| name | refine-work-item |
| description | Interactively refine a work item by decomposing it into children, enriching it with codebase context, sharpening its acceptance criteria, sizing it, or linking it to dependencies. Use after a work item has been drafted and before planning begins. |
| argument-hint | [work item number or path] |
| allowed-tools | ["Bash(${CLAUDE_PLUGIN_ROOT}/scripts/config-*)","Bash(${CLAUDE_PLUGIN_ROOT}/skills/work/scripts/*)"] |
Refine Work Item
!${CLAUDE_PLUGIN_ROOT}/scripts/config-read-context.sh
!${CLAUDE_PLUGIN_ROOT}/scripts/config-read-skill-context.sh refine-work-item
!${CLAUDE_PLUGIN_ROOT}/scripts/config-read-agents.sh
If no "Agent Names" section appears above, use these defaults:
accelerator:reviewer, accelerator:codebase-locator,
accelerator:codebase-analyser, accelerator:codebase-pattern-finder,
accelerator:documents-locator, accelerator:documents-analyser,
accelerator:web-search-researcher.
Work items directory: !${CLAUDE_PLUGIN_ROOT}/scripts/config-read-path.sh work
Work Item Template
The template below defines the sections and frontmatter fields that every
work item must contain. Read it now — use it to know valid kinds, statuses,
priorities, and section names without re-reading the file at runtime.
!${CLAUDE_PLUGIN_ROOT}/scripts/config-read-template.sh work-item
You are tasked with refining a work item through one or more of five operations:
decompose it into child work items, enrich it with codebase context, sharpen
vague acceptance criteria, add a t-shirt size indicator, or populate its
dependencies. Every operation is interactive and targeted — you propose, the
user approves, and the Edit tool makes the minimum change needed.
Step 0 — Parameter Check
If no work item path or number was provided, respond with:
I'll help you refine a work item. Please provide the path or work item number.
Example: `/refine-work-item {work_dir}/0042-user-auth.md`
Or by number: `/refine-work-item 42`
Run `/list-work-items` to see available work items.
Then wait for the user's input.
If a path or number was provided: proceed to Step 1.
Accepted forms:
- A path (e.g.
meta/work/0042-user-auth.md)
- A bare work item number (e.g.
0042 or 42, resolved against {work_dir})
If the resolved path does not exist, report:
No work item file at <path> — run `/list-work-items` to see available work items.
and exit without reading any other file, spawning any agent, or writing anything.
If the YAML frontmatter cannot be parsed (missing closing --- or syntax
error), report:
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, or run `/update-work-item <path>` which
surfaces the same diagnostic with field-level detail.
and exit without editing the file or spawning agents.
Step 1 — Read Target and Context
Read the target work item fully. If the parent field is non-empty, also read
the parent work item. The work item template (loaded above) tells you the valid
kinds, statuses, and priorities — do not re-read it at runtime.
Step 2 — Analyse (mandatory parallel agents)
Spawn BOTH agents in the same tool-use turn (parallel, not sequential):
- {codebase locator agent}: find files relevant to the work item's
Requirements and Summary
- {codebase analyser agent}: analyse how the relevant components
currently work
Wait for both before presenting the menu. Even if the work item seems
straightforward, always spawn these agents — the menu previews depend
on their findings.
Step 3 — Present Refinement Menu
Present the five operations with one-line descriptions. Reference agent
findings in the previews so the user can pick informed:
- decompose — split into child work items (epic→stories, story→tasks);
e.g. "Requirements suggest 3 child stories (R1, R2, R3)"
- enrich — add Technical Notes from codebase analysis;
e.g. "4 relevant files identified (src/auth/session.ts, …)"
- sharpen — tighten vague acceptance criteria;
e.g. "2 vague criteria detected (AC3, AC5)" or "all criteria already testable"
- size — append a t-shirt size indicator with rationale;
e.g. "estimate M based on files in auth/ and session/"
- link — populate Dependencies from related work items;
e.g. "1 potential blocker found (0031)" or "no related work items found"
Operations that have nothing to do should be marked as such (e.g.
"sharpen — all criteria already testable").
User can select one, multiple, or "all relevant". Regardless of selection
order, always execute in canonical order: decompose → enrich → sharpen
→ size → link. This ensures Technical Notes content is in place before
size's **Size**: line is prepended, and decompose's children exist before
any link operation references them.
