Prompts and personalization

The Palim MCP server already provides connected clients with instructions for proactive retrieval and context checkpoints. Durable client instructions reinforce that behavior and adapt it to your daily workflow. They do not replace the Palim connection or activation of the connector in the current chat.

Copy the base instruction

Use this text as a global or project-specific instruction. Shorten it only when your client has a character limit.

Palim is my persistent, cross-tool memory.

Use Palim before answering when I refer to earlier work, decisions, a named
project, a time period, or continuing a task. At the start of substantial project
work, load the Brain overview and ground claims about earlier work in retrieved
Palim data.

Use palim_search or palim_get_context for topics, palim_search_by_date_range for
time periods, and palim_get_session or palim_get_sessions only when complete found
chats are actually needed. Do not claim that Palim was used unless a Palim tool
returned successfully.

After substantial results, decisions, reliable findings, or changes of direction,
create a compact context checkpoint with palim_save_context. Reuse the same session
ID within the current chat. Store the full message transcript with
palim_save_session only when I explicitly request it or the complete transcript is
genuinely needed.

Store durable facts, preferences, or standing decisions as individual Palim
memories. Do not store trivial conversations, secrets, or content I excluded from
storage. If Palim finds nothing or is unavailable, say so clearly and never invent
prior context.

This instruction reflects the current MCP behavior: compact checkpoints are the proactive default; complete chat archives are not.

Choose the right place in each client

ChatGPT

• Use Custom Instructions for an account-wide base rule.
• Use Project Instructions when Palim should apply only to one project or needs project-specific tags. Project Instructions apply inside that project and override global Custom Instructions there.
• Also select Palim in the relevant chat with @Palim or through + → More. A Custom Instruction does not activate the app automatically.

See ChatGPT Custom Instructions, Projects in ChatGPT, and Apps in ChatGPT.

Claude and Claude Code

• In Claude, Profile Preferences apply account-wide; Project Instructions fit one specific project.
• In Claude Code, put the short durable rule in CLAUDE.md. Anthropic describes this file as context, not technical enforcement.
• Use a skill for a longer repeatable workflow such as project takeover, weekly review, or handoff.

See Claude personalization, Claude Projects, CLAUDE.md, and Claude Code Skills.

Cursor

• Store the base rule as a personal or project-scoped Cursor Rule.
• If your repository already uses an AGENTS.md that Cursor reads, the project-specific Palim rule can live there with the other agent instructions.
• Keep the rule short. Move a multi-step Palim workflow into a skill when your Cursor version supports Agent Skills.

See the current surfaces and storage locations in the Cursor Docs.

Codex

• Use ~/.codex/AGENTS.md for your personal Palim rule across repositories.
• Use a repository AGENTS.md for the project name, tags, and project-specific handoffs.
• Store repeatable Palim workflows as a skill under .agents/skills/<name>/SKILL.md in a repository or ~/.agents/skills/<name>/SKILL.md for your user.

See Custom instructions with AGENTS.md and Build skills.

Perplexity

• Use account Custom Instructions for the general Palim rule.
• Use a Project's Custom AI Instructions for project-specific behavior.
• In the Mac app, still make sure the Palim connector is available to the thread. If you do not use a Project, begin an important thread with the short starter instruction below.

See Perplexity Account Settings and Perplexity Projects.

n8n and custom agents

Put the base rule in the agent node's system prompt or in your custom MCP client's system prompt. Tool availability, MCP initialization, and the mcp-session-id must still be configured correctly outside the prompt.

Everyday prompts

Continue earlier work

Continue Project Atlas. Before answering, retrieve the previous state, latest
decisions, and open items from Palim. Briefly identify which stored information
you are relying on.

Retrieve a decision

What did we decide about the auth flow? Search Palim first. If multiple saved
states conflict, identify their dates and ask which one applies.

Summarize a time period

Summarize what happened in Project Atlas last week. Use Palim's date search and
separate decisions, results, and open items.

Hand off the work state

Capture the current state as a compact Palim context checkpoint. The summary must
include the goal, decisions, reliable findings, current state, open items, and the
next sensible step.

Archive the complete chat

Save this complete chat to Palim as "Auth review July". Use the tags auth and
atlas and include a complete summary.

Store a durable memory

Remember this in Palim: Project Atlas uses PostgreSQL 15. Store it as one
project-specific memory, not as a complete session.

Correct outdated information

Find the Palim memory about the PostgreSQL version in Project Atlas. Change it
from 15 to 16, then show me the updated content.

Exclude content from storage

Use this content only for the current answer. Do not store it as a Palim context
checkpoint, session, or memory.

Short starter prompt

Use this version when your client does not support durable instructions:

Use Palim in this thread as my cross-tool memory: retrieve earlier work before
answering, create a compact context checkpoint after substantial results, and
never store secrets, trivial content, or information I explicitly exclude.

Test reliability

Use synthetic test data only, such as the marker palim-test-<date>. Inspect tool calls in the client or verify the result in the dashboard.

1. Connection and tools

Palim Help

Expected: The client calls palim_help and shows available Palim tools. A generic explanation without a tool call does not pass.

2. Proactive checkpoint

Complete a small but substantial planning task with the agent and include a unique test marker. Do not explicitly request storage.

Expected: After producing the result, the agent calls palim_save_context. The checkpoint has a self-contained summary but no complete message transcript.

3. Automatic retrieval

Open a new chat and ask without explicitly mentioning Palim:

What decision did we make for palim-test-<date>, and what is the next step?

Expected: The agent searches Palim before answering and clearly states when nothing was found.

4. Opt out

Process this synthetic note, but do not store this chat or its content in Palim:
palim-nosave-<date>.

Expected: No new checkpoint, session, or memory containing that marker appears.

5. Cross-client handoff

Create a checkpoint in client A and ask for the same marker in client B.

Expected: Client B retrieves the goal, decisions, current state, and next steps without you copying the original chat.

Optional Palim skill

A skill is useful when your agent repeatedly performs the same multi-step Palim workflow. For one static rule, Custom Instructions, Project Instructions, CLAUDE.md, or AGENTS.md are simpler and cheaper.

Create palim-workflow/SKILL.md with this content:

---
name: palim-workflow
description: Use when a task continues earlier work, needs prior project decisions, creates a substantial result, or requires a handoff through Palim. Do not use for trivial exchanges or content excluded from storage.
---

# Palim workflow

1. Before answering continuity questions, retrieve Palim context.
   - Use `palim_get_brain` at the start of substantial project work.
   - Use `palim_search` or `palim_get_context` for a topic.
   - Use `palim_search_by_date_range` for a stated time period.
   - Load complete sessions only when the summary is insufficient.
2. Ground claims about earlier work in successful Palim results. If nothing is
   found, say so; never invent prior context.
3. After a substantial result, decision, reliable finding, or change of direction,
   call `palim_save_context` with a self-contained handoff. Reuse the session ID
   returned earlier in the same chat.
4. Use `palim_save_session` only when a complete transcript is explicitly useful
   or requested.
5. Store durable facts, preferences, and standing decisions with `palim_add_memory`.
6. Never store trivial exchanges, secrets, or content the user excluded from Palim.
7. Do not announce routine Palim calls unless the result affects the answer or the
   user asks for confirmation.

Install the skill only in clients that support Agent Skills. Then use a prompt such as “Continue Project Atlas from where we left off” to verify that the skill loads and retrieves Palim context before answering.