Create Work Item

!accelerator config context --skill create-work-item --fail-safe !accelerator config agents --fail-safe

If no "Agent Names" section appears above, use these defaults: accelerator:reviewer, accelerator:codebase-locator, accelerator:codebase-analyser, accelerator:codebase-pattern-finder, accelerator:documents-locator, accelerator:documents-analyser, accelerator:web-search-researcher.

Work items directory: !accelerator config path work --fail-safe Active integration: !accelerator config work integration --fail-safe ID pattern: !accelerator config work id_pattern --fail-safe

The Active integration line gates the post-draft push offer (Step 5). The accelerator config work integration command above prints an empty line when no integration is configured, so branch on the string: a non-empty value means integration configured (offer the push); an empty value means not configured (no push offer — behave exactly as today). This is the single resolution of the active tracker for this invocation; the push dispatcher is told which tracker to use from this value, so the gate and the route cannot diverge.

When ID pattern is {tracker}, the tracker owns every item's id: a pushed item takes the issue's key as its id, and an item no tracker has confirmed is saved as a draft in drafts/ under a provisional draft- ID, which /sync-work-items promotes later.

Work Item Template

The template below defines the sections and frontmatter fields that every work item must contain. Read it now — use it to guide what information you gather in Step 1 and what structure you produce in Steps 3–4.

!accelerator config template work-item --fail-safe

You are tasked with guiding the user through creating a well-formed work item — a structured document capturing a feature, bug, task, spike, or epic for tracking and implementation. This is a collaborative, challenging conversation: the model contributes its own knowledge and research alongside the user's input rather than simply transcribing what the user says. The goal is a work item that is well-reasoned and well-specified — one that would stand on its own without needing the author present to explain it.

The work item template has three sections that look similar but capture different things — populate each deliberately:

• Assumptions are interpretations you made that affect what gets built. Flag one only when using the wrong interpretation would lead someone to build something different. Example: "Treating 'users' as end users rather than internal staff — if wrong, scope changes."
• Open Questions are genuine unknowns a reader or implementer needs resolved before work can proceed. Example: "What does 'better results' mean — improved relevance, faster delivery, or both?"
• Drafting Notes capture interpretations you made while filling out the work item — business-context calls, scope decisions, or technical choices that someone should review if they turn out to be wrong. Actively populate this section. If you inferred who the stakeholders are, what a vague term means, where the scope boundary sits, or which technical approach is implied, write it down. Routine field selections (kind, priority, tags) don't need an entry unless the choice reflects a substantive scope or meaning interpretation a reviewer should be aware of.

Step 0: Parameter Check

When this command is invoked:

1. If an argument was provided:

   First, try to resolve the argument as a reference to an existing work item by invoking the resolver:

   accelerator work resolve <argument>

   The resolver respects work.id_pattern and accepts paths, full IDs (PROJ-0042), legacy bare numbers (0042), and short numbers (42, resolved against work.key when set).

   • Exit 0 (single match): the resolver echoes the absolute path on stdout. Continue to frontmatter validation below.
   • Exit 1 (input invalid — neither path, full ID, nor bare number): proceed to topic-string handling below; treat the argument as a topic.
   • Exit 2 (ambiguous match): the resolver lists every candidate with a source-category tag (legacy, project-prepended, pattern-shape, or a project code). Present the list to the user and ask them to disambiguate by re-running with a full ID (e.g. PROJ-0042) or a path.
   • Exit 3 (no match): print "No work item matching <argument> — interpreting as topic string. If that's wrong, abort and re-run with a different argument (or /list-work-items to find a valid reference)." and proceed to topic-string handling below.
   • Exit 6 (path outside the work directory): the argument names a real file that lies outside the managed work directory, so it is not a topic string. Stop, print the resolver's error, and offer to run /list-work-items.

   Frontmatter validation (only reached after a file was successfully resolved). Run:

   accelerator work show <path> --field id

   The own-identity field is id on unified-shape files and work_item_id on legacy files; work show resolves both transparently, so either key name passed here would return the same value.