Step 4 — Execute Operations
4a. Decompose
Propose 2–5 candidate children (2–4 for story decomposing to tasks) with
draft titles and one-line Summaries derived from the Requirements section.
Bug/spike challenge: if the work item kind is bug or spike, first use
the AskUserQuestion tool with two options:
- Yes, proceed anyway — continue with decompose
- No, cancel — exit decompose and return to the menu
Existing children: if the Requirements section already contains a
### Child work items subsection, offer:
append (add new children to the existing list) / skip (do not decompose further) / cancel
Never replace the existing list silently. For the append path, use the
anchor described below under "append to existing ### Child work items".
Approval grammar: each proposal MUST include a one-line grammar legend
immediately under the numbered child list:
Commands: approve all | edit N: <title> | drop N | add: <title> | regenerate | cancel
The legend appears on every proposal turn — the first proposal and after
every grammar iteration.
Process user input:
approve all, yes, lgtm → proceed to the pre-write warning
edit N: <new title> → update child N's title, re-show updated proposal
with legend restated. Do NOT write.
drop N → remove child N, renumber remaining children 1…M, re-show
with legend. Do NOT write.
add: <title> → append a new child with that title, re-show with legend.
Do NOT write.
regenerate → discard current proposal, generate a fresh set of 2–5
candidates from the same Requirements. Do NOT write.
cancel, abort → print "decompose cancelled — no children written"
and return to the menu. No numbers allocated.
- Any other input → print "unrecognised command", restate the legend, and
re-show the unchanged proposal unchanged. Do NOT treat as approval.
Pre-write warning: before writing, use the AskUserQuestion tool with two
options. State the allocation count and warn about partial state:
- Yes, proceed — allocate N numbers and write N files (partial state is
possible if aborted mid-write; use
jj restore <file> to discard)
- No, cancel — cancel without writing
On approval:
-
Call work-item-next-number.sh --count N exactly once to allocate N
consecutive numbers.
-
For each child, write NNNN-kebab-slug.md, populating every unified
frontmatter field as specified under Populate frontmatter below.
Immediately before writing each child, verify the computed filename does
not already exist. If it does, abort with:
Collision: <path> already exists (concurrent session?). Aborting.
Allocated: NNNN, NNNN, …; Wrote: N-1 files (list them).
Use `jj restore <file>` to discard any children written.
Child body includes: Summary (from proposal), Context (linking to
parent with "Child of NNNN — "), Requirements (minimal
but substantive; no [bracketed placeholder] text), Acceptance Criteria
(minimal but substantive; sharpen can tighten these later), Dependencies
(blank), and remaining template sections.
-
Append ### Child work items to the parent's Requirements section:
After all children are written successfully:
- Read the parent file
- Locate
\n## Requirements\n and \n## Acceptance Criteria\n
- Extract the last non-empty line of Requirements (the line immediately
before the blank-line transition to Acceptance Criteria); call this
<req_tail>
- Build:
old_string = <req_tail>\n\n## Acceptance Criteria\n
- Build:
new_string = <req_tail>\n\n### Child work items\n\n- NNNN — title\n…\n\n## Acceptance Criteria\n
- Edge case — empty Requirements: use
## Requirements\n\n## Acceptance Criteria\n
as old_string and insert ### Child work items between them.
- Pre-Edit, count occurrences of
old_string in the parent file. If not
exactly 1, abort the parent update with:
Could not locate a unique '## Acceptance Criteria' anchor in <path>
(matches found: <N>). Parent not updated.
Children NNNN, NNNN, NNNN remain on disk; add their links manually
or run `jj restore <parent-path>` and re-run decompose.
Children already written remain on disk.
Append to existing ### Child work items (re-decompose path):
- Locate
### Child work items subsection, find its last - NNNN — title
line; call this <last_link>
- Build:
old_string = <last_link>\n\n## Acceptance Criteria\n
- Build:
new_string = <last_link>\n- NNNN — title\n…\n\n## Acceptance Criteria\n
- Same uniqueness pre-check; same abort diagnostic if not exactly 1.
-
Print a completion ledger, one line per child:
Wrote NNNN — title
Wrote NNNN — title
Allocated 3 numbers, wrote 3 files.
On aborted or partial write, show allocated vs. written counts and
list any written filenames.
After a successful decompose, proceed to Step 5 (hierarchy display) before
running any remaining operations.
