new-task

Name: new-task
Author: minghinmatthewlam

34stars4forks

How to install new-task?

minghinmatthewlam/agent-guards/new-task

$npx -y skills add minghinmatthewlam/agent-guards --skill new-task

Installs into the current project.

Summary

Iteratively clarify requirements for a new task until reaching 90%+ confidence, then auto-invoke plan-loop.

Files · 1

View on GitHub

SKILL.md

1	# New Task
2
3	Iterative clarification loop: refine understanding of a new task until it is precise, testable, and ready for planning.
4
5	Consider `$ARGUMENTS` if provided.
6
7	## Loop
8
9	Each round:
10
11	1. Present current understanding: problem statement, requirements, assumptions. Show confidence as a single percentage.
12	2. Ask 3-7 targeted questions, prioritized by what will increase confidence most. Focus on: edge cases, product decisions, integration points, success criteria.
13	3. Wait for user answers — update understanding and recalculate confidence.
14
15	Keep iterating until confidence reaches 90%+.
16
17	## Question Strategy
18
19	- Start broad (problem space, users, goals), narrow each round (edge cases, constraints, acceptance criteria).
20	- Make questions concrete with examples when ambiguity is high.
21	- Group related questions; don't repeat answered ones.
22
23	## Transition to Planning
24
25	Once confidence reaches 90%+:
26
27	1. Explicitly call out the specific use cases and features you'll test end-to-end (`self-test`).
28	2. Use the Skill tool to invoke `plan-loop` with the finalized requirements and e2e test targets as context. Do NOT plan on your own — `plan-loop` uses multi-agent research, dual-model review, structured verification design, and `self-test` integration that ad-hoc planning skips.
29
30	If `plan-loop` is unavailable, tell the user and stop — do not substitute your own planning.
31
32	## Gotchas
33
34	- The #1 failure mode: agent reaches 90% confidence and starts planning/implementing directly instead of invoking `plan-loop` via the Skill tool. This defeats the entire purpose of `new-task` — the value is the structured handoff, not just the clarification.
35	- Don't inflate confidence to reach 90% faster. If key questions remain, stay in the loop.
36	- Don't ask questions the user already answered in earlier rounds.

People who liked this also viewed

6 picks

Type

new-task

34stars4forks

View on GitHub

How to install new-task?

minghinmatthewlam/agent-guards/new-task

$npx -y skills add minghinmatthewlam/agent-guards --skill new-task

Installs into the current project.

Summary

Iteratively clarify requirements for a new task until reaching 90%+ confidence, then auto-invoke plan-loop.

Files · 1

View on GitHub

SKILL.md

1	# New Task
2
3	Iterative clarification loop: refine understanding of a new task until it is precise, testable, and ready for planning.
4
5	Consider `$ARGUMENTS` if provided.
6
7	## Loop
8
9	Each round:
10
11	1. Present current understanding: problem statement, requirements, assumptions. Show confidence as a single percentage.
12	2. Ask 3-7 targeted questions, prioritized by what will increase confidence most. Focus on: edge cases, product decisions, integration points, success criteria.
13	3. Wait for user answers — update understanding and recalculate confidence.
14
15	Keep iterating until confidence reaches 90%+.
16
17	## Question Strategy
18
19	- Start broad (problem space, users, goals), narrow each round (edge cases, constraints, acceptance criteria).
20	- Make questions concrete with examples when ambiguity is high.
21	- Group related questions; don't repeat answered ones.
22
23	## Transition to Planning
24
25	Once confidence reaches 90%+:
26
27	1. Explicitly call out the specific use cases and features you'll test end-to-end (`self-test`).
28	2. Use the Skill tool to invoke `plan-loop` with the finalized requirements and e2e test targets as context. Do NOT plan on your own — `plan-loop` uses multi-agent research, dual-model review, structured verification design, and `self-test` integration that ad-hoc planning skips.
29
30	If `plan-loop` is unavailable, tell the user and stop — do not substitute your own planning.
31
32	## Gotchas
33
34	- The #1 failure mode: agent reaches 90% confidence and starts planning/implementing directly instead of invoking `plan-loop` via the Skill tool. This defeats the entire purpose of `new-task` — the value is the structured handoff, not just the clarification.
35	- Don't inflate confidence to reach 90% faster. If key questions remain, stay in the loop.
36	- Don't ask questions the user already answered in earlier rounds.

People who liked this also viewed

6 picks

Type