   If it exits non-zero (frontmatter missing or unclosed), print:

   Could not parse frontmatter in <path> — the file may be corrupted.
   Re-open it and check that the YAML frontmatter is bracketed by two
   `---` lines and contains all nine required fields.

   and exit without spawning agents or writing any file. (The file clearly resolved, so the user intended a path — no fallback applies.)

   If validation passes, read the file fully (frontmatter and body) and cache the identity fields in conversation state: id (or work_item_id on legacy files), date, author, status, title, kind, priority, parent, tags. For any missing optional field, use the template default — do not abort. Set the conversation into enrich-existing mode with existing_work_item_path cached, and skip directly to Step 1 in that mode — do not run the vagueness check.

   Topic-string handling: if no resolution succeeded (discriminators did not trigger, or they triggered with a fallback warning), check whether the argument is too vague (no clear deliverable or subject).

   • Vague examples: "improve things", "fix the API", "add more features"
   • Clear examples: "add full-text search to the docs index page", "fix login timeout after password reset"
   • If vague, ask at least one clarifying question and wait for a more specific description before proceeding to Step 1.
   • If clear, proceed directly to Step 1.
2. If no argument was provided, respond with:

I'll help you create a well-specified work item through a collaborative conversation.

To get started, describe what you want to achieve — what's the problem or goal?
For example: "add full-text search to the docs index page", or "users can't log in after resetting their password"

I'll ask a few questions to understand the problem space, do some research, then
work with you to shape a thoroughly reasoned work item.

Then wait for the user's input. Apply the same vagueness check to their response — if still vague, ask a clarifying question before proceeding to Step 1.

Step 1: Gather Business Context

Once the topic is clear, ask 3–5 open questions to understand the business context before investigating. Consult the work item template sections above to understand what the work item will need — use them to guide which questions are worth asking. Tailor the questions to the topic, but cover:

• What pain point or problem does this address, and who experiences it?
• What is the desired outcome — what changes for people once this is done?
• Does the user have a solution or approach in mind? If so, surface it and ask whether it's a constraint or a preference to be explored further.
• Are there any constraints, deadlines, or dependencies worth knowing?
• For bug topics: what is the impact, and is it a blocker?
• Is there anything you are uncertain about, or that I should research?

Ask all relevant questions at once rather than one at a time. Wait for the user's answers before proceeding to Step 2.

In enrich-existing mode

Do not ask the broad discovery questions above. Instead:

1. Identify which body sections of the existing file are substantive (real content beyond [bracketed placeholder] text) and which are gaps (empty, placeholder-only, or missing entirely). Tag each gap with exactly one of the literal tokens:

   • empty — section is absent or contains no content
   • placeholder-only — section contains only template [...] blocks
   • instructional-prose — section contains the template's instructional prose carried over verbatim (e.g. "Describe the business value of this work item here…")
   • partial — section contains one or more substantive sentences alongside residual placeholders

   Use the literal token; do not paraphrase. The eval grader pattern-matches on these exact strings.

2. Present the gap analysis briefly:

   I've read the existing work item (<resolved path>). Here's what looks
   complete and what still needs work:
   
   Complete: [section list, or "none"]
   Gaps:
     - <Section> (<tag>)
     - ...
   
   I'll ask targeted questions about the gaps.

3. Ask only questions that address the identified gaps. Do not re-ask questions whose answers are already substantively present in the file.
4. If the existing content is rich enough that no obvious gaps remain, briefly confirm this and ask what the user wants to add or improve, then proceed to Step 2.

Step 2: Investigate

Using the topic and business context from Step 1, run investigation agents in parallel:

1. Spawn {documents locator agent} to search {work_dir} only — look for existing work items with similar titles, descriptions, or scope. Do not search research documents, plans, or other directories — work items capture business requirements, not implementation details.
2. Spawn {web-search-researcher agent} when there is uncertainty or a need for richer context about any aspect of the topic — business rules, domain concepts, competitive landscape, industry standards, external technology, or anything the model lacks confidence on. Skip this agent only when the topic is entirely self-contained and well-understood from the user's description. When in doubt, prefer to spawn research — over-asking is cheaper than producing a vague or poorly-grounded proposal.

