conversation templates

templates for common task types, derived from analysis of 4,656 threads. optimized for the patterns that correlate with resolution.

debug

goal: identify and fix a specific issue

@path/to/problematic/file.ts

[symptom]: describe what's happening
[expected]: describe what should happen
[reproduction]: steps or command to trigger

hypothesis: [your best guess, if any]

why this works:

• file anchor (+25pp success rate)
• structured context (300-1500 char sweet spot)
• hypothesis signals collaborative debugging vs delegation

follow-up pattern (socratic, concise_commander-style):

• “what did you find?”
• “try running [specific test]”
• “ok, what’s next?“

feature

goal: implement new functionality

@path/to/relevant/area.ts @path/to/similar/example.ts

add [feature] that [does what]

acceptance:
- [ ] criterion 1
- [ ] criterion 2

constraints: [tech choices, patterns to match, things to avoid]

why this works:

• multiple anchors establish context
• explicit acceptance criteria reduce steering
• constraints prevent scope creep

anti-pattern: don’t front-load walls of context. let agent discover details. high initial context correlates with steering.

refactor

goal: improve code structure without changing behavior

@path/to/target.ts

refactor [what] to [goal]

keep: [behaviors that must not change]
pattern: [desired structure, or link to example]

why this works:

• explicit preservation constraints prevent breakage
• pattern reference gives target shape
• narrow scope (one file) prevents sprawl

confirmation gate: before major refactors, ask agent to outline plan. approval:steering ratio of 2-4:1 predicts success.

review

goal: evaluate code quality and correctness

@path/to/file.ts

review for: [specific concerns]
context: [why this matters, what changed]

follow-up options:

• “apply fixes” (if changes needed)
• “explain [specific concern]” (if unclear)
• approval signal: “lgtm” / “ship it”

why this works:

• focused review beats open-ended “review this”
• context prevents generic feedback
• clear exit signals (approval) close the loop

meta-patterns

optimal thread shape

• 26-50 turns: highest resolution rate (75%)
• steering recovery: if 2+ consecutive corrections, pause and ask “should we change approach?”
• don’t abandon: approval:steering ratio >1:1 usually recovers

prompting style

task spawning

use Task tool for:

• multi-layer changes (frontend + backend + api)
• token-heavy operations
• parallel independent work

avoid spawning for single-file changes. max productive depth: 6.

quick reference