Guide · Agent skill
How to use the Structured Deep Reasoning agent skill (think)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The think skill, published in the claude-obsidian repository under skills/think, makes an AI agent work through a hard or ambiguous decision in a structured way. It runs a ten-stage review loop that checks evidence, questions the agent's own biases, looks for hidden risks and argues against its preferred answer. This guide explains what the skill does, how to install it, how to trigger it with natural requests, and what to watch out for, so you can use it on architecture reviews, postmortems and irreversible choices.
Structured Deep ReasoningAgriciDaniel/claude-obsidian · MIT Download the skill Quick start
- 1Download the ZIP. Get the think skill ZIP from this skill's page in the directory.
- 2Unzip the skill folder. Extract the archive so the think folder containing SKILL.md is ready to copy.
- 3Place it in the skills folder. Copy the folder to ~/.claude/skills/ for personal use, or to .claude/skills/ inside a project.
- 4Or upload it in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the skill.
- 5Describe the decision. Write the decision, its context, constraints and the option you prefer in plain language.
- 6Use a trigger phrase. Include a phrase such as think this through, architecture review or postmortem so the skill is applied.
- 7Review the output and revisit. Check the highest-risk claim against the evidence, then use the growth loop to set when to reassess.
What the think skill makes the agent do
The skill applies a loop of ten stages: OBSERVE (external), OBSERVE (internal), LISTEN, THINK, CONNECT (lateral), CONNECT (system), FEEL, ACCEPT, CREATE and GROW. Each stage pushes the agent to look at a different angle, such as the actual inputs and constraints, the biases likely to distort the decision, the user's intent, upstream and downstream effects, and what feedback should later update the decision.
Before the loop, the skill asks the agent to separate observed facts, inferences, unknowns and preferences. The ACCEPT stage requires the agent to state unfavourable findings plainly and present the strongest counterargument to its preferred conclusion. The output is a recommendation plus a plan to revisit it, not a finished change to your project.
- Observed: directly supported by an input, source, test or measurement
- Inferred: a conclusion drawn from observations, with the bridge stated
- Unknown: missing evidence that could change the decision
- Preference: a user or designer choice, not a factual claim
Install the think skill from the ZIP or the Claude apps
Download the ZIP for the skill from the skill's page in this directory, then unzip it. For a personal install, place the unzipped skill folder in ~/.claude/skills/. To scope it to one project, place it in that project's .claude/skills/ folder instead.
If you use the Claude apps, you can upload the skill under Settings → Capabilities → Skills rather than copying files manually. The skill's files are SKILL.md only, so the folder should contain that file.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside the project
- Claude apps: Settings → Capabilities → Skills
~/.claude/skills/think/SKILL.mdTrigger the skill with natural requests
The skill's description lists phrases that should activate it: think this through, deep think, architecture review, postmortem, tradeoff analysis and challenges to assumptions. You do not need special syntax. State the decision, the context and what you need back, and include one of these phrases so the agent knows to apply the loop.
Give the agent the material it needs. The skill can only reason from the evidence you supply, so paste the relevant error logs, design notes or constraints where you can.
- "Think this through: should we move our job queue to a managed service? Here are the constraints and costs."
- "Do a deep think on whether we should drop support for the v1 API."
- "Run a postmortem-style review of this outage timeline, and challenge my assumptions about the root cause."
- "Give me a tradeoff analysis of the two database options below, and argue against my preferred one."
Examples of when the skill helps
Use the skill when a decision is consequential, ambiguous or hard to reverse. It is useful for architecture reviews, where you want the design checked against upstream inputs, downstream consumers and failure modes. It also suits postmortems, where the ACCEPT and CONNECT stages can surface contributing causes you may have skipped.
For a simple lookup or a typo fix, skip it. The skill itself says trivial, reversible tasks should skip the formal loop, and running it there only adds overhead.
- Architecture review of a service split before committing to it
- Postmortem on an incident, looking for hidden connections and risks
- Irreversible choice, such as retiring a public feature
- Devil's-advocate pass on a plan before a team signs off
What the skill needs and its safety notes
The skill lists no requirements. It does not run scripts and does not use the network. It is read-only: it may inspect the evidence available to it, but it does not edit files, commit to Git or publish anything. If its CREATE stage recommends a change, the source says that a separate mutation skill should be used only after you authorise that operation.
Because the skill is read-only and runs no scripts, the main safety question is what you paste into the conversation. Remove secrets, credentials and personal data from logs or documents before sharing them with the agent.
- Requirements: none listed
- Runs scripts: no
- Uses network: no
- Read-only: it reasons and recommends, it does not change files
Tips for getting useful output
Supply concrete evidence and say what success looks like. The skill scales its depth to novelty, stakes, uncertainty and irreversibility, so state those up front. Ask it to seek disconfirming evidence, since the verification guidance favours counterexamples over supporting examples.
Treat the output as a reasoning aid. Its conclusions depend on the evidence you give it, so check the highest-risk claim yourself before acting on it. Ask for the growth loop section too, so you get a revisit condition and a feedback signal.
- Name the decision and the deadline in the first sentence
- Include the constraints and the option you currently prefer
- Ask for the strongest counterargument explicitly
- Ask for the revisit condition and the feedback signal
What people use it for
- Architecture review of a planned service change, checked against upstream and downstream effects
- Postmortem of an outage, challenging the assumed root cause
- Tradeoff analysis between two technology options, with the preferred option argued against
- Devil's-advocate review of a product decision before a team commits to it
- Structured reasoning about an irreversible choice such as deprecating a public API
Frequently asked questions
Is the think skill free to use?
The skill is published under the MIT licence in the claude-obsidian repository. The source does not mention pricing, so check the licence file and your own agent platform's terms for any costs.
What does the think skill require to run?
The skill's listed requirements are none. It does not run scripts and does not use the network, and it is a single SKILL.md file.
How do I install the think agent skill?
Download the ZIP from the skill's page and unzip it. Place the folder in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder, or upload it in the Claude apps under Settings → Capabilities → Skills.
Does the think skill change my files or publish anything?
No. The skill is explicitly read-only. It may inspect available evidence, but it does not edit files, commit to Git or publish. Any change is a separate operation you authorise.
Is the think skill a health check or repair tool for my vault?
No. The skill describes itself as a reasoning review, not a deterministic vault health check or repair workflow. Use it for judgement on decisions, not for guaranteed automated checks or fixes.
When should I not use the think skill?
Skip it for simple, reversible lookups or typo fixes, since the skill itself says those should skip the formal loop. It is also unsuitable where you need a guaranteed deterministic check, and its output is not verified fact.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
The think skill suits engineers preparing architecture reviews or postmortems, and anyone facing a consequential or irreversible decision who wants assumptions and counterarguments surfaced. It is a read-only reasoning framework, so its conclusions are only as good as the evidence you supply. If you need a deterministic check, automated repair or a quick answer to a simple question, pick a different tool.
Download the skill