Run both agents in parallel where both are warranted.

Once agents complete, synthesise findings — what the model knows from training, what research turned up, what prior work items exist — as the foundation for Step 3.

If a similar existing work item is found: Check its status.

• If the status is done, abandoned, or superseded — mention the work item briefly and note it is already closed, then continue creating the new one.

• Otherwise — surface the work item and offer the user numbered options, for example:

  1. Proceed with a new work item (if the scope genuinely differs)
  2. Exit and update the existing work item instead (use /update-work-item once available; for now, update it manually)
  3. Continue creating a new work item linked to the existing one as a parent

  Adapt the options to what makes sense given the work item's kind and status. Do not silently continue or modify the existing work item inline. Wait for the user's choice before proceeding.

In enrich-existing mode: exclude the resolved input file from the similarity scan. The {documents locator agent} search of {work_dir} will find the file being enriched; do not surface it as a "potential duplicate" of itself. Other near-duplicates discovered are handled per the unchanged rules above.

Step 3: Propose and Refine

Using the business context and investigation synthesis, the model leads with a structured proposal. Do not ask the user to generate requirements or acceptance criteria from scratch — propose them and invite challenge.

1. Recommend a work item kind with a brief rationale. Valid kinds come from the work item template's kind field (loaded at the top of this skill) — do not hardcode the list. Default to story when the kind is genuinely ambiguous.
2. Draft requirements drawn from the business context and research.
3. Draft acceptance criteria — specific and testable; prefer Given/When/Then format for story/task. Draw on domain knowledge and research to make these thorough, not just what the user explicitly mentioned.
4. Surface interpretations and gaps using the right section:
   • Put scope-changing interpretations you made in Assumptions and invite the user to confirm or correct them.
   • Put genuine unknowns that block progress in Open Questions.
   • Put every meaningful drafting-time interpretation (who stakeholders are, what vague terms mean, scope boundaries, implied technical approach) in Drafting Notes so a reviewer can challenge them.

Present as a structured proposal:

Based on what you've told me and my research, here's what I think this work item needs:

**Suggested kind**: [kind]
**Rationale**: [one sentence]

**Requirements I'd suggest**:
- [requirement]

**Out of scope** (explicitly not captured here):
- [item]

**Acceptance criteria I'd suggest**:
- Given/When/Then...

**Assumptions I've made** (scope-changing calls — confirm or correct):
- [assumption]

**Drafting Notes** (interpretations a reviewer should see):
- [interpretation — e.g. who stakeholders are, what a vague term means,
  scope boundary, implied technical approach]

**Open questions** (unknowns to resolve before work begins):
- [anything the user should clarify before drafting]

Wait for the user to validate, push back, or refine. Challenge vague or untestable responses — if an acceptance criterion is not measurable (e.g., "it works correctly"), first ask a clarifying question to understand the intent ("what would a passing test actually look like here?"). Only after understanding the intent should you help reformulate it into something testable. Do not accept vague criteria into the draft. Iterate until the proposal is well-specified and agreed.

In enrich-existing mode

Do not lead with a from-scratch proposal. Instead present a section-by-section review and augmentation:

Here's how the existing work item reads against my research, with proposed
additions for the gaps:

**[Section name]** — [complete | needs improvement: <reason> | missing]
[existing content excerpt or note that it is missing]
[proposed addition or replacement, when applicable]

[repeat per section]

**Title**: [keep / propose new title with rationale]
**Kind**: [keep existing <kind> / propose change to <kind> with rationale]
**Priority**: [keep / propose change with rationale]
**Parent**: [keep / propose change]
**Tags**: [keep / propose additions]
**Status**: [keep <cached> — say so if you'd like to transition it]

Apply the canonical Identity Field Rules (see Quality Guidelines): propose only the fields marked Proposable; never propose changes to immutable fields; surface a status transition only if the user makes a direct explicit request for one (the listing offers the affordance but never proposes a change unsolicited).

The refinement loop in this step (challenging untestable criteria, vague requirements, etc.) applies equally to existing and proposed content.

