Your thoughts don’t need to arrive organized. Dump freely. Play a card when you need the thought to move.
Think It Through is an open-source deck for developing ideas with an agent. Write the thought as it arrives, even when fragmented or overloaded. Play a card to make the next move explicit without composing another prompt.
This repository remains the stable home of this deck. New experimental Games, Decks, and Cards now develop in the central Control Decks catalogue.
I use AI before an idea is ready, so my messages start doing two jobs:
The product may be a method, an interface, and a protocol. Some of those ideas
overlap, but perhaps I am forcing them together.
Separate the ideas. Clarify each one without merging them. Preserve useful
connections, and do not choose a direction yet.
The first paragraph is the thought. The second is a response contract I kept rewriting.
DISTILL turns it into a card:
The product may be a method, an interface, and a protocol. Some of those ideas
overlap, but perhaps I am forcing them together.
/distill
The card carries that contract. Without one, the agent responds as usual.
The deck follows the way the work unfolds:
dump freely
→ DISTILL
→ DISCUSS or FURTHER
→ GRILL
→ PROPOSE
pause
→ RECAP
→ continue
Use only the steps the thought needs. Their names stay stable when you pause or resume.
Human
The product may be a method, an interface, and a protocol. Some of those ideas
overlap, but perhaps I am forcing them together.
/distill
Agent
🎯 Latest message → 🧪 DISTILL
Distilled
- A method is a repeatable way of working.
- An interface is how a person directs and observes the work.
- A protocol supplies rules that different implementations can follow.
Connections
The three ideas may describe layers of one system. The thought does not establish that they must share one product boundary. A method can prescribe a way of working while a protocol stays neutral about the method.
Response
The unresolved question is which layer people would adopt first: the method, the interface, or the protocol. The other two may support it without belonging to the same initial product boundary.
DISTILL clarified the material without choosing a direction.
Install the HACP session adapter once, then the deck. Without a compatible adapter the commands may appear, but the session is not fully HACP-conforming.
codex plugin marketplace add control-decks/human-agent-control-protocol
codex plugin add hacp@hacp
codex plugin marketplace add control-decks/think-it-through
codex plugin add think-it-through@think-it-throughUse $think-it-through:help or play
$think-it-through:distill.
claude plugin marketplace add control-decks/human-agent-control-protocol --scope user
claude plugin install hacp@hacp --scope user
claude plugin marketplace add control-decks/think-it-through --scope user
claude plugin install think-it-through@think-it-through --scope userUse /think-it-through:help or play
/think-it-through:distill.
The examples below use the portable shorthand /<card>.
Six cards cover the recurring thinking moves.
| When your thinking needs... | Play | The agent will... |
|---|---|---|
| structure | 🧪 DISTILL |
separate and clarify the thought |
| room to develop | 💬 DISCUSS |
develop implications already present |
| one new edge | 🚀 FURTHER |
add one grounded extension, then stop |
| pressure | 🔥 GRILL |
test one weak branch per exchange |
| orientation | 🗺️ RECAP |
recover a map and synthesis |
| a direction | 🧭 PROPOSE |
offer one choice with its tradeoff and risk |
DISCUSS stays inside the thought and develops implications already present.
FURTHER crosses its current edge once with a supported extension, marks the
leap, then stops. Repeat, switch, or return to normal conversation as needed.
If a long session has lost its shape, send:
/recap
RECAP uses the conversation by default. The same vocabulary reduces prompt
writing and recovery work across sessions. Cards control the requested move,
not correctness or cross-session memory; you still verify the result.
Cards pass their results from left to right:
The product may be a method, an interface, and a protocol.
/distill
+ /propose
🎯 Latest message → 🧪 DISTILL → 🧭 PROPOSE
DISTILL clarifies first. PROPOSE receives that result and selects one
direction with its tradeoff and risk.
Some cards need more than one exchange:
I think the interface should be the first product.
/grill
🎯 Current testable idea → 🔥 GRILL
Recommendation
Treat the interface as the adoption surface, but keep the method optional.
Question
If the protocol vanished tomorrow, what value would the interface still
deliver on its own?
