Second Thought

The anti-hype agent skill. Critique-first, decision-checkpoint engine for AI agents.

A portable agent skill that challenges every high-commitment decision before execution — catching flaws, surfacing risks, questioning assumptions, and pushing back when an agent is about to confidently accelerate in the wrong direction. Works across Claude Code, Cursor, Codex, Gemini CLI, OpenCode, and GitHub Copilot CLI.

A lot of LLMs and agents have the same default personality:

“Hell yeah brother👏🙌, incredible idea🧠💪, absolutely cooked🔥, let’s ship it 🚀🚚.”

Unfortunately, enthusiasm is not judgment.

Too many agents are eager to agree, amplify, and move forward, even when the idea is undercooked, the plan drifted from the original ask, or the human really needed pushback instead of applause.

Second Thought is the anti-hype skill.

Because not every idea deserves a standing ovation. Some ideas deserve a second look.

When to invoke

Trigger words / moments: second-thought, double check, sanity check, before we ship, before we commit, before we deploy, before we merge, is this right?, let's lock this in, ship it, final plan, destructive or irreversible actions, major refactors/migrations.

Skip it for: low-stakes exploration, early brainstorming, casual Q&A, or anything the user has explicitly bypassed with [st:skip <reason>] or #st-skip.

What is this?

Second Thought is a reusable decision checkpoint engine that can be added to many kinds of agents.

It is a standalone decision checkpoint engine for agents that challenges decisions before they turn into action. Instead of overhyping every plan, it revisits the original human intent, questions weak assumptions, checks whether the current direction is actually the right one, and recommends when the agent should revise, challenge, or prompt the human / engineer before moving forward.

Its job is simple:

before an agent commits to a direction, plan, decision, or action, make it stop and give it a second thought: is this actually the right move?

This skill is designed to work across different agent types and workflows. It is not limited to coding agents or implementation tasks. It can be attached to agents that brainstorm, plan, write, research, design, analyze, or take tool-based actions.

The point is not to make agents slower for the sake of it.

The point is to make them less blindly agreeable, more goal-aligned, and more intellectually honest.

How to use

Second Thought can auto-trigger when a conversation starts hardening into a high-commitment decision or action.

You can also type second-thought at any point to run a manual checkpoint.

Installation

Installation differs by platform.

Codex (CLI or App)

Paste this to your agent:

Follow the installation instructions here: https://github.com/gg-mo/second-thought-agent-skill/blob/main/.codex/INSTALL.md

Gemini CLI

gemini extensions install https://github.com/gg-mo/second-thought-agent-skill

OpenCode

Follow .opencode/INSTALL.md.

Claude Code

Option A: Claude Code Plugin (recommended)

From within Claude Code, add the marketplace and install the plugin:

/plugin marketplace add gg-mo/second-thought-agent-skill
/plugin install second-thought@second-thought

Option B: local development marketplace (if you're testing local changes)

git clone https://github.com/gg-mo/second-thought-agent-skill.git ~/.claude/second-thought
/plugin marketplace add ~/.claude/second-thought/.claude-plugin/marketplace.json
/plugin install second-thought@second-thought

Option C: CLAUDE.md (per-project)

New project:

curl -o CLAUDE.md https://raw.githubusercontent.com/gg-mo/second-thought-agent-skill/main/CLAUDE.md

Existing project (append):

echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/gg-mo/second-thought-agent-skill/main/CLAUDE.md >> CLAUDE.md

Cursor

Option A: install from marketplace (if available in your Cursor environment)

Open Cursor Agent chat.
Run:

/add-plugin second-thought

Confirm install in the plugin manager.

Option B: local plugin install (recommended for this repo right now)

Clone locally:

git clone https://github.com/gg-mo/second-thought-agent-skill.git ~/.cursor/second-thought

In Cursor Agent chat, run:

/add-plugin

Choose local/plugin manifest install and select:

~/.cursor/second-thought/.cursor-plugin/plugin.json

Reference manifest in this repo: .cursor-plugin/plugin.json

GitHub Copilot CLI

Follow docs/README.copilot.md.

Why this exists

LLMs and agents are often very good at sounding helpful while being a little too eager to agree.