Step 4: Draft Work Item

1. Draft a complete work item from the agreed proposal using the template structure loaded at the top of this skill. Use NNNN as the placeholder work item number throughout. Do NOT call accelerator work next-number at this step.
2. Kind-specific content placement:
   • story/epic: open the Summary section with a user story statement — "As a [role], I want [goal], so that [benefit]." — before the descriptive sentences
   • bug reproduction steps, expected/actual behaviour → Requirements section
   • spike research questions, time-box, exit criteria → Requirements section
   • epic initial stories → Requirements section as a list
   • Do not rename or add sections beyond those defined in the work item template
3. Populate the References section with any external material that informed the work item — research artefacts surfaced by the web researcher, related work items found in Step 2, design docs, or source specs. Leave it empty if nothing external was consulted; do not invent references.
4. Populate the Drafting Notes section with the interpretations carried over from the agreed proposal. If there were none (the user confirmed every call), the section may be omitted or left as an empty list — but do not leave placeholder bracketed text in it.
5. Present the full draft to the user:

Here's my draft work item. Please review and let me know if you'd like any
changes before I write it to disk:

[draft content]

6. Continue to challenge during the review loop — flag untestable criteria, vague requirements, or gaps surfaced by research that remain unaddressed. Iterate until the user explicitly approves. accelerator work next-number is never called during this loop.

In enrich-existing mode

1. Produce a complete updated draft incorporating the existing content plus the additions approved in Step 3.
2. Apply the canonical Identity Field Rules (see Quality Guidelines) when filling frontmatter — the rules are the single source of truth for which fields are immutable, preserved, or proposable.
3. Apply the H1 sync rule (see Quality Guidelines).
4. Apply the Script avoidance rule (see Quality Guidelines).

5. Present the full updated draft, framed as an update preview:

   Here's the updated work item. The Step 5 confirmation will name the file
   path and ask for explicit approval before any write:
   
   [draft content]

6. Continue to challenge during review — apply the same rules as the normal-path Step 4. Iterate the draft until the user is happy with it. Do not treat a "looks good" mid-iteration as approval to write — that approval is gated by Step 5's single confirmation.

Step 5: Write Work Item

A re-key in progress. accelerator work next-number and accelerator work create (with or without --push) refuse before reading or writing anything while an accelerator work rekey run is unfinished. Before any other branch on a non-zero exit from one of them, check its stderr for these codes; in every case nothing was written:

• E_REKEY_IN_PROGRESS (exit 4) — tell the user a re-key is unfinished: run /rekey-work-items to finish or undo it. Stop.
• E_REKEY_PLAN_UNREADABLE (exit 4) — relay the printed remedy as is, never suggesting the re-key be finished. Stop.
• E_IDENTITY_RUN_BUSY (exit 1) — another identity command is still running; retry once it ends.

Under {tracker}, skip steps 1–3: there is no projected ID. The tracker assigns the key on a push, and accelerator work create mints the draft ID itself, so its first stdout line is the only path to report.

1. Call accelerator work next-number to get the next full ID under the configured pattern:

accelerator work next-number

The output is the full ID (0001 under default {number:04d}, PROJ-0001 under {key}-{number:04d} with work.key: "PROJ"). The allocator reads the pattern and local ID prefix from configuration; pass --project CODE if the user explicitly wants a non-default prefix for this one call.

If the command exits non-zero (e.g., overflow, missing project value), abort immediately and surface the error message verbatim — do not proceed.

2. Resolve the target path: {work_dir}/<full-id>-kebab-slug.md where <full-id> is the allocator's output and the slug is a meaningful kebab-case summary of the title (not raw input text).
3. Check that the target path does not already exist. If it does, abort:

Path {path} already exists — another session may have written a work item
concurrently. Please re-run /create-work-item.

4. Create the work items directory if it does not exist.
5. Resolve author: configuration if present, then the current git/jj user identity, then — only if both fail — ask the user once. Never pass a placeholder. This value feeds both branches below.

6. No integration configured — write via accelerator work create. Skip this step (go to step 7) when an integration is configured; the push state machine in step 8 governs that case's single write instead.

