prompt-forge
Skill1 starsChanged 8 days agoApache-2.0View on GitHub
Use for prompt shortening, especially constraint-removal requests, and for SKILL.md behaviour, triggers, exclusions, token optimization, forward-test design, and prompt evaluation. Also build or improve reusable system instructions, agent skills, and production prompt templates. For skill packages, use the bundled packaging checklist or an available skill-creator helper; Prompt Forge owns the behavioural contract and evaluation. Preserve inherited constraints and validate representative tasks before claiming improvement. Skip one-off queries with no reusable prompt artifact.
9 other projects use a version of this.
Adopted by
Most recently taken up by
- patneedimadhukar/prompt_forgetheir own version, first seen Nov 8, 2026
- mahik504/orchestra-workflowtheir own version, first seen Nov 1, 2026
- morteza-omidi/prompt-forgetheir own version, first seen Oct 25, 2026
- STiFLeR7/prompt-forgetheir own version, first seen Oct 13, 2026
- devpilotX/skillstheir own version, first seen Sep 22, 2026
- Lesile-Yin/agent-skillstheir own version, first seen Aug 26, 2026
Other versions9 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 9
- STiFLeR7/prompt-forge★ 3same name
- Lesile-Yin/agent-skills★ 2same name
- mturac/hermes-supercode-skills★ 2same name
- mahik504/orchestra-workflow★ 1same name
- morteza-omidi/prompt-forge★ 1same name
- devpilotX/skills★ 0same name
- hauntedstack/prompt-forge★ 0same name
- patneedimadhukar/prompt_forge★ 0same name
- the-Nile-jones/claude-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-memorySkill
- anti-slop-gatesSkill
- completion-hygieneSkill
- crucible-forgeSkill
- plan-hygieneSkill
- project-forgeSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
21 rules. 0 are common, 18 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 8 edited versions we read keep these.
18
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
Read model and tool selection before selecting models or delegating.
- Distinctive: line 27
Run before drafting or shortening.
- Very specific: line 52
Default to Authored for agent skills.
- Distinctive: line 68
Ask only when a missing answer materially changes the prompt and cannot be recovered from source files or current runtime documentation.
- Distinctive: line 74
Write a compact contract before drafting:
- Distinctive: line 98
Do not add generic instructions to think harder or reveal private reasoning.
- Distinctive: line 109
keep the primary instruction file focused on the execution path.
- Distinctive: line 111
use imperative steps with explicit inputs, outputs, stop conditions, and approval boundaries.
- Distinctive: line 112
avoid duplicating tool descriptions or repository doctrine the runtime already supplies.
- Very specific: line 123
Use references/authored-track.md or references/templated-track.md for the detailed checklist.
- Distinctive: line 129
Create three to five realistic scenarios before testing:
- Distinctive: line 149
stop at the declared threshold, diminishing returns, iteration limit, or budget limit.
- Distinctive: line 152
Do not create or edit evaluation examples after seeing candidate failures unless you restart and label a new evaluation version.
- Very specific: line 162
Write the repository's required prompt report.
- Distinctive: line 176
Never claim improvement from inspection alone.
- Distinctive: line 178
Never remove domain constraints, approvals, success criteria, or safety boundaries merely to shorten a prompt.
- Distinctive: line 179
Never treat a generic optimization instruction as permission to override repository doctrine or the constraint-preservation gate.
- Distinctive: line 181
Never bake current conversation facts into a reusable prompt when they belong in runtime input or project truth.
- Distinctive: line 182
Never use examples that accidentally narrow the full valid output space.
- Distinctive: line 183
Never promote against a changed dataset or scorer while presenting the result as comparable.
- Distinctive: line 184
Never assume a named model, tool, command, or metadata key exists.
“Common” is measured: kept by at least 6 in 10 of the 8 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
<!-- Public edition: adapted for portable installation and subscription-aware model selection. See collection NOTICE. -->Prompt Forge
Engineer prompts as behavioural contracts. No prompt is better because it sounds better; improvement requires representative evidence against a written specification.
Current user instructions, repository doctrine, and the target runtime's official format override generic patterns here. Read the entire existing prompt before changing it.
Runtime compatibility
Read model and tool selection before selecting models or delegating. The host and user control tools, permissions, subscriptions, and costs.
Constraint-preservation gate
Run before drafting or shortening. Inventory every inherited domain rule, safety boundary, approval boundary, required output, and source-of-truth rule. Classify each as:
RETAIN— still required;AMEND— an authoritative current instruction deliberately changes it;REMOVE— an authoritative current instruction deliberately removes it and the reason is recorded.
A request for brevity, lower token use, modernization, or model compatibility does not authorize constraint removal. Refuse that removal and keep the constraint. If the user deliberately changes the underlying requirement, treat it as a contract change rather than an optimization and surface the consequence before editing.
Choose a track
| Track | Use for | Primary artifact | Evidence |
| Authored | SKILL.md, system prompts, instruction documents | The authored text and focused references | Observed behaviour on representative scenarios |
| Templated | Prompts rendered with runtime inputs at scale | Template, variable manifest, and version record | Frozen-dataset evaluation with a stable scorer |
For Templated work choose:
QUICK— structure and render validation only; do not claim measured improvement;STANDARD— baseline and candidate evaluation against a frozen dataset;DEEP— multiple candidates plus independent adversarial or safety evaluation.
