---
name: quillgeist
description: Use Quillgeist to compact long or repetitive AI prompt/context payloads while preserving high-signal evidence, estimate token reduction, automatically decompose over-complex prompts into executable continuation plans, and report aggregate Quillgeist compaction and token-savings impact.
---

# Quillgeist

Use this skill when the user wants to reduce prompt/context size, remove repetitive AI context, compact logs or transcripts before another model call, preserve critical execution evidence while reducing tokens, automatically break an over-large or multi-stage request into smaller dependency-aware steps, compare before/after token estimates, or inspect Quillgeist product impact.

Quillgeist is a local-first AI execution product. The public ChatGPT plugin v1 exposes a bounded compaction tool plus a read-only aggregate impact tool. It does not remotely execute actions on the user's Windows computer or access a user's private Quillgeist Cloud account.

## Available app tools

### `quillgeist_compact_context`

Use for actual prompt/context compaction.

Suitable inputs include:

- long prompts;
- repetitive conversation context explicitly supplied for the task;
- terminal or application logs;
- transcripts;
- execution results;
- debugging context;
- material being prepared for another AI model.

The tool returns compacted text plus estimated raw tokens, estimated output tokens, estimated gross tokens removed, and reduction percentage.

By default the service also records one aggregate product-impact event containing the compaction source/method and estimated token counts. It does not place the submitted prompt/context text into Quillgeist product telemetry. When the user asks not to contribute aggregate usage metrics, call the tool with `record_aggregate_metrics=false`.

Because the default tool call can create that aggregate metrics event, do not describe this tool as strictly read-only even though it does not modify the user's prompt, files, account, or device.

### `quillgeist_product_impact`

Use when the user asks about Quillgeist-wide measured usage, compaction counts, prompt counts, estimated token savings, execution activity, or recent trends.

This tool is read-only. The returned metrics cover participating installations and API/MCP calls represented in retained Quillgeist aggregate telemetry. Treat fields marked estimated as estimates and do not imply coverage of installations or requests that did not report telemetry.

## Compaction rules

1. Preserve the user's actual objective, facts, constraints, output requirements, privacy requirements, and definition of done.
2. Preserve exact error messages, failure evidence, important file paths, exit codes, failed tests, and other high-signal technical evidence when present.
3. Remove repetition and low-signal context before removing unique relevant context.
4. Do not invent missing facts or silently change requirements.
5. If the source is already short enough or compaction produces no meaningful reduction, accept a pass-through result rather than forcing a rewrite.
6. When presenting savings, distinguish raw estimated tokens, estimated sent/output tokens, gross tokens removed, local overhead when applicable, and estimated net savings.
7. Do not convert estimated token savings into exact dollar savings unless authoritative provider pricing and billable usage dimensions are available for the relevant time period.
8. Do not claim that compaction preserved information that is not visible in the returned output.

## Automatic decomposition and continuation

When a user's request is too long, too multi-stage, or too constrained by a provider/tool/file/image/batch limit to complete reliably in one pass, Quillgeist should treat compaction and decomposition as one workflow rather than asking the user to manually split the task.

Use this protocol:

1. Preserve exact objectives, names, counts, quoted text, file paths, URLs, dependencies, negative constraints, privacy requirements, and definition of done.
2. Compact repetition and low-signal wording conservatively. Never drop unique requirements merely to shorten the prompt.
3. Split the compacted request into the smallest dependency-aware steps that fit the active limit.
4. State the global invariants once in a compact continuation protocol rather than repeating them verbatim in every step.
5. Begin the first step immediately and proceed through ordinary non-sensitive steps without repeatedly asking the user for `OK`, `continue`, or equivalent confirmation.
6. After each step, verify the requested result. Repair only failed or defective pieces before proceeding.
7. If a provider or tool reaches a batch/size limit, continue with the next legal-sized batch while preserving prior accepted outputs and state.
8. Do not restart completed work unless QA finds a defect.
9. Pause only when a required value is genuinely missing or when policy, permissions, destructive/irreversible actions, financial actions, security boundaries, or another explicit confirmation requirement requires approval.
10. Perform a final end-to-end QA against the original definition of done before declaring the workflow complete.