   1. Write the drafted body (H1 through References, everything approved across Steps 1–4) to a scratch file, with the NNNN placeholder still literally present everywhere it appears (the H1 line and anywhere else) — do not substitute it yourself. work create performs the one real NNNN → allocated-ID substitution as part of the same atomic write that decides the ID, so the approved content can never end up with an id/H1 mismatch.
   2. Call the dispatcher directly:

   accelerator work create "<title>" <kind> <priority> \
     --status draft \
     --author "<resolved author>" \
     --producer create-work-item \
     --body-file <scratch-file> \
     [--parent "work-item:NNNN"] \
     [--tag <value>]... [--block "work-item:NNNN"]... \
     [--blocked-by "work-item:NNNN"]... \
     [--derived-from "plan:NNNN"]... [--relates-to <ref>]... \
     [--source "issue-research:NNNN"] [--project <code>]

   Include each bracketed flag only when that field has a value — work create itself omits the corresponding frontmatter key when the flag is absent, matching the template's own omit-when-empty convention, so no separate omission logic is needed here. --tag/ --block/--blocked-by/--derived-from/--relates-to are each repeatable for multiple values. external_id has no flag — work create never writes it (nothing to sync yet); this matches today's behaviour of omitting it from a fresh draft. 3. If the command exits non-zero, surface its stderr message verbatim and abort — do not retry and do not attempt a manual Write. Otherwise its stdout is the path of the written file. Skip to step 8.

7. Push state machine (only when Active integration is non-empty; otherwise this whole step is skipped and the file was already written in step 6). The drafted frontmatter is held in memory — no file exists yet.

   1. Offer the push using the fail-safe gate. Preview the fields the push would resolve through the dispatcher's own dry-run so the single keystroke maps to an understood, non-destructive result. Write the drafted body to a temp file and run:

      accelerator work create "<title>" <kind> <priority> \
        --push --dry-run --body-file <body-tmp>

      It resolves the tracker from Active integration, contacts it to resolve the fields, writes nothing, and creates no remote issue. Branch on its exit and the line it prints on stdout:

      • Exit 0, jira → jira\t<issue type>\t<type source>\t<project>\t<project source>. Render the type and project with their sources. A <project source> of unresolvable means the configured jira.project_key names a project the tracker does not hold — surface that as a pre-create failure before the gate (name jira.project_key) and do not offer the push; save locally instead, through the step 6 call (no --push). A default source is a benign fallback, not a failure.
      • Exit 0, linear → linear\t(no user-resolvable type/project fields) (the team and issue-type catalogue are fixed by /init-linear). Offer the push.
      • Exit 70 → the tracker was unreachable (distinct from an unresolvable value). Do not hard-block: tell the user the preview could not be resolved and offer to save locally now, or retry the preview. The push itself carries its own retry (below), so a transient outage here must not strand the draft.
      • Any other non-zero exit → the tracker is not usable for a push (not available, unrecognised, or unconfigured). Relay the stderr message and save locally.

      Every "save locally" above is the step 6 accelerator work create call with no --push, so under {tracker} it writes a draft that /sync-work-items promotes later. Never write the file with the Write tool, and never call accelerator work next-number for it.

      Then, when a push is still on the table:

      Use the AskUserQuestion tool with two options:

      1. Yes, push to [tracker] now — create the remote issue and save locally
      2. No, save locally only — save as unsynced; push it later with /sync-work-items (it offers to create unsynced local drafts on the remote), which shares this external_id contract.

      Under {tracker} the second option reads No, save as draft — save it as a draft in drafts/; /sync-work-items promotes it onto a tracker key later.

   2. On decline → write the file now with external_id omitted (unsynced) via the accelerator work create call from step 6 (no --push), then go to step 8. Under {tracker} that call writes a draft.

