sequential-thinking

Sequential Thinking

Same intent as MCP-based sequential-thinking tools — externalize a numbered chain of thoughts with revision and branch semantics — implemented as visible Markdown so reasoning-capable models can follow it without an extra tool round-trip.

When to use

• The problem has 3+ interacting constraints or sub-decisions.
• An earlier conclusion may need revising as new evidence arrives.
• More than one viable approach is on the table and they should be compared, not silently chosen.
• The user explicitly asked for stepwise/structured reasoning.

If none of these hold, answer directly — protocol overhead on a simple question wastes tokens and looks performative.

Why a visible protocol (not just internal CoT)

Modern reasoning models already think internally. They need a shared on-the-page format so:

• the user can audit which step a wrong conclusion came from,
• revisions can point at the specific thought they replace,
• branches can be compared side by side instead of overwriting each other,
• a downstream agent or reviewer can parse the structure.

Do not try to suppress, expose, or paraphrase the model's hidden reasoning trace. Use this protocol in addition to it — write the durable, reviewable summary as Thought blocks; let the internal trace stay internal.

Protocol

Open with a one-line plan, then emit numbered Thought blocks. Use these exact markers so output is parseable:

**Plan:** N thoughts (estimate). Subject: <one-line problem statement>.

### Thought 1
<the step — observation, deduction, sub-decision, or question>

### Thought 2
<...>

### Thought 3 (revises 1)
<corrected version, with one sentence on why 1 was wrong>

### Branch A from 2
<alternative path>

### Thought 4 (Branch A)
<continuation of A>

### Thought 5 (main)
<continuation of the main line>

### Final
<the conclusion or recommendation. Names the winning branch if any.>

Block rules

• Numbering: monotonic. A revision gets a new number; it does not overwrite an old one. (revises N) declares which thought it supersedes.
• Branches are labelled Branch <letter> from <N>. Subsequent thoughts on that branch tag themselves (Branch <letter>). The default trunk needs no tag.
• Estimate: the opening Plan: is an estimate, not a contract. If more thoughts are needed, write ### Thought N (extending plan to M) and continue. Don't pretend you knew all along.
• One thought, one move: a thought makes one observation, one deduction, or asks one question. Walls of text under a single number defeat the point.
• Final: mandatory. Stop with ### Final. Without it, callers can't tell whether you finished or got cut off.