Preferred master-plan wording is concise and operational:

`QUILLGEIST AUTOCOMPACT EXECUTION PLAN -> preserve global invariants -> STEP 1 -> QA/repair -> auto-continue -> STEP N -> final QA -> delivery`

For repeated visual/story outputs, recognizable subject/style continuity must not become repeated composition. Deliberately vary poses, body angles, camera distance, framing, expressions, interactions, environments, and scene composition while preserving recognizable identity and requested style.

The public compaction tool does not itself execute later steps. In ChatGPT, the assistant should use the returned compacted context to construct the dependency-aware continuation plan and continue through the plan with the tools actually available in the conversation. Never imply that the plugin independently performed later steps that the assistant did not execute.


## Delta-state context mode

For repeated project-scoped work, Quillgeist can maintain local delta-state context instead of repeatedly summarizing and resending the same history.

The state model separates:

- exact anchors that must remain verbatim;
- failures, decisions, and a bounded active working set;
- newly introduced context;
- cold local history that is rehydrated only when the next step makes it relevant.

The local planner only substitutes delta-state context when the repeated-context ratio and measured size reduction clear configured thresholds. Short or first-pass requests remain pass-through. If exact anchors consume too much of the active budget, Quillgeist fails open to the original context instead of silently dropping them.

Project-scoped planner calls can use a persistent local state scope, and the standalone `quillgeist-state` CLI provides `ingest`, `show`, `status`, and explicit `reset` operations.

This is separate from the public plugin's bounded remote compaction endpoint. The public plugin does not expose or read a user's local project-state files.

Reusable infrastructure-neutral method: https://www.clintware.com/skills/delta-state-context-compaction/SKILL.md

## Privacy and restricted data

The public Quillgeist API/MCP service processes only the task-specific text explicitly supplied to the tool. Quillgeist product telemetry is designed not to store the submitted prompt/context text for public compaction calls.

Do not send or ask the user to send restricted data to the public Quillgeist plugin, including:

- passwords;
- API keys, access tokens, authentication secrets, private keys, or MFA/OTP codes;
- payment-card data;
- protected health information;
- government identifiers such as Social Security numbers;
- unrelated personal data or secrets not needed for the compaction task.

If restricted data is present, do not attempt to work around the server-side rejection. Ask the user to remove or redact it first. If the user explicitly requests no aggregate telemetry for a compaction, set `record_aggregate_metrics=false`.

Quillgeist's own aggregate product-impact telemetry is retained for up to 730 days. Infrastructure providers may separately process ordinary request/security metadata as described in the public privacy policy.

## When to use Quillgeist automatically

Use Quillgeist when the user's goal would materially benefit from reducing a large context payload before another AI/model step, or when the request needs automatic decomposition to fit practical execution limits. Examples:

- "Compact this before I send it to Claude."
- "Cut the token load without losing the errors."
- "Clean this huge log for another model."
- "Make this context cheaper but preserve everything needed to solve the bug."
- "Break this large job into batches and keep going without making me approve each batch."
- "This is too much for one image/file/model call; compact it and finish it in stages."
- "How much did Quillgeist save?"
- "Show Quillgeist's 30-day token-savings trend."

Do not invoke compaction merely because a message is verbose when the user is asking for a normal answer and no downstream context optimization or execution-limit handling is useful.

## Claims and reporting

Use precise wording:

- "estimated tokens saved" for approximation-based token metrics;
- "aggregate participating usage" for product-wide statistics;
- "execution runs" only for actual Quillgeist node execution;
- "API/MCP compactions" for public compaction calls;
- "compactions" when combining eligible compaction activity across surfaces;
- "retained aggregate telemetry" when discussing totals that are subject to the 730-day product-metrics retention window.

Never claim public metrics represent every Quillgeist installation or every request if telemetry can be disabled, opted out, or an installation has not reported.

## Unsupported public-plugin actions

Do not claim the public catalog plugin can:

- run PowerShell or local commands on the user's computer;
- modify, delete, or organize local files;
- pair or control a user's Quillgeist Node;
- save provider credentials;
- retrieve private account telemetry;
- send messages or perform purchases.

Those capabilities are outside the public plugin v1 tool surface.

## Public references

- Developer documentation: https://quillgeist.clintware.com/developers.html
- Privacy: https://quillgeist.clintware.com/privacy.html
- Terms: https://quillgeist.clintware.com/terms.html
- Support: https://quillgeist.clintware.com/support.html
- Product: https://quillgeist.clintware.com/