   3. On accept → run the same accelerator work create call with --push added (never a re-derived tracker — the CLI reads Active integration itself):

      accelerator work create "<title>" <kind> <priority> \
        --status draft --author "<resolved author>" --producer create-work-item \
        --push --body-file <body-tmp> [--parent …] [--tag …]… [linkage flags…]

      This is one call that creates the remote issue and writes the local file atomically — the retryable/terminal decision, the retry, and the duplicate-create guard all live inside work create --push; do not drive a retry loop yourself. On stdout it prints the written path on the first line and <keyword>\t<external_id> on the second, where the keyword is one of write-once, local-save, loud-terminal, rejected, created-blocked, created-unwritten or retirement-incomplete. The first line is empty only for a created-unwritten under a numeric ID pattern, where no file exists. Render per the table below.

      Under {tracker} the call first saves the item as a draft, then creates the issue and promotes the draft onto its key, so the written path is <KEY>-<slug>.md on success and the draft's path otherwise. If it exits 4 with E_PUSH_PENDING or E_DRAFT_EXISTS on stderr, an earlier create of the same content is still pending and nothing was sent: relay the named draft and the recovery the message gives, and do not re-run the create.

   Outcome table (work create --push returns the keyword; you render the message). The 70-retryable / 71-terminal / 75-rejected contract the keywords derive from is documented in exit_codes and cross-referenced by work create --help:

   Second-line keyword	Exit	Action
   write-once	0	The remote issue was created and the file written with its external_id. Confirm success (step 8), echoing the returned identifier.
   local-save	0	The push could not be sent (the tracker is unreachable, retryable and retried once, or not available/recognised/configured) and nothing left the machine — the file was written unsynced. Tell the user it saved unsynced and can be pushed later via /sync-work-items; but if stderr carries a warning: -: configure failed: line, relay its cause; the draft cannot be pushed or promoted until it is fixed, and sync will refuse until then.
   loud-terminal	71	A terminal failure at or after the create — a remote issue may already exist. Print loud non-idempotent guidance naming the saved path: do not blindly re-run /create-work-item. Under a numeric ID pattern the file was saved unsynced: check the tracker, and if the issue exists reconcile via /sync-work-items or set external_id: <KEY> by hand. Under {tracker} the item is a draft whose ID is the draft- prefix of its filename: check the tracker, then run accelerator work promote <draft-id> --adopt <KEY> if the issue exists, otherwise accelerator work promote <draft-id> --create; never set external_id by hand on a draft.

| rejected | 75 | The client refused the request as invalid before sending it, so no remote issue exists, but re-sending the same request would be refused again. The file was saved unsynced (a draft under {tracker}). Relay the cause printed on stderr and tell the user the request must change (for example, the body) before it is pushed via /sync-work-items. | | created-blocked | 4 | The remote issue exists, but a local item already claims its key. The draft is kept. Relay the holder and the remedy printed on stderr; the user resolves the collision, then runs accelerator work promote <draft>. | | created-unwritten | 71 | The remote issue exists and no local item carries it yet. The next /sync-work-items links it (under {tracker}, by promoting the draft onto it). Do not re-run the create. | | retirement-incomplete | 71 | The remote issue exists, and moving the draft onto its key could not restore every file it had written. Relay the paths and the recovery directory printed on stderr; the user restores them, then runs /sync-work-items. | | E_PUSH_PENDING (stderr) | 4 | An earlier create of the same content is still pending. Relay the named draft and the stage-specific recovery the message gives. | | E_DRAFT_EXISTS (stderr) | 4 | The same draft already exists. Tell the user to run accelerator work promote <draft> or /sync-work-items. |

Under {tracker}, local-save and rejected leave the item as a draft, and loud-terminal leaves the draft with a record that a create may have reached the tracker.

Because the create and the local write are one atomic CLI operation, there is no in-skill window where a remote issue exists but no file does — that window, and its pending_push recovery marker, are handled inside the CLI. The pre-create preview is a live, credentialed round-trip now, so a clean preview does not guarantee the push succeeds; the authoritative outcome is the keyword above.