That shows up in all kinds of ways:

cheering on weak ideas,
locking into a plan too early,
mistaking momentum for correctness,
following the latest instruction without checking it against the original goal,
and pushing forward when they should have challenged, clarified, or escalated.

One of the most common failure modes is not lack of intelligence. It is over-obedience.

An agent hears an idea and responds with:

“Yes”
“Great idea”
“Let’s do it”
“That makes sense”

But sometimes the correct response is:

“Maybe”
“Not yet”
“This drifted from the original goal”
“This sounds good in the moment, but are we solving the right problem?”
“We should challenge this before locking it in”

Second Thought exists to create that moment of resistance.

Not useless negativity.
Not random contrarianism.
Not fake caution theater.

Just real critique, at the right time.

The core idea

Agents should not blindly act on their first thought.

They also should not blindly validate every idea that emerges from a back-and-forth with a human or engineer.

A conversation can feel productive and still drift. A decision can feel settled and still be wrong. A plan can sound polished and still fail the original objective.

Second Thought is built to catch that.

It asks the agent to:

revisit the original human intent,
compare the current proposal against that intent,
question whether the latest direction is actually the right one,
surface hidden assumptions or risks,
consider whether a better path exists,
and decide whether it should proceed, revise, or prompt the human / engineer.

What it does

At meaningful checkpoints, the agent sends its current proposal through Second Thought.

That proposal might be:

an idea,
a chosen direction,
a design decision,
a plan,
a tool call,
a risky next step,
or an action the agent is about to take.

The skill then critiques that proposal in context.

It asks questions like:

Does this still match the human’s original intent?
Did the conversation drift toward something easier, louder, or more convenient rather than more correct?
Is this decision based on assumptions that have not been verified?
Is the agent prematurely locking in a direction?
Is there a simpler, safer, or sharper alternative?
Is this action reversible?
Is this the moment to prompt the human or engineer instead of continuing?
Is the agent being thoughtful, or just being agreeable?

It then returns structured guidance that helps the agent:

proceed
revise
or prompt the human / engineer

Where this skill matters most

1. During brainstorming

During ideation, agents should be collaborative and helpful, but not mindlessly validating everything.

If a human suggests a direction, the agent should explore it. But once a direction starts hardening into a real decision, the agent should step back and challenge it.

Agreement is not proof. Momentum is not proof. Excitement is not proof.

2. During decision convergence

This is one of the biggest targets of the skill.

A human and an agent may go back and forth, refine ideas together, and gradually converge on a plan. At that moment, many agents simply accept the result and move on.

Second Thought does the opposite.

When a design, decision, or plan appears to be settling, it re-evaluates that outcome against the original intent and asks:

Is this really it?
Are we solving the actual problem?
Did we converge because the direction is right, or because it was easy to agree on?
Did anyone challenge the weak parts?

3. Before execution

Before a meaningful action is taken, the agent should run the decision through critique.

This is especially useful before:

tool calls,
code changes,
file edits,
destructive actions,
external writes,
public output,
high-cost moves,
or steps that are hard to reverse.

Why this is different

Second Thought is not just a pre-execution reviewer.

It is a decision checkpoint engine that specifically targets convergence risk: moments where agreement forms quickly and can be mistaken for correctness.

Philosophy

The best agents are not the ones that always move fast.

They are not the ones that always agree.

They are the ones that know when to stop, revisit the original intent, question the current direction, and ask whether the next move actually deserves to happen.

Second Thought is built around that mindset.

It is not trying to make agents timid.
It is trying to make them responsible.

It is not trying to make agents argumentative.
It is trying to make them honest.

It is not trying to make agents block progress.
It is trying to make sure the agent is not confidently accelerating in the wrong direction.

Design principles

Goal-first, not action-first

Every critique starts by grounding itself in the original human goal, not just the latest message in the conversation.

Challenge convergence, not just execution

The skill should not wait until implementation time. It should also challenge decisions when they begin to harden during discussion.

Push back when needed

A good agent should not be a hype machine. Sometimes the right move is to challenge the direction rather than cheer it on.

Prompt humans intentionally

If the path is ambiguous, high-impact, irreversible, or under-specified, the agent should be willing to pause and ask the human or engineer.

