| name | bead-polishing-ralph |
| description | Iteratively refine beads for ralph-tui execution through 4-6 polishing rounds until convergence, ensuring each bead fits one agent context window, has quality gates appended, uses correct bd CLI syntax, and follows schema-backend-UI dependency ordering. Use when user wants to polish beads for ralph, refine ralph tasks, iterate on beads before running ralph-tui, review bead quality for ralph execution, or check if beads are ralph-ready. Triggers on: polish beads ralph, refine beads for ralph, ralph bead review, check beads, are beads ready, polish for ralph-tui, iterate ralph beads. |
Bead Polishing for Ralph-TUI
Iteratively refine beads through 4-6 polishing rounds until convergence, enforcing all constraints that ralph-tui needs for successful autonomous execution.
This skill combines the convergence-tracked polishing loop from the Agentic Coding Flywheel with the specific format and sizing constraints required by ralph-tui's beads tracker.
Why This Matters
ralph-tui spawns a fresh agent instance per bead with no memory of previous work. A bead that is too big, vague, or has broken dependencies will cause the agent to either run out of context, produce the wrong thing, or block the entire pipeline. Polishing beads before execution is the single highest-ROI activity because:
- Plan mistakes cost 1x to fix (edit the markdown)
- Bead mistakes cost 5x to fix (rewrite + re-link + re-order)
- Code mistakes cost 25x to fix (debug + refactor + retest across files)
A 30-second fix to a bead description prevents a 30-minute failed ralph-tui iteration.
Before Starting
Detect CLI
Use bd (beads) as the CLI for all bead operations. The polishing process uses bd commands throughout.
Gather Context
- Read the source plan/PRD that the beads were created from. Polishing without the source is guessing.
- Extract Quality Gates from the PRD's "Quality Gates" section. If none exists, ask the user what commands should pass (e.g.,
pnpm typecheck, pnpm lint).
- Check if beads exist:
bd list --json
Create Beads if None Exist
If no beads exist yet, create them from the source PRD before polishing. This makes the separate "convert to tasks" step optional — bead-polishing-ralph can handle the full plan-to-polished-beads pipeline.
-
Create an epic linking back to the PRD:
bd create --type=epic \
--title="[Feature Name]" \
--description="$(cat <<'EOF'
[Feature description from PRD]
EOF
)" \
--external-ref="prd:./path-to-prd.md"
-
Create one bead per user story/requirement from the PRD. For each story:
- Title matches the story title
- Description includes story context, acceptance criteria, quality gates, and the
<promise>COMPLETE</promise> instruction
- Priority reflects dependency order (schema=1, backend=2, UI=3, polish=4)
- Pass
--parent=<epic-id> at creation time so the parent link exists from the start
-
Add dependencies between beads: bd dep add <issue> <depends-on>. NEVER add bd dep add <epic> <child> ("epic depends on child"). The parent-child link already represents "epic completes when children complete" — adding a reverse dep creates a cycle with the parent-child relation and breaks bd ready.
-
Identify the epic: Which epic are we polishing?
CRITICAL — Epic type. The epic MUST be --type=epic (not task / feature / bug). bd represents --parent=<epic> as a parent-child dependency on the child, and bd ready treats any open ancestor as a blocker UNLESS the ancestor is type=epic. A feature-typed parent will silently block every child from appearing in bd ready, which causes ralph-tui to report "no more tasks available after starting." Verify with bd show <id> — the line should read [epic].
This initial creation is intentionally rough — it's the "first pass" that the polishing rounds will refine. Don't over-invest here; the convergence loop catches everything.
The Five Ralph-Specific Checks
Every round applies all five checks. Each check is tuned for ralph-tui's execution model.
Check 1: Size — One Context Window Per Bead
ralph-tui's #1 constraint: each bead must be completable in one agent iteration (~one context window). The agent starts fresh with no memory of previous work.
Right-sized:
- Add a database column + migration
- Add a UI component to an existing page
- Update a server action with new logic
- Add a filter dropdown to a list
Too big (split these):
- "Build the entire dashboard" → schema, queries, UI components, filters
- "Add authentication" → schema, middleware, login UI, session handling
- "Refactor the API" → one story per endpoint or pattern
Rule of thumb: If you can't describe the change in 2-3 sentences, it's too big for ralph-tui.
Split oversized beads into vertical slices that each produce a working increment. Use bd create to make the new beads and bd close + recreate if an existing bead needs splitting.
Check 2: Quality Gates Appended
Every bead's acceptance criteria must end with the quality gates from the PRD. This is not optional — ralph-tui agents check these criteria to verify their own work.