8. Print a confirmation:

Work item created: `<path>`

<path> is always the first stdout line of the accelerator work create call that wrote the file — whether step 6 (no integration) or step 7 (--push) — since its slugification is authoritative; do not re-derive it or assume it matches step 2's preview path. When the item was saved unsynced (decline, local-save, or loud-terminal), say so: Work item created (unsynced): … — push later with /sync-work-items, or, when a warning: -: configure failed: line came with a local-save, Work item created (unsynced): … — fix <cause> before syncing. Under {tracker}, report the tracker key on write-once (Work item created: <KEY> at <path>), and otherwise the draft path and why it is still a draft (Work item saved as draft: <path> — <reason>; /sync-work-items promotes it), where the reason is "push declined", "tracker unreachable", or the outcome table's cause; with that warning it ends — fix <cause> before syncing instead.

In enrich-existing mode

1. Do not call accelerator work next-number. The target path is the resolved existing_work_item_path cached in Step 0.

2. At-write identity-swap check (best-effort guard): immediately before the confirmation prompt, re-read the target file's own identity from disk via:

   accelerator work show <existing_work_item_path> --field id

   (This returns the value of id: on unified-shape files and falls back to work_item_id: on legacy files.)

   • If the command exits non-zero (file gone or frontmatter unparseable since Step 0), abort with: "Error: <path> is no longer present or its frontmatter is unparseable. Your proposed draft is below — copy it before re-running /create-work-item: <draft>"
   • If the on-disk identity differs from the value cached in Step 0, abort with: "Error: <path> changed identity since Step 0 (was <cached>, now <current>). Your proposed draft is below — copy it before re-running /create-work-item to refresh: <draft>"

3. Single confirmation gate: present the path and a per-section change summary, then require explicit y/n approval before any write:

   I'm about to overwrite <existing_work_item_path>.
   
   Sections changed:
     Added:    [...]
     Modified: [...]
   
   Frontmatter changed:
     <field>: <cached> → <new>     # repeated per modified field
     (status: <cached> unchanged)  # always show status explicitly
   
   Sections preserved verbatim: <count> (<terse list or "none">)
   Frontmatter fields preserved verbatim: id, date, author
     [+ any unmodified proposable fields]
   

   Then use the AskUserQuestion tool with two options:

   1. Yes, proceed — overwrite the work item with the changes shown
   2. No, cancel — make no changes
4. Confirmation interpretation (fail-safe):
   • Exactly y or Y (after trimming whitespace): proceed to step 5.
   • Exactly n or N: go to step 8.
   • Anything else (empty input, "yes", "go ahead", "looks good but also...", paragraph of feedback): treat as n. Print: "Did not recognise <response> as y/n — staying in review. What change would you like before I overwrite?" and go to step 8.
5. On y — substitute cached immutable fields, then write: immediately before invoking the Write tool, re-read the cached immutable identity fields (id, date, author) from conversation state and overwrite the corresponding frontmatter lines in the draft text with those exact values, even if they appear unchanged. Also write last_updated: to the current Current Date/Time (UTC): value and last_updated_by: to the resolved author. This textual substitution defends against drift during Step 4 iteration.
6. Write the file to existing_work_item_path. The path-existence guard from the normal flow does not apply — overwrite is the intended behaviour.

7. Print:

   Work item updated: <existing_work_item_path>

8. On n or unrecognised: stay in Step 4 / Step 5 review and iterate. Do not re-run the identity-swap check until the next y; do not write.

Validate the frontmatter: after the work item is written, run

accelerator corpus frontmatter validate --file <written work item path>

If it exits non-zero, the document violates the canonical frontmatter standard; report the emitted violation and fix the frontmatter before completing.