GRILL stays active until its verdict or until you stop it. INTERVIEW uses
the same multi-exchange shape to build shared understanding.
- Keep writing as usual when you do not need a card.
- You play each card. The agent does not choose a move for you.
- Resolve combos from left to right against one Binding.
- Clear most cards after one agent response.
- Keep
INTERVIEWandGRILLactive until they finish or you stop them. - Create briefs and plans when you play their cards.
Ten advanced cards select Bindings, gather information, recommend actions, create artifacts, or change a result's presentation.
/help gives exact commands without playing a card:
/help
/help distill
/help "I need to choose a direction"
Binding cards target a conversation, topic, or axis. Output cards create a brief or plan. Presentation cards explain or render the current Working Object. For example:
/recap + /with-diagrams
🎯 Conversation → 🗺️ RECAP + 📊 DIAGRAMS
/explain + /with-diagrams
EXPLAIN preserves claims and caveats; DIAGRAMS adds one useful visual.
/to-plan prepares an accepted or provisional direction for review. It does
not authorize execution.
/explain can consume any compatible HACP result:
$work-this-way:implement
+ $think-it-through:explain
EXPLAIN presents the observed implementation result without changing it.
Think It Through controls response shape. Methods, project rules, and tools govern substance and actions.
The Human-Agent Control Protocol Draft 0.4 loads the shared session rules: Binding, Working Object transfer, control state, duration, and visible resolution. Cards stay human-only; the agent cannot infer or replay them.
| Layer | Owns |
|---|---|
| Think It Through | purpose, card-local mental model, cards, and defaults |
| HACP | Binding, object transfer, control state, order, and clearing |
| Methods and project rules | reasoning and quality constraints |
| Providers and tools | instruction loading, context, and actions |
Complete card reference
There is no deck initializer or root resolver. help explains the deck but is
not a card. Conversation → Topics → Axes appears only in the binding, recap,
and help surfaces that need it.
| Card | Play when | Default binding | Result | Duration |
|---|---|---|---|---|
| 🧪 Distill | Thoughts need structure | Latest human message | Clear thoughts | One agent turn |
| 💬 Discuss | Exploration should stay open | Current thought | Developed thought | One agent turn |
| 🔎 Interview | Understanding is missing | Smallest unclear subject | Shared understanding | Multiple exchanges |
| 🔥 Grill | An idea needs pressure | Current testable idea | Verdict and risks | Multiple exchanges |
| 🗺️ Recap | Orientation is lost | Available conversation | Map and synthesis | One agent turn |
| 🧭 Propose | An open question needs direction | Current open decision | One proposal | One agent turn |
| ⚡ Next | Action should follow | Latest actionable result | One to three actions | One agent turn |
| 🚀 Further | A useful idea should be pushed beyond its current edge | Current Working Object | One grounded creative extension | One agent turn |
| Card | Chooses | Duration |
|---|---|---|
| 🎯 Conversation | All available topics and axes | One combo |
| 🎯 Topic | One topic | One combo |
| 🎯 Axis | One axis | One combo |
| Card | Creates | Default binding |
|---|---|---|
| 📄 Brief | Portable thinking checkpoint | Available conversation |
| 📋 Plan | Execution plan for review | Accepted or provisional direction |
| Card | Presents | Default binding |
|---|---|---|
| 💡 Explain | Concise contextual explanation | Current Working Object |
| 📊 Diagrams | Smallest useful visual | Final or latest useful result |
| 🧠 Reasoning map | Supported reasoning structure | Final reasoning or current decision |
Start with an instruction you repeat:
repeated instruction
→ define one effect and default binding
→ define result, duration, and limits
→ test positions and cross-deck transfer
→ keep, revise, merge, or remove
A Draft 0.4 card contract records:
ID → Kind/Mode/Traits → Default Binding → Accepts
→ Requires (when needed) → Produces → Duration → Effect → Limits
A different deck can use the same interaction rules for another purpose and mental model. Share an instruction you repeat in a GitHub issue.
Grill Me was the seed: one short command for one reusable conversation contract. Think It Through extracted more repeated instructions into cards. I derived the HACP draft from this first implementation; one deck cannot prove a universal standard.
Think It Through is available under the MIT License.

