round-prompting

name	Round Prompting
description	Write focused single-round execution prompts for autonomous iteration — the atomic unit of phase-prompting.

Round Prompting

Use this skill when writing a single execution round — the atomic unit of work within a larger phase. A round is typically 2-3 hours, 5-7 waves, with one dominant delivery goal.

For multi-round phase planning, see phase-prompting. For local LLM platform-specific rounds, see local-llm-phase-prompting.

When to Use

• You need to write the next round of autonomous iteration
• A phase prompt exists but the next concrete step needs scoping
• You're resuming work after a break and need to capture current state → next action

Round Structure

# Round N: <Verb> <Object> — <Measurable Outcome>

## Current State (verified)
- What works right now (cite test counts, commit hashes)
- What is partially done (be specific about gaps)
- What is blocked (and by what)

## Goal
One sentence: what is different when this round is done?

## Waves

### Wave 1: <name> (~XX min)
- Task 1: <specific action>
- Task 2: <specific action>
- Checkpoint: <how to verify wave success>

### Wave 2: <name> (~XX min)
...

### Wave N: Verify + Commit (~10 min)
- Run full test suite
- Verify no regressions
- Commit with descriptive message
- Update progress tracking

## Hard Rules
- <rule 1>
- <rule 2>

## Branch Conditions
- If <condition>: <alternative path>
- If blocked on <X>: skip to Wave N, note blocker

## Out of Scope
- <thing that might be tempting but isn't this round>

## Done When
- [ ] <measurable criterion 1>
- [ ] <measurable criterion 2>
- [ ] Tests pass (count: N → N+M)
- [ ] Committed and pushed

Writing Rules

1. Start from Verified State

Never write "Current State" from memory. Run commands, check test output, read recent commits. Cite evidence.

2. One Goal Per Round

If you have two goals, you have two rounds. Split them.

3. Every Wave Has a Checkpoint

A wave without a checkpoint is a wave that can silently fail. The checkpoint should be verifiable (test passes, command output, file exists).

4. Time-Box Waves

Estimate time per wave. If a wave exceeds estimate by 2x, it's a signal to split or simplify.

5. Branch Instead of Bloat

If Wave 3 depends on Wave 2 succeeding, write an explicit branch:

• If Wave 2 passes: proceed to Wave 3
• If Wave 2 fails: pivot to Wave 3-alt (diagnose + minimal fix)

6. Close with Verification

The final wave always includes: run tests, verify no regressions, commit, update tracking. Never skip this.

Anti-Patterns

Anti-Pattern	Fix
"Improve X" (vague)	"Add 3 tests for edge case Y in X" (specific)
10+ waves	Split into 2 rounds
No checkpoints	Add verification after each wave
Goal depends on external response	Mark as blocked, write alternative round
Repeating previous round's work	Check commit log before writing

Example

# Round 131: Wire skill match into system prompt — auto-inject on task start

## Current State (verified)
- Skill store: 49 skills, CRUD + export/import working (1618 tests)
- TF-IDF matcher: scoreMatch() returns top-N skills by relevance
- System prompt injection: not yet wired (translate/request.ts untouched)
- Last commit: b49f581 (Round 130)

## Goal
When a new task starts, automatically match relevant skills and inject
their guidance into the system prompt sent to the backend LLM.

## Waves

### Wave 1: Injection point (~20 min)
- Add injectSkillContext() to src/translate/request.ts
- Takes task description, returns augmented system prompt
- Checkpoint: unit test passes for injection with 0, 1, 3 skills

### Wave 2: Auto-match trigger (~30 min)
- Hook into onRequest plugin to extract task context
- Call matcher → inject top 3 skills above threshold 0.3
- Checkpoint: integration test with mock skill store

### Wave 3: Verify + Commit (~10 min)
- npx vitest run — all green
- Manual test: run the harness with a task that should trigger the debugging skill
- Commit: "feat: auto-inject matched skills into system prompt — Round 131"

## Hard Rules
- Never inject more than 3 skills (context budget)
- Threshold 0.3 minimum (avoid noise)
- Log injected skills to session meta (observability)

## Branch Conditions
- If matcher returns 0 results: proceed without injection (no-op)
- If system prompt exceeds 4000 tokens after injection: truncate oldest skill

## Out of Scope
- Embedding-based matching (future round)
- User-configurable skill preferences

## Done When
- [ ] Skills auto-injected on task start
- [ ] Integration test passes
- [ ] Tests: 1618 → 1625+
- [ ] Committed and pushed