Quality Guidelines

• Never write a file without explicit user approval.
• Never call accelerator work next-number before the user approves the draft.
• The slug must be a meaningful kebab-case title, not raw input text.
• Work item kind must come from the work item template's kind field (loaded at the top of this skill), not a hardcoded list. Default to story when the kind is genuinely ambiguous.
• All frontmatter fields defined in the work item template must be populated in every written work item: type (work-item), id matching the assigned full ID (a quoted YAML string — see contract below), title matching the user-approved title, date, author, producer (create-work-item), status (draft), kind, priority (medium unless the user specified otherwise), parent (empty string unless a parent was established), external_id (empty unless set), tags (a YAML array, possibly empty), last_updated, last_updated_by, and schema_version: 1. No field may contain unfilled placeholder text like [author] or NNNN. The body H1 format is # <full-id>: <title> where <full-id> is whatever the configured work.id_pattern produces — kept in sync with the frontmatter title: field.
• id frontmatter type contract: the own-identity field is always a quoted YAML string, regardless of the configured pattern. Files created under default {number:04d} write id: "0001"; files created under {project}-{number:04d} write id: "PROJ-0001". This contract is uniform so consumers can read the field as a string without coercion. See skills/config/configure/SKILL.md > work for the full contract. Legacy files carry the same value under work_item_id:; work show returns either key transparently so consumers do not need to know which shape is on disk.

Identity Field Rules (apply in enrich-existing mode):

• Immutable — id (or work_item_id on legacy files), date, author. Cached from the source file in Step 0; never proposed for change; the model substitutes the cached values back into the draft frontmatter at write time (Step 5 step 5) as a defence against drift during Step 4 iteration. This is a textual substitution the model performs — the eval suite verifies the written file's values match the cached values, which is the strongest guarantee the grader- mediated harness can provide.
• Preserved unless explicitly changed — status. Defaults to the cached value. May only change if the user makes an explicit, direct request during the conversation (e.g. "set status to in-progress"). The model must not propose a status change unsolicited; if the user makes only an oblique reference (e.g. "this is now in flight"), the model asks a clarifying question rather than infer the transition. A proposed transition is shown explicitly (e.g. "draft → in-progress") and requires confirmation in Step 3 before acceptance. The status: draft default applies only to newly-created work items, not to enrichment.
• Proposable in Step 3 — title, kind, priority, parent, tags. Default to the cached values; can be replaced after explicit user agreement in Step 3's augmentation review.

H1 sync (enrich-existing mode): the body H1 is # <id>: <title>, using the cached immutable id (the full ID produced by the configured pattern, never a placeholder; sourced from work_item_id on legacy files) and the title as confirmed or replaced in Step 3.

Script avoidance (enrich-existing mode): accelerator work next-number is never called. The number is already cached. The path-existence guard in Step 5 does not apply — overwrite is the intended behaviour once the at-write identity-swap check passes.

• date must use the work item template's YYYY-MM-DDTHH:MM:SS+00:00 format in UTC (e.g. obtained via date -u +%Y-%m-%dT%H:%M:%S+00:00).
• author is sourced in this order: configuration if present, then the current git/jj user identity, then — only if both fail — ask the user once before writing the file. Never write [author] or any placeholder.
• Summary and Requirements must have substantive content — no [bracketed placeholder text] in the final written file. The same applies to every other body section that is populated: if a section would only contain a bracketed placeholder, remove the placeholder and leave the section empty (or omit optional sections entirely).
• Acceptance criteria must be specific and testable; prefer Given/When/Then format. Challenge any criterion that is not measurable before accepting it.
• Populate Drafting Notes with every meaningful interpretation made while filling out the work item — who stakeholders are, what a vague term means, where scope boundaries sit, which technical approach is implied. Use Assumptions for scope-changing calls the user should confirm. Use Open Questions for genuine unknowns that block progress. These three sections serve different purposes; don't collapse them.
• Populate References with any external material that informed the work item — research artefacts, related work items, design docs, source specs. Leave it empty if nothing external was consulted; do not invent references.
• The model must contribute its own knowledge and research to the proposal — not simply transcribe the user's answers. Bring domain expertise, surfaced research, and reasoned suggestions to every work item.
• Restrict {documents locator agent} to {work_dir} only. Do not search research documents, plans, or the codebase — work items are about business requirements, not implementation details.
• Spawn {web-search-researcher agent} whenever there is uncertainty about any aspect of the topic — business rules, domain concepts, competitive landscape, industry standards, external technology, or otherwise. When in doubt, prefer to spawn research — over-asking is cheaper than producing a vague or poorly-grounded proposal.
• If accelerator work next-number exits non-zero, abort and surface the error message verbatim.

!accelerator config instructions create-work-item --fail-safe