Structure per bead:
## Acceptance Criteria
- [ ] [Story-specific criterion 1]
- [ ] [Story-specific criterion 2]
- [ ] pnpm typecheck passes <- universal gate
- [ ] pnpm lint passes <- universal gate
- [ ] Verify in browser <- UI gate (only for UI stories)
When all acceptance criteria are met, output <promise>COMPLETE</promise>
CRITICAL: The <promise>COMPLETE</promise> signal is how ralph-tui detects that a bead is done. Exit code 0 alone does NOT trigger completion — only this explicit signal does. Every bead description MUST include this instruction, or the agent will finish its work but ralph-tui will not advance to the next bead.
Check for:
- Beads missing the
<promise>COMPLETE</promise> instruction (most critical)
- Beads missing quality gates entirely
- Beads with quality gates but missing story-specific criteria
- UI beads missing browser verification gates
- Quality gates that don't match the PRD's Quality Gates section
Check 3: Dependency Integrity + Parent Linkage
Two independent graph properties must both be correct for ralph-tui to schedule anything. They look similar but answer different questions: parent linkage answers "which beads belong to this epic?" (ralph-tui's --epic <id> filter), and dependencies answer "in what order can they run?". Missing parent links and missing deps fail in different ways.
Parent linkage (ralph-tui's epic filter)
CRITICAL: ralph-tui --epic <id> filters by parent, NOT by the dependency graph. Beads that block the epic via bd dep add <epic> <child> are invisible to ralph-tui unless they also carry parent=<epic>. Symptom: ralph-tui reports "no tasks or epic found" despite a populated bd ready queue. This is the #2 cause of failed ralph runs (after missing <promise>COMPLETE</promise> signals).
Verify:
- The epic itself has
type=epic (see bd show <epic-id> — the header should read [epic]). A feature/task-typed parent silently blocks every child from bd ready.
- Every child bead has
parent = <epic-id> set explicitly.
bd show <epic-id> lists every child under a CHILDREN section. If a bead you expect to see is missing, its --parent was never set.bd ready returns at least the schema/leading bead. If it returns "No ready work found" despite open children, suspect (1) wrong epic type and/or (2) a redundant bd dep add <epic> <child> creating a cycle with the parent-child relation.
Check with: bd show <epic-id> and confirm the CHILDREN list matches the intended child set. Then bd ready --explain to see why anything is reported as blocked — if a child is "blocked by [the epic]", the epic is mistyped or the cycle exists.
Fix with:
bd update <epic-id> --type=epic if the epic is mistyped.bd update <child-id> --parent=<epic-id> for every child missing the link.bd dep remove <epic-id> <child-id> for any redundant "epic depends on child" entry that creates a cycle with parent-child.
Dependency ordering
ralph-tui respects dependencies strictly within an epic: it will never select a bead whose dependencies are still open. Broken dependencies block the entire pipeline.
Verify:
- No circular dependencies
- Correct ordering: schema -> backend -> UI -> integration
- No bead depends on something that should come after it
- No missing dependencies (bead B uses what bead A creates, but no dep link exists)
- No unnecessary dependencies (bead marked blocked when it could start immediately)
Check with: bd list --json and examine the dependsOn arrays. Mentally simulate: "If ralph-tui executes these in dependency order, does each agent have everything it needs?"
Fix with: bd dep add <issue> <depends-on> or bd dep remove <issue> <depends-on>
Check 4: Description Completeness
Each bead description must give a fresh agent (with zero prior context) everything it needs to execute. ralph-tui includes the bead description in the agent prompt — that description is all the agent gets.
Must include:
- What to change and where (specific files, tables, components)
- Why (enough context to make reasonable decisions)
- Acceptance criteria (verifiable, not vague)
- Quality gates (appended)
Red flags:
- "Works correctly" — not verifiable
- "Good UX" — not verifiable
- "Handle edge cases" — which ones?
- Missing file paths or component names
- Assumes knowledge from a previous bead (the agent won't have it)
The key test: Could an agent with access to the codebase but no other context complete this bead using only its description?
Check 5: Coverage Against Source Plan
Compare beads against every section, feature, and requirement in the source PRD.
- Walk through the PRD section by section
- For each requirement, confirm at least one bead addresses it
- Flag requirements with no bead as gaps -> create new beads
- Flag beads with no corresponding requirement as orphans -> investigate
Round-by-Round Emphasis
All five checks run every round, but each round leads with a different lens.
Round 1: Structural Audit (Checks 1 + 5)
Focus on shape. Are beads the right size for one context window? Are there obvious duplicates? Does the bead set cover the PRD? This round produces the most changes — splits, merges, new beads.
Round 2: Dependencies + Quality Gates (Checks 2 + 3)
Now that the structure is clean, verify the dependency graph and quality gate completeness. Run bd list --json and trace the execution order. Append missing quality gates. This round typically fixes 3-5 dependency links and adds gates to 2-3 beads.
Round 3: Description Depth (Check 4)
With the right beads identified, invest in making each one self-contained. Rewrite vague criteria. Add file paths, API references, table names. Remember: a fresh agent with no memory will read this description and nothing else.