Stay portable

This is meant to be a standalone skill that can be used by many different agents, stacks, and workflows.

Improve thinking without killing momentum

Critique should be sharp and useful, not bloated or performative.

What a strong critique looks like

A strong critique does not just say “this might be bad.”

It does things like:

reconnect the current direction to the original ask,
point out where the conversation drifted,
identify weak or untested assumptions,
flag risk, cost, or irreversibility,
suggest a better alternative,
call out premature certainty,
recommend a clarification question,
or explain why human input is needed before continuing.

A weak critique is vague, generic, or purely oppositional.

A strong critique is specific enough to improve the next step.

Decision outcomes

Every review from Second Thought should generally end in one of three outcomes.

Proceed

The current proposal appears aligned, reasonable, and safe enough to continue.

Revise

The current proposal has flaws, drift, weak assumptions, or a better alternative exists. The agent should improve it before moving forward.

Prompt Human / Engineer

The path is ambiguous, under-specified, high-impact, or likely to benefit from human judgment. The agent should pause and ask instead of pretending certainty.

Example use cases

Coding agents

Before changing code, the agent can verify that the proposed change actually solves the user’s real problem, not just the symptom discussed most recently. It can also decide whether a broader change should be confirmed with the engineer first.

Planning agents

Before finalizing a workflow, roadmap, or strategy, the agent can test whether the proposed path still matches the original objective and whether it is prematurely committing to a convenient but weak plan.

Writing agents

Before drafting, rewriting, or finalizing content, the agent can verify that tone, structure, and argument still reflect the original ask rather than drifting into something polished but misaligned.

Research agents

Before summarizing findings or making a recommendation, the agent can challenge whether the available evidence is actually sufficient and whether uncertainty should be stated more honestly.

Design agents

Before settling on a product or UX direction, the agent can ask whether the chosen approach addresses the true user need or simply became the most discussable idea during collaboration.

General-purpose assistants

Any agent involved in brainstorming, synthesis, or multi-step work can use this skill to avoid becoming a passive “yes machine.”

Brainstorming and decision convergence

This skill is especially important in human-agent collaboration.

A lot of systems assume the dangerous part happens at execution time. But many bad outcomes begin earlier, at the moment when discussion turns into commitment.

A human proposes something.
The agent builds on it.
The conversation gains momentum.
Everyone starts feeling good.
A direction gets chosen.

That can look like progress. Sometimes it is.

Sometimes it is just unchallenged agreement.

Second Thought is designed to catch that exact moment.

When a design, decision, or plan starts to solidify, the skill should step back and ask:

Does this still reflect the original intent?
Did the discussion drift?
Are we locking this in because it is right, or because it is the latest thing everyone accepted?
Are we rewarding confidence over correctness?
Should the agent challenge this before moving forward?

The role of the skill is not to be annoying.
The role of the skill is to prevent shallow convergence from being mistaken for good judgment.

Intended integration style

Second Thought is meant to be portable.

An agent should be able to call it whenever it has:

a candidate next step,
a proposed decision,
a plan,
a direction emerging from brainstorming,
a tool call,
a risky action,
or uncertainty about whether to continue.

That means the skill should stay:

standalone,
reusable,
lightweight enough to wrap around many workflows,
and independent of any one framework or stack.

This is not meant to be locked into a single implementation style. It should be usable by all kinds of agents.

Ideal moments to invoke the skill

Use Second Thought when:

a discussion is starting to turn into a decision,
an agent wants to finalize a design or direction,
the plan sounds good but has not been checked against the original goal,
the agent is about to take a meaningful action,
the proposed step is costly or hard to reverse,
the agent is operating on incomplete context,
the user’s real intent may have been lost,
or the agent feels suspiciously “too sure.”

If the agent is about to say “great idea, let’s do it,” that is often a very good time to run Second Thought.

Non-goals

This skill is not trying to:

reject every idea,
slow down every workflow unnecessarily,
be contrarian for the sake of it,
replace human judgment,
or create fake rigor through generic caution.

It is also not only for implementation review.

Its job is broader than that:

challenge emerging decisions,
catch goal drift,
and improve the quality of commitment before action happens.