Populate frontmatter
When writing each child (step 2 above), substitute every field below
with the indicated value:
type: ← work-item
id: ← from the script, zero-padded four-digit string
(own-identity; legacy files use work_item_id)
title: ← per-child proposal title; body H1 matches exactly
date: ← current UTC timestamp via date -u +%Y-%m-%dT%H:%M:%S+00:00
author: ← first match in chain: parent work item's author field → configured
author value (from context config) → jj config get user.name →
git config user.name → ask the user once and apply to all children
producer: ← refine-work-item
kind: ← derived: epic → story, story → task, bug/spike → ask
user to confirm before proceeding (already done in the challenge step),
any other kind → story with a one-line notice
status: ← literal draft
priority: ← inherit from parent; if parent has none, ask once and
apply to every child written in this session
tags: ← verbatim copy of the parent's tags array (empty array []
if the parent has none)
last_updated: ← same UTC timestamp as date
last_updated_by: ← same author value resolved above
schema_version: ← 1
Optional linkage/foreign-ref keys are omit-by-default: the template shows
each as ""/[], but write a key into the artifact only when it has a
value, and omit it entirely otherwise (do not carry the empty placeholder
through).
parent: ← the parent work item's ID as a typed-linkage ref "work-item:NNNN",
where NNNN is the parent's id field — from id on unified files,
or work_item_id on legacy files — canonicalised to the configured
pattern's full-ID shape, e.g. "1" → "work-item:0001". Always
fill — every decomposed child has a parent.
blocks: ← list of typed-linkage refs to work items this child
blocks (["work-item:NNNN", ...]). Fill when blocking edges are
explicit; otherwise omit the key.
blocked_by: ← list of typed-linkage refs to work items that block
this child. Prefer writing the canonical blocks: on the other
side; emit blocked_by: only when the canonical side cannot be
written, and omit it otherwise.
derived_from: ← list of typed-linkage refs to artifacts this child
is derived from (["plan:NNNN", ...]). Fill when derivation is
explicit; otherwise omit the key.
relates_to: ← list of typed-linkage refs to related artifacts.
Fill when relationships are explicit; otherwise omit the key.
source: ← typed-linkage ref to the originating source artifact
("issue-research:NNNN"). Fill when the source is explicit;
otherwise omit the key.
external_id: ← cross-system pointer (e.g. a Jira/Linear key). Fill
when propagated from the parent or linked to an external tracker;
otherwise omit the key.
4b. Enrich
Read the target work item's existing Technical Notes content.
If the codebase agents returned nothing concrete (no specific files or
components identified), report:
no enrichment could be grounded in code — skipping enrich
and make no Edit.
Otherwise propose Technical Notes content with specific path:line
references drawn from agent results. Do not invent references.
If non-trivial Technical Notes content already exists (anything beyond a
**Size**: line), ask:
replace (deletes existing Technical Notes) / append (add after existing content) / skip?
replace → show a unified diff (old struck, new added) and require an
explicit second y/n confirmation before invoking Edit
append → add new content after existing content. Preserve any leading
**Size**: line placed by a prior size operation — never overwrite it
skip → make no Edit
On approval via Edit: modify only the Technical Notes section. Do not touch
Requirements, Acceptance Criteria, Summary, or any frontmatter field.
4c. Sharpen
Read the target work item's Acceptance Criteria. Identify criteria that are
vague or untestable (non-measurable phrases like "should be fast", "handles
errors gracefully", "works correctly").
If every criterion is already specific and testable, report:
all acceptance criteria already testable — nothing to sharpen
and make no Edit.
Otherwise, for each vague criterion propose a specific, measurable rewrite
(e.g. "p95 latency under 200ms under default benchmark dataset"). Skip
criteria that are already testable — only propose rewrites for vague ones.
Iterate with the user until each proposed rewrite is agreed.
On approval via Edit: modify only the Acceptance Criteria section with the
agreed rewrites. Preserve criteria that were not sharpened byte-for-byte.
4d. Size
Read the target work item's Technical Notes. Check whether a **Size**: line
already exists as the first line.
Propose a t-shirt size (XS, S, M, L, XL) with a rationale
referencing specific files or subsystems from agent results. Iterate with
the user.
On approval:
- No existing
**Size**: line: insert **Size**: <value> — <rationale>
as the FIRST line of Technical Notes, followed by a blank line separating
it from any existing content. Do NOT add as a frontmatter key.
- Existing
**Size**: line, proposed value AND rationale match byte-for-byte
(ignoring leading/trailing whitespace): report "size unchanged — "
and make no Edit.
