
(type-specific)
Middle
Background, observed behaviour, root cause, proposed solution, scope, etc. - varies per type.
Requirements
Near end
Observable outcomes in MUST / MUST NOT form. Each independently verifiable.
[Self-checks]
Optional
Automated (CI assertions) and Manual (operator-runnable scenarios). Encouraged for Bug, Story, Task.
Context
Bottom
Origin (where the ticket came from), verbatim operator/reviewer quote, related tickets, tracking links.
The per-type templates in assets/ realize this shape with the right middle sections for each type. Some sections are bracketed […] in the templates - drop them when there is no content.
Type classification
Pick exactly one. When two rows match, ask one question that decides between them.
MUST have a parent. JQL project = {key} AND issuetype in (Story, Task, Bug) AND statusCategory != Done → present keys → ask.
Story / Task / Bug
Parent Epic is optional. If the user did not name one, run project = {key} AND issuetype = Epic AND statusCategory != Done ORDER BY rank ASC and offer the list; "none" is a valid answer.
Epic / Spike
No parent.
Before creating a ticket
Duplicate check (BLOCKING)
Run before every createJiraIssue call. Use searchJiraIssuesUsingJql with a keyword query:
project = {KEY} AND text ~ "<keywords from the user's request>" AND statusCategory != Done
ORDER BY created DESC
Exact duplicate. Stop. Report the existing ticket key.
Partial overlap. Mention the related ticket. Ask whether to proceed.
No match. Proceed.
Create
Body rules
Language. English.
Tone. Professional, concise, no filler.
Privacy. Never include usernames, API keys, internal URLs, tokens, or personal information, regardless of whether the project is private or read-restricted. Jira is observed by people outside the immediate authoring team.
Self-contained. A reader who has never seen this work must understand the ticket from the body alone.
Cite source by file:line. When the ticket refers to existing code, constants, or behaviour, anchor each claim to a path and line range (e.g. src/services/foo.ts:123-145). File:line citations turn the ticket into a self-auditable record.
Preserve verbatim quotes. When the ticket originates from a specific operator, reviewer, or customer report, include the exact wording in the Context section. Paraphrasing loses the reporter's mental model and removes the literal phrase a future searcher will type.
No hard wrapping mid-sentence. Write each paragraph as a single line; Jira handles flow at render time. Hard-wrapping mid-paragraph creates spurious paragraph breaks in wiki markup and noisy diffs on edit.
Bugs describe problems, not solutions. Steps to reproduce, Expected, Observed describe the problem. When the reporter investigated and has a concrete fix in mind, it goes into the optional Proposed solution section as a suggestion, not as a mandate that bypasses Requirements.
Requirements. Required for Story and Bug. Optional for Task, Spike, Epic, Sub-task (include when a concrete completion signal exists). Write each requirement in MUST / MUST NOT form. Each is independently verifiable. Reserve MAY for genuinely optional outcomes.
Markup. Jira wiki markup via jira-syntax (NOT Markdown).
Never reference internal artefacts. ADR numbers, architecture section IDs, doc paths, ticket IDs in source-code comments - none of these belong inside the ticket body unless the user explicitly asked. Those identifiers live in specs, not in work items.
Body templates
Pick the template for the classified type from the table in "Type classification" above. Copy the ## Template block verbatim, then fill every section. Drop bracketed {...} placeholders that do not apply. Required sections (no brackets) stay.
On rejection of a custom field: drop that field, retry with required-only, surface the omitted field in the report so the user can backfill via editJiraIssue.
Batch creation
When creating multiple related tickets (e.g. an Epic with child Stories):
Present all planned tickets as a numbered list (title, type, parent, labels) before creating any.
Wait for user confirmation.
Create sequentially. Report each ticket key after creation.
Print a summary table when done:
| # | Key | Type | Parent | Title | Labels |
|---|-----|------|--------|-------|--------|
Confirm before create
Print this block. Wait for explicit user approval before calling createJiraIssue.
Draft:
Project: {KEY} ({name})
Type: {type}
Parent: {KEY or "none"}
Title: {composed title}
Labels: {labels or "—"}
Body:
{full body, wiki markup}
Reply "create" to file the ticket, or send edits.
Silence, "ok", or unrelated follow-ups are not confirmation.
Report after create
Created: {KEY} — {title}
URL: {browseUrl}
Type: {type}{, parent: PARENT-KEY}
Labels: {labels or "—"}
Inferred:
- {field}: {value} # each field set without explicit user input
Links:
- {KEY-A} {linkType} {KEY-B} # each created link, or "—"
Inferred is the user-care line: it lists every field the agent chose without being told. If empty, write —.
Search, Edit, Transition
MCP invocations for these three operations are catalogued in references/jira-recipes.md. Load that file when the user asks to find, modify, or transition a ticket. For broad searches over a populated project, follow the "Querying large projects without overflowing tool output" recipe in that file: pass narrow fields, cap maxResults, and avoid whole-project ORDER BY dumps. The rules below are policy and bind regardless of which recipe is used:
Edit. Read current state via getJiraIssue before destructive edits (body replacement, label replacement, type change). Confirm destructive edits with the user before executing. Pass only the changing fields - do not resend unchanged fields.