Summary

Second Thought is a standalone decision checkpoint engine for agents.

It helps agents stop overhyping every idea, revisit the original human intent, challenge weak or drifting decisions, and determine whether they should proceed, revise, or prompt the human / engineer before moving forward.

In short:

it helps agents think twice before they do something once.

Implementation in this repo

This repository now includes both:

a reusable Python critique engine (second_thought/), and
portable skill/plugin scaffolding for multiple coding agents (skills/, plugin manifests, and install docs).

Core code

second_thought/models.py - typed input/output schemas
second_thought/engine.py - modular rule-based critique engine
second_thought/prompts.py - LLM critique prompt and prompt builder
scripts/second_thought_cli.py - CLI wrapper (stdin or --input JSON)

Skills

skills/second-thought/SKILL.md - core decision checkpoint contract
skills/using-second-thought/SKILL.md - startup/process wrapper for consistent usage

Platform integration files

Claude plugin manifest: .claude-plugin/plugin.json
Cursor plugin manifest: .cursor-plugin/plugin.json
Codex plugin manifest: .codex-plugin/plugin.json
Codex install guide: .codex/INSTALL.md
Gemini extension manifest: gemini-extension.json + GEMINI.md
OpenCode plugin: .opencode/plugins/second-thought.js + .opencode/INSTALL.md
Session-start hooks: hooks/

Hybrid trigger model (auto-gate + manual escape hatch)

Second Thought now uses a hybrid trigger model:

Automatic gates at high-commitment moments.
Manual invocation remains available for explicit checkpointing.

Automatic gates are currently implemented in OpenCode plugin runtime transforms, and startup behavior is enforced across other integrations through using-second-thought bootstrap instructions.

High-commitment trigger examples:

commit/merge/deploy/release/publish intent
destructive or irreversible actions
costly/high-impact execution steps
major edits/refactors/migrations

One-shot bypass (manual escape hatch):

[st:skip <reason>]
#st-skip

Bypass should be acknowledged explicitly and used intentionally, not as default behavior.

Run the critique engine

python3 scripts/second_thought_cli.py --input payload.json

Or:

cat payload.json | python3 scripts/second_thought_cli.py

Auto-fill mode (default, no manual schema fields required):

echo "We are about to deploy a risky migration; are we sure?" | python3 scripts/second_thought_cli.py

Strict manual mode:

python3 scripts/second_thought_cli.py --input payload.json --input-mode manual

For agent chat UX, keep JSON as an internal contract and return a short conversational critique by default (a few sentences). Only emit raw JSON when explicitly requested for debugging/integration.

Payload schema is the same cross-platform skill contract described in the second-thought skill and in the Python models.

Schemas: see schemas/second_thought_input.schema.json and schemas/second_thought_output.schema.json for integration validation.

Manifest source of truth

Platform manifests are generated from config/platform_manifest_config.json.

Regenerate with:

./scripts/generate_platform_manifests.py

Benchmarking

Run fixture benchmark scenarios:

./scripts/run_benchmark.py

The benchmark currently enforces a minimum pass rate of 75%.

SDK helpers

Python helper: sdk/python/second_thought_sdk.py
TypeScript helper: sdk/ts/index.ts

API contract

Versioned contract: docs/api/v1.md
Contract changelog: docs/api/changelog.md

Confidence calibration

./scripts/calibrate_confidence.py
./scripts/run_benchmark.py

Name		Name	Last commit message	Last commit date
.claude-plugin		.claude-plugin
.codex-plugin		.codex-plugin
.codex		.codex
.cursor-plugin		.cursor-plugin
.github		.github
.opencode		.opencode
benchmarks		benchmarks
commands		commands
config		config
docs		docs
hooks		hooks
schemas		schemas
scripts		scripts
sdk		sdk
second_thought		second_thought
skills		skills
tests		tests
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
GEMINI.md		GEMINI.md
INSTALL.md		INSTALL.md
LICENSE		LICENSE
README.md		README.md
examples.py		examples.py
gemini-extension.json		gemini-extension.json
package.json		package.json

Latest commit

History

Folders and files

Repository files navigation