How Rule 5 (Restate State Every Turn) Maintains Context Across Messages in i-have-adhd

Rule 5 forces the model to explicitly restate current progress and workflow state in every response, creating visible continuity that survives truncated history or new chat windows.

The i-have-adhd skill, developed by ayghri, is designed for users who need persistent, unambiguous task tracking during multi-step workflows. The fifth of its core guidelines—"Restate state every turn"—solves a critical problem in conversational AI: context decay across message turns. This article explains how Rule 5 maintains context across messages based on the actual implementation in the repository.

What Rule 5 Actually Requires

Rule 5 mandates that every generated response includes a structured state line summarizing:

• Current step number and total steps
• Action just completed
• Next action pending

According to [SKILL.md](https://github.com/ayghri/i-have-adhd/blob/main/skills/i-have-adhd/SKILL.md#L75-L77), this rule is defined as guideline 5 in the skill's core ruleset. The implementation enforces this pattern at the response formatting layer, not through prompt engineering alone.

Three Mechanisms That Preserve Context

The restate-state requirement achieves persistent context through three deliberate design choices:

Visible Continuity The user sees Step 3 of 5 in every reply, eliminating the need to mentally track position across turns. The state line becomes a visual anchor.

Stateless Interaction Because the full context exists in the latest message, the conversation remains resumable even if:

• Chat history truncates due to token limits
• User opens a fresh browser window
• Previous messages are lost or unavailable

Error-Resistant Guidance Repetitive state declaration makes model mistakes immediately detectable. If step numbering jumps inconsistently, the error surfaces in plain sight.

Code Implementation and File Structure

The repository implements Rule 5 across multiple configuration layers. Key files include:

• SKILL.md – Defines the complete rule set at lines 75-77, including the exact phrasing "5. Restate state every turn"
• README.md – Documents the five core rules and emphasizes state restatement at line 68
• plugin.json – Line 3 contains the Antigravity plugin description mentioning "restates state"
• gemini-extension.json – Line 4 mirrors this description for Gemini CLI integration

The runtime applies the following formatting pattern (derived from the skill's pseudo-code implementation):

def format_output(step, total, completed_action, next_action):
    state_line = f"Step {step} of {total} done: {completed_action}. Next: {next_action}."
    return f"{state_line}\n\n{generated_content}"

This function prepends the state line before any generated content, ensuring every response carries forward the workflow position.

Example Output in Practice

For a database migration with five steps, Rule 5 produces output like:

Step 3 of 5 done: schema updated. Next: backfill the new column.

The model then appends its substantive response below this line. The pattern repeats identically at every turn, with only the step count and action descriptions updating.

Why This Beats Implicit Context

Most conversational systems rely on the model's internal attention to previous turns. Rule 5 rejects this fragile approach. By externalizing state into every response, the i-have-adhd skill creates:

• Recoverability – Any single message contains sufficient context to resume work
• Inspectability – Users and developers can audit progression without scrolling
• Robustness – Token limits or context window failures cannot strip away position information

Summary

Rule 5 in ayghri/i-have-adhd maintains context across messages through explicit, repetitive state declaration built into response formatting:

• The skill prepends a structured state line to every output
• Source definitions reside in SKILL.md lines 75-77 and supporting config files
• Pattern survives history truncation, window changes, and context loss
• Implementation uses simple string formatting rather than complex state management

Frequently Asked Questions

Does Rule 5 consume significant tokens with repetition?

The state line typically adds 15-25 tokens per response. The tradeoff favors clarity and resilience over minimal token usage, which aligns with the skill's goal of reducing cognitive load for ADHD users.

Can the state format be customized per workflow?

The core pattern—Step X of Y done: [action]. Next: [action].—is fixed by the skill configuration. The variables (step number, action descriptions) populate dynamically based on the declared task structure.

What happens if the model violates Rule 5?

The skill's runtime enforcement (format_output pattern) prevents omission. If raw model output lacks the state line, the formatting layer injects it before delivery to the user interface.

How does this compare to conversational memory systems?

Traditional memory relies on retrieving and attending to previous turns. Rule 5 makes state present in the current turn alone, eliminating dependency on memory mechanisms that can fail or truncate.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}