Transition. Match the user's intent ("close", "done", "in progress", "in review") to a transition name returned by getTransitionsForJiraIssue; transition IDs vary per project workflow. After transitioning, verify the new status via getJiraIssue.
Linking. Use getIssueLinkTypes to confirm the type name and read its inward/outward labels. For "Blocks", inwardIssue is the blocker and outwardIssue is the blocked: a link created with inwardIssue = X and outwardIssue = Y renders in the UI as "Y is blocked by X". The inward/outward field names are not self-evident, so do not reason direction out from them - confirm against the type's own labels, and after creating any directional link verify orientation by re-reading one endpoint via getJiraIssue and inspecting issuelinks, or by checking the rendered relationship in the UI. This MCP exposes no delete-issue-link tool: a wrong-direction link succeeds silently and cannot be removed programmatically, so get direction right before the call and flag any stray link id for the user to delete manually. If the link fails, report the ticket as created and the link as pending - do not delete the ticket to retry from scratch.
Triage
When triaging a finding from code review, logs, discussion, or PR feedback into the backlog:
Extract the concern. State it in one sentence.
Duplicate check. Run the BLOCKING JQL query from "Before creating a ticket".
Classify type per the table in "Type classification".
Resolve parent per "Parent matching".
Draft the body using the matched template.
Present the full createJiraIssue payload for review before executing.
Quality checklist
Before executing createJiraIssue, verify:
getAccessibleAtlassianResources returned a cloudId
Project key is explicit (named by user or unambiguous from getVisibleJiraProjects)
Chosen type appears in getJiraProjectIssueTypesMetadata
Required custom fields from getJiraIssueTypeMetaWithFields are filled
Duplicate check performed (JQL keyword search)
Title: imperative, ≤ 80 chars, no trailing period, no [type] prefix, code identifiers backtick-wrapped
Body matches the type's template; bracketed sections dropped if empty
Requirements present for Bug and Story, in MUST / MUST NOT form
Source-code references cite file path and line range
Verbatim quote preserved in Context when the ticket originated from a specific report
Self-checks present for non-trivial Bug and Story tickets (Automated + Manual)
Parent set for Sub-task (and for Story/Task/Bug if the user named an Epic)
No usernames, keys, internal URLs, tokens, or personal information in the body
No solution prescribed inside the problem-description sections of a Bug (use Proposed solution instead)
No ADR numbers, doc paths, or ticket IDs in the body unless explicitly requested
Labels exist on prior issues in the project (no invented labels) - or omitted
Assignee/priority/reporter only set when the user named them
Body passed through jira-syntax (wiki markup, not Markdown)
Error recovery
Error
Recovery
Atlassian MCP unavailable
Stop. Report that Jira requires MCP; ask user to enable it.
getVisibleJiraProjects returns multiple
List keys and short names; ask which to use.
Classified type missing from getJiraProjectIssueTypesMetadata
List accepted types; ask which fits the request.
createJiraIssue rejects a required field
Re-read getJiraIssueTypeMetaWithFields; ask user for the value; retry.
createJiraIssue rejects an unknown field
Drop the field; retry; report omission so the user can backfill.
Sub-task created without parent (parent silently ignored)
Set parent via editJiraIssue immediately; report in the post-create summary.
createIssueLink creates a reversed link (succeeds silently, or rejects with "wrong direction")
For "Blocks", inwardIssue = blocker, outwardIssue = blocked; verify by re-reading issuelinks via getJiraIssue or checking the rendered UI. A silently-reversed link cannot be deleted via this MCP - flag its link id for the user to delete and create the correct link.
lookupJiraAccountId returns multiple users
Ask the user to disambiguate by email or unique substring.
transitionJiraIssue rejects the transition ID
Re-run getTransitionsForJiraIssue (workflow may have changed); pick from the fresh list.
User-named issue key not found by getJiraIssue
Verify project key prefix; if still missing, ask whether the key was typed correctly.
Constraints
One ticket per createJiraIssue invocation. Loop for batches; never run a hidden create-loop.
Never auto-confirm. The user's explicit "create" is the only gate.
Never invent labels. Use labels the user named, or labels already present on issues read from the same project. When unsure, omit.
Never set priority, assignee, or reporter unless the user named them. Resolve display names via lookupJiraAccountId; never write a raw email into an account-id field.
Never reference internal artefacts (ADR numbers, architecture section IDs, doc paths, ticket IDs) inside the ticket body unless the user explicitly asked. Those identifiers belong in specs, not in work items.
Never use gh or curl for Jira operations. MCP is the only path.
All field text passes through jira-syntax first. Jira fields are wiki markup, not Markdown.
After a destructive edit (body replacement, label replacement, status transition), verify by re-reading the ticket via getJiraIssue and include the verified state in the report.