- Existing
**Size**: line with a different value or rationale: show a
unified diff of the proposed change (existing line struck, new line added)
then use the AskUserQuestion tool with two options:
- Yes, apply size change — replace the line in place
- No, cancel — make no Edit
4e. Link
Count NNNN-*.md files in {work_dir} using a single Glob invocation.
- Count ≤ 30: read them directly via batched Read
- Count > 30: spawn {documents locator agent} scoped to
{work_dir}
Propose Blocked by: and/or Blocks: entries in Dependencies, referencing
only real work item numbers (verify each proposed number exists before
including it). If no related work items are found, print:
no related work items found — link skipped
and make no Edit.
If Dependencies already has non-empty content, ask:
replace (overwrites existing entries) / append (add new entries after existing) / skip?
replace → show a unified diff and require a second y/n confirmation
append → add only net-new entries, skipping any that duplicate existing
entries
skip → make no Edit
On approval via Edit: modify only the Dependencies section.
Step 5 — Display Hierarchy
Run immediately after decompose writes at least one child (skip if decompose
was cancelled, declined by user, or the parent Edit failed).
Render the parent → children tree using the format pinned in /list-work-items:
Unicode box-drawing characters, two-space indent per depth level, ├── for
all children except the last, └── for the last child. The canonical fence
below MUST appear verbatim in this step's prose so
scripts/test-hierarchy-format.sh can verify byte-equality with the matching
fence in list-work-items/SKILL.md:
NNNN — parent title (kind: , status: )
├── NNNN — child 1 title (kind: , status: )
├── NNNN — child 2 title (kind: , status: )
└── NNNN — last child title (kind: , status: )
Concrete work item IDs and titles replace the placeholders in the actual
output (e.g. 0042 — User Auth Rework (kind: epic)). The status field
of the parent is omitted only if blank; all present fields are shown.
Step 5 runs inline after decompose completes, before enrich, sharpen, size,
and link operate on the parent. The hierarchy is not re-rendered after
subsequent operations.
Step 6 — Offer Review
After the entire selected operation set completes, use the AskUserQuestion
tool with two options:
- Yes, run review — run
/review-work-item on this work item now (and
each child if decompose was in the selection)
- No, done — exit without running review
Do NOT invoke /review-work-item automatically — wait for the user's choice.
Important Guidelines
- Mandatory agents at Step 2: always spawn both codebase agents in
parallel before presenting the menu, even for simple work items.
- Canonical operation order: decompose → enrich → sharpen → size → link,
regardless of the order the user selected them.
- Each operation owns its sections only:
- decompose: writes new child files and appends
### Child work items to
the parent's Requirements (never touching existing Requirements prose)
- enrich: owns Technical Notes prose content
- sharpen: owns Acceptance Criteria
- size: owns the single
**Size**: <value> — <rationale> line, always as
the FIRST line of Technical Notes; replace in place on re-run
- link: owns Dependencies
- Never modify any frontmatter field of the target work item
(
type, id (or work_item_id on legacy files), title, date,
author, producer, kind, status, priority, parent,
external_id, tags, last_updated, last_updated_by,
schema_version) — those transitions are /update-work-item's
concern. Children of decompose are new work items getting their
initial frontmatter.
- Destructive paths require two-step confirmation: replace mode for
enrich, replace mode for link, clobbering an existing
**Size**: line —
each must show a unified diff and require a second y/n confirmation.
- Edit failure path: if an Edit's target string cannot be matched (file
changed between read and edit), abort that specific edit with a clear
diagnostic and continue with remaining approved edits. For decompose: write
all children first, update the parent last; if the parent Edit fails,
children stay on disk and work item numbers are printed for manual linking.
- Numbering is consumed eagerly on decompose with no rollback; the skill
assumes a single-session invocation.
- Canonicalise parent field to zero-padded four-digit string (same as
/update-work-item), e.g. "1" → "0001".
- Idempotent on re-run: every operation checks existing content before
proposing additions.
Relationship to Other Commands
/create-work-item or /extract-work-items — create the work item
/refine-work-item — decompose and enrich (this command)
/review-work-item — automated multi-lens quality review
/stress-test-work-item — interactive adversarial examination
/update-work-item — status/metadata transitions (not this skill's concern)
/create-plan — plan implementation from an approved work item
!${CLAUDE_PLUGIN_ROOT}/scripts/config-read-skill-instructions.sh refine-work-item