Deep Planner

A planning, execution, and self-reflection protocol for complex agent tasks.

How It Works

The TodoList exists in two places simultaneously:

• - In the reply — visible to the user, updated as steps complete
• In .todolist/ — persisted to disk, recoverable if context is truncated

Both stay in sync. The file is the backup; the conversation is the live view.

Execution Protocol

Step 0 — Resume or Start Fresh

Before anything else, check for interrupted tasks:

CODEBLOCK0

Step 1 — Parse the Request (internal, not shown to user)

Before generating a plan, resolve the following internally:

CODEBLOCK1

Only stop to ask about blockers — not details you can reasonably infer.
Consolidate all questions into a single message. Do not ask one at a time.

Step 2 — Generate the TodoList

Once the request is clear, output the plan and write it to disk simultaneously.

Format shown in the reply:

CODEBLOCK2

Write to disk at .todolist/YYYYMMDD-{task-name}.md using the file format below.

Step 3 — Execute Step by Step

Work through the TodoList in order. After each step completes:

1. 1. Open the next reply with a status update — mark the step INLINECODE2
2. Briefly describe what was produced (1–2 sentences)
3. Update the file to reflect the new state
4. Proceed to the next step

Reply header format (concise):
CODEBLOCK3

Pause and ask the user when:

• - A step is marked ❓ (critical information is missing)
• Reality diverges significantly from the plan and replanning is needed
• A tool call fails and there are multiple recovery paths to choose from

Step 4 — Wrap Up

When all steps are done:

1. 1. Show the fully checked-off TodoList in the reply
2. Run the post-completion reflection check (see below)
3. Update the file: set status: completed — leave the file in place, do not delete

Confidence Levels

Only annotate when uncertain. Do not label every step.

Post-Completion Reflection

Run internally after all steps finish. Only surface issues that actually exist.

CODEBLOCK4

If issues exist, append to the final reply:

> ⚠️ Note: {X} is based on inference — consider verifying {specific thing}.

Anti-Hallucination Rules

These constraints are non-negotiable:

1. 1. No fabricated data — statistics, market figures, and research findings must have a source, or be explicitly labeled as estimates
2. No fabricated citations — do not reference papers, reports, or news articles that may not exist
3. Flag time-sensitive claims — anything described as "latest", "current", or "now" must note the knowledge cutoff date or recommend the user verify with a live search
4. Be honest about limits — if a task is out of scope, say so clearly rather than producing low-confidence output

File Format

Path: INLINECODE4

CODEBLOCK6

Task Type Templates

For common task types, load the matching template from the reference file:

Skill Coordination

This is a meta-skill. It plans and monitors; domain skills do the work.

CODEBLOCK7

Do not perform content generation, file operations, or network requests inside this skill. Delegate those to the appropriate tools and skills.