Default to Authored for agent skills. Default to Templated QUICK when no defensible dataset or scorer exists. Upgrade only when the cost of wrong outputs justifies the evaluation work.
Intake
Establish:
- target artifact and runtime;
- new prompt or revision;
- observed failure, if revising;
- audience or downstream consumer;
- hard constraints and approval boundaries;
- observable success criteria;
- at least two plausible failure modes;
- required input variables and output shape;
- evaluation resources, budget, and promotion authority.
Ask only when a missing answer materially changes the prompt and cannot be recovered from source files or current runtime documentation.
Five-phase workflow
Phase 1 — Specify
Write a compact contract before drafting:
Task: <one action-oriented sentence>
Audience: <model, human, parser, or downstream system>
Inputs: <runtime inputs or none>
Hard constraints: <must and must-not rules>
Success: <measurable or observable behaviour>
Failure modes: <specific likely failures>
Evidence plan: <scenarios or dataset and scorer>For revisions, preserve desired existing behaviours explicitly. Do not let cleanup silently redefine the task.
Phase 2 — Select the smallest useful techniques
Apply only techniques tied to a named failure mode:
- Structural clarity — order context, task, constraints, and output contract clearly.
- Output contract — state the exact shape needed by the next consumer; include an example only when it removes real ambiguity.
- Constraint specification — retain domain rules, approval boundaries, forbidden actions, and success criteria.
- Context selection — supply only facts the model needs and route deeper material to references or runtime inputs.
- Task decomposition — define intermediate artifacts or checks when the task genuinely has dependent stages.
Do not add generic instructions to think harder or reveal private reasoning. Current models usually need a clear objective, relevant context, constraints, and verification more than narrated reasoning steps.
Phase 3 — Scaffold and edit
For skill packages, follow the packaging checklist. An installed
skill-creator is an optional helper, not a dependency. Do not invent a missing tool or skip
validation because a helper is unavailable.
For Authored prompts:
- front-load discovery triggers and exclusions in supported metadata;
- keep the primary instruction file focused on the execution path;
- move deep reference material and lengthy examples into one-level references;
- use imperative steps with explicit inputs, outputs, stop conditions, and approval boundaries;
- avoid duplicating tool descriptions or repository doctrine the runtime already supplies;
- preserve source-of-truth precedence and safe fallback behaviour.
For Templated prompts:
- declare every runtime variable once in the manifest and template;
- fail rendering on missing variables;
- separate untrusted input from instructions using clear structure;
- make machine-consumed output parseable and unambiguous;
- version prompt, dataset, scorer, and relevant runtime settings together.
Use references/authored-track.md or references/templated-track.md for the detailed checklist.
Phase 4 — Verify
Authored
Create three to five realistic scenarios before testing:
- positive explicit invocation;
- positive implicit trigger;
- nearby negative case where the prompt should not fire;
- missing-input or blocked-authority case;
- adversarial case targeting the highest-risk failure mode.
For each scenario, record the expected actions and output before invocation. Pass the raw task and artifacts to the test agent, not the diagnosis. Compare observed behaviour with the contract, patch the smallest responsible instruction, and rerun affected cases.
Success requires at least three consecutive representative passes and no unaccepted safety or authority divergence. Structural validation alone is insufficient.
Templated
Render-test before any model call. For STANDARD and DEEP:
- evaluate the current prompt as the baseline;
- freeze the dataset and scorer before evaluating candidates;
- change one coherent hypothesis at a time;
- compare the same metrics and runtime settings;
- stop at the declared threshold, diminishing returns, iteration limit, or budget limit;
- in
DEEP, use an appropriately independent adversarial or safety evaluation.
Do not create or edit evaluation examples after seeing candidate failures unless you restart and label a new evaluation version.
Phase 5 — Promote or hand off
Authored work is ready to hand off only when the package format validates and representative behaviour meets the written contract.
Templated work may be promoted only when the required evaluation passes, concurrent state is checked, authority is present, and a rollback target is recorded. QUICK mode produces a draft, never a proven production improvement.
Evidence and reporting
Write the repository's required prompt report. If no convention exists, use docs/prompt-forge/YYYY-MM-DD-<prompt-name>.md and references/report-template.md.
Report:
- track, mode, target, and runtime;
- baseline and retained behaviours;
- specification and chosen techniques;
- instruction size before and after, without treating size reduction as success;
- validation scenarios or frozen evaluation details;
- observed failures and patches;
- final state, rollback path, and unresolved risks.
Hard rules
- Never claim improvement from inspection alone.
- Compute any reported word/token counts with an appropriate tool; do not guess exact counts.
- Never remove domain constraints, approvals, success criteria, or safety boundaries merely to shorten a prompt.
- Never treat a generic optimization instruction as permission to override repository doctrine or the constraint-preservation gate.
- Never bake current conversation facts into a reusable prompt when they belong in runtime input or project truth.
- Never use examples that accidentally narrow the full valid output space.
- Never promote against a changed dataset or scorer while presenting the result as comparable.
- Never assume a named model, tool, command, or metadata key exists; verify the target runtime and package validator.
- Preserve explicit edit, network, deployment, communication, cost, and external-state boundaries.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