Round 4: Adversarial Review (All Checks)
Read every bead as if you are a ralph-tui agent encountering it for the first time with zero context beyond the description. Ask:
- "Could I complete this in one context window?"
- "Do I have enough information to start without asking questions?"
- "Would a different agent interpret this description the same way I do?"
- "If I run the quality gates after implementing, will they actually pass?"
Rounds 5-6: Only if Needed
Run additional rounds only if convergence is below 75% after round 4.
Convergence Tracking
After each round, record metrics:
## Polishing Log
### Round 1
- Beads added: 3
- Beads removed: 1 (merged duplicate)
- Beads modified: 6
- Beads split: 2 (too big for one iteration)
- Dependencies changed: 3
- Quality gates added: 4
- Total changes: 19
- Notes: Split "build dashboard" into 4 slices, merged duplicate schema beads
### Round 2
- Beads added: 1
- Beads modified: 5
- Dependencies changed: 4
- Quality gates added: 2
- Total changes: 12
- Change velocity: -37% (converging)
### Round 3
- Beads modified: 4
- Quality gates added: 0
- Total changes: 4
- Change velocity: -67% (converging)
- Notes: Expanded descriptions with file paths and component names
### Round 4
- Beads modified: 1
- Total changes: 1
- Convergence: ~85%
- Notes: Clarified one ambiguous acceptance criterion
When to Stop
- ~75% convergence (4 rounds): Good default. Further polishing yields diminishing returns.
- ~90% convergence: Maximum useful refinement.
- Change velocity near zero for 2 consecutive rounds: Stop.
- Changes increasing: Something is structurally wrong — step back and check the source PRD.
Applying Changes
After each round, apply changes using the appropriate CLI:
bd update <bead-id> --description="$(cat <<'EOF'
[updated description with quality gates]
EOF
)"
bd dep add <issue> <depends-on>
bd dep remove <issue> <depends-on>
bd create --parent=<epic-id> \
--title="[Story Title]" \
--description="$(cat <<'EOF'
[description with acceptance criteria + quality gates]
EOF
)" \
--priority=[1-4]
bd close <bead-id>
CRITICAL: Always use <<'EOF' (single-quoted) for HEREDOCs. This prevents shell interpretation of backticks, $variables, and () in descriptions.
Presenting Results
After the final round, present:
- Summary: Rounds run, beads before/after, key structural changes
- The polishing log (all rounds with metrics)
- Ralph-readiness assessment:
- Are all beads one-context-window sized?
- Are quality gates on every bead?
- Is the dependency graph clean?
- Are all descriptions self-contained?
- Remaining concerns: Anything needing human input
Then suggest:
Ready to run: ralph-tui run --tracker beads --epic <epic-id>
Anti-Patterns
Polishing without the source PRD. You cannot verify coverage without the original requirements.
Beads that assume prior context. ralph-tui agents start fresh. If bead 3 says "continue the work from bead 2," the agent will have no idea what bead 2 did. Each description must be self-contained.
Vague quality gates. "Tests pass" is not a quality gate. pnpm typecheck && pnpm lint is.
Over-polishing. If you're on round 7 and still finding issues, the PRD is underspecified. Fix the PRD, regenerate beads, then polish.
Missing completion signal. Every bead must end with When all acceptance criteria are met, output <promise>COMPLETE</promise>. Without this, ralph-tui cannot detect that the agent finished and the loop stalls. This is the #1 cause of "stuck" ralph-tui runs.
Missing parent linkage. Wiring children only via bd dep add <epic> <child> is not enough — ralph-tui's --epic <id> filter reads the parent field, not the dependency graph. Symptom: ralph-tui reports "no tasks or epic found" even though bd ready shows work. Always run bd show <epic-id> and confirm every intended child appears under CHILDREN before suggesting the ralph-tui command.
Mistyped epic. Creating the epic with --type=feature (because bd prime only lists task|bug|feature in its quick-reference) causes bd to treat the parent-child link from each child to the epic as a regular blocking dependency. bd ready returns "No ready work found", ralph-tui reports "no more tasks available after starting", and the session sits at iteration 0. Fix: bd update <epic-id> --type=epic. The epic type is supported by bd update --type= even when bd prime doesn't advertise it.
Cyclic epic-child deps. Adding bd dep add <epic> <child> "so the epic closes when children close" creates a cycle: parent-child says child depends on epic, bd dep add says epic depends on child. bd ready --explain will show every node "blocked by" something in the cycle. Fix: remove the explicit bd dep add entries — parent-child + closing children already gives you "epic closes when children close" semantically.
Skipping the size check. The most common ralph-tui failure is beads that are too large for one context window. When in doubt, split.
Checklist
Before declaring beads ralph-ready:
