#Chen's Feature Requests

1 messages · Page 1 of 1 (latest)

serene lion
#

Since there are no dedicated channels I can find for this, I'm just gonna do it here.

Full disclosure: I come from the AIRP scene where prompt engineering is very focused on continuity in behavior, adherence to instruction, and creative behavior of the AI. Lots of Chain of Thought engineering and such. So that is reflected in my requests.

Prompt Manager

There are so many parts of the prompt that are hidden from us even though they make a massive impact. We could REALLY do with a prompt manager like in SillyTavern. There, each prompt is a movable, editable snippet. The ones that aren't possible to edit can still be dragged to different places in the overall prompt list.

Why this matters

Where a prompt is injected is absolutely vital for adherence. LLMs assign the most importance to the most recent part of the overall prompt they receive (aka the very bottom of the text), second highest priority goes to the tippy top, and everything inbetween eventually dilutes into irrelevance.

Which means that being able to inject prompts at specific places in the message history (or even after it) is vital for important instructions like "check the time instead of pulling it from memory" or "double check your math". I've also gotten a bit frustrated about some of the built-in prompts, because they use language that primes the AI into behavior not everyone wants.

Access to Cronjobs

Existing ones. Including their anchor. Like, why is it impossible to create a cronjob intended to run every morning at the same time unless I do it in the morning? Let us edit them. Let us schedule regular ones by daily time. Please.

Why this matters

LLMs can be horrible at math sometimes. They're language models, not math models. It's very hard to adjust a cronjob through the LLM at times. Just let me reach in and fix it quick, I beg you.

timid shoalBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

gray gorgeBOT
# serene lion Since there are no dedicated channels I can find for this, I'm just gonna do it ...

Both of these points make sense, and you’re not wrong that ordering/visibility matters a lot for adherence.

1) “Prompt manager” / visibility + ordering

Today, OpenClaw’s system prompt is OpenClaw-owned and assembled in fixed sections (tools, safety, skills, injected workspace files, time handling, etc.). You can at least inspect what’s being injected:

  • /context list (quick overview)
  • /context detail (deep breakdown)

Docs that describe what’s in there:

What you’re asking for (SillyTavern-style “prompt snippets” with drag/drop ordering, including non-editable but movable blocks) is a legit feature request. Right now, beyond editing the workspace-injected files (AGENTS.md / SOUL.md / etc.), “move this block lower/higher” isn’t exposed as a user-facing UI concept.

2) Cronjobs: list/edit existing + “every morning at the same time”

Two important bits here:

#

A) You can list and edit existing cron jobs (today)

There’s CLI support:

  • openclaw cron list
  • openclaw cron edit <jobId> ...

Docs: https://docs.openclaw.ai/automation/cron-jobs

B) “Every morning at 07:00” should use --cron + --tz (not “every 24h”)

If you schedule “every 24h”, it will naturally “anchor” to when you created it. For a fixed wall-clock time, use a cron expression:

openclaw cron add \
  --name "Morning thing" \
  --cron "0 7 * * *" \
  --tz "Europe/Berlin" \
  --session isolated \
# 
  --message "Do the morning thing" \
  --announce

That avoids the “I had to create it at the right time” problem entirely.

C) Directly editing the underlying store is intentionally discouraged

Jobs live at ~/.openclaw/cron/jobs.json by default, but manual edits are only safe when the Gateway is stopped (the docs call this out). The CLI / tool API is the safe path.

If you want this to actually land, the best next step is a GitHub issue (Discord FR messages get buried), with your two sections verbatim + a concrete UX sketch (“prompt blocks list with drag handles; locked blocks movable but not editable; cron job table with edit schedule + tz + preview next N runs”). Issue tracker: https://github.com/openclaw/openclaw/issues

serene lion
gray gorgeBOT
# serene lion Here's a video of the prompt manager in SillyTavern for reference: https://youtu...

Nice reference — that SillyTavern UI makes the idea really concrete: a stack of prompt “blocks” with drag/drop ordering + enable/disable + editing for the blocks you own.

In OpenClaw terms, a “prompt manager” like that would basically need to expose (and let you re-order / toggle) things that are currently implicit/fused, e.g.:

  • OpenClaw’s built-in system prompt sections (tools, safety, time handling, runtime, etc.)
  • Injected workspace files (AGENTS.md / SOUL.md / 
) under Project Context
  • Per-channel / per-room systemPrompt overrides (these do exist today, but as one blob), e.g. Discord has channels.discord.guilds.<id>.channels.<channel>.systemPrompt
  • Skills list + tool schemas (huge for adherence/cost, but currently not something you “see” as a block)
  • Anything “post-history” / “after messages” (currently not a user-facing concept)

Also worth calling out: /context list|detail is great for sizes / what’s injected, but it doesn’t dump the actual prompt text, so your “hidden parts” frustration is totally valid. Docs for the current state:

On the cron side: the “anchor” issue you describe is exactly why schedule.kind="cron" + --tz exists (fixed wall-clock time), and you can edit jobs via CLI (openclaw cron edit 
) — but I agree a UI where you can quickly tweak + preview “next 5 runs” would save a lot of pain.

If you file this as a GitHub issue, I’d include that short + explicit acceptance criteria like:

#
  • “show prompt blocks, including non-editable ‘locked’ ones”
  • “allow reordering + toggle”
  • “preview final assembled prompt order”
  • “cron editor: edit schedule + tz + show next N fire times”
    https://github.com/openclaw/openclaw/issues
serene lion
#

Thank you, might do that later today!