Appendix

3 min read · 641 words · 6 sections

Tool names and product behavior in this appendix follow the official documentation and the case-study repository.[1][2][3]

A. Glossary

Term	One-line explanation
Agent	The working assistant: model + harness + current context
Harness	The Agent's workbench: tools, sessions, permissions, plugins
Model	The brain: understanding, reasoning, generation
Session	One continuous conversation
Workspace	The folder the Agent works in
Context Window	The maximum text the model can process at once
Compaction	Compressing context when it fills up
Tool	An action the Agent can take
Skill	Knowledge/procedure for how to do something
Preset	A capability composition for one Agent
Plugin	A runtime capability extension
Bundle	An installable package of plugins/configuration
Profile	An installation environment
Sandbox	A mechanism limiting file/command access
Approval	Confirmation before a sensitive action
MCP	A protocol for connecting external tools/data
PTC	Programmatic Tool Calling
Creator	The Preset for extending DSH (internal id `cordis`)

B. Ten Prompt Templates

1. Read-only inventory

Read the current workspace and list files grouped by type.
Do not modify anything or run commands that change files.
At the end, tell me which tools you used.

2. Plan first

Do not execute yet. Give me a 3–5 step plan:
what each step does, which files it reads/writes, and how to verify.
Wait for my confirmation.

3. Generate a document

Read [files] and produce [output].
Requirements: [structure/length/language/links].
Do not modify the originals.
After writing, verify the output exists and list the files you read.

4. Code review

Review this change for bugs, edge cases, and missing tests.
Order findings by severity, with file and line references.
Do not edit code yet; report only.

5. Debugging

Reproduce the problem, then propose at most three hypotheses.
Test one hypothesis at a time and show evidence.
Do not make broad changes without verification.

6. Research

Research [topic], preferring primary sources.
Output: conclusions, source links, counterarguments, and open questions.
Do not invent sources.

7. Data cleanup

Read [input] and convert it to [output format] using [fields].
Do not modify the source data.
Give me the field mapping and cleaning rules first.

8. Writing

Write a [length] piece on [topic] for [audience], in [tone].
Avoid generic filler; support each claim.
Give me an outline first.

9. File organization

List the files in [directory] read-only and propose an organization plan.
Do not delete or move anything until I confirm.
Explain where each file goes and why.

10. Plugin request

I want to add this capability to DSH: [description].
Do not write code yet. Check:
1. Which layer it belongs to;
2. Which APIs/templates are available;
3. What the minimal version is;
4. How to verify and uninstall.

C. Before-You-Start Checklist

[ ] Is the goal clear in one sentence?
[ ] Are the input files/folders known?
[ ] Is the output location and format known?
[ ] What is forbidden?
[ ] How will we verify success?
[ ] Do we need a backup or commit?
[ ] Can we start read-only?
[ ] Does it involve secrets or private data?
[ ] Does it require a plugin or elevated permission?

D. Task-Completion Checklist

[ ] Does the output file exist?
[ ] Does it meet the requirements?
[ ] Were the originals modified?
[ ] Was any input missed?
[ ] Were any facts or sources invented?
[ ] Is there test/verification evidence?
[ ] Was the final deliverable presented?
[ ] Were temporary files cleaned up?
[ ] Do we know how to roll back?

Closing

DSH is not hard to use. The hard part is learning to control it before it becomes powerful.

Remember three sentences:

Read-only first, then write.
Start small, then expand.
Secrets never go into chat; back up before destructive actions.

Once those are habits, Presets, Plugins, Bundles, and Creator stop being a pile of nouns. They become a workbench you can compose on demand.

References

All links are public sources; product behavior may change between versions.