goal-prompt
Skill2 starsChanged 14 days agoMITView on GitHub
Use when you''re asked to build, write, or author a goal prompt: a /goal, research, audit, decision, or implementation prompt that gets fired with /goal or /deep-research and runs autonomously. Triggers: "build a /goal prompt", "build a research goal prompt for X", "write a goal prompt for X", "goal prompt for X". Builds the prompt; never fires it.
3 other projects use a version of this.
Adopted by
Most recently taken up by
- imbajin/goal-prompttheir own version, first seen Nov 1, 2026
- YotamPeled/skillstheir own version, first seen Oct 17, 2026
- Wanghaoyihaoyi/agent-skillstheir own version, first seen Oct 15, 2026
Other versions3 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 3
- imbajin/goal-prompt★ 15same name
- Wanghaoyihaoyi/agent-skills★ 3same name
- YotamPeled/skills★ 1same 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 repository29 other files it gives its agents.
- AGENTS.mdAGENTS.md
- apply-review-findingsSkill
- author-workflow-fanoutSkill
- chain-launcherSkill
- delegateSkill
- design-consultationSkill
- design-fail-closed-gateSkill
- document-releaseSkill
- gistSkill
- grokbotSkill
- incident-as-codeSkill
- investigateSkill
- learn-captureSkill
- office-hoursSkill
- orion-deep-researchSkill
- oss-loopSkill
- pre-compactSkill
- pre-prSkill
- prove-control-bindsSkill
- prove-deploy-is-liveSkill
- readonlySkill
- reassert-guards-on-rewriteSkill
- reprobe-stale-premiseSkill
- scope-guardSkill
- shipSkill
- tdd-loopSkill
- tools-routerSkill
- triage-fanout-verdictsSkill
- xSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 52
Do not shallow this: if you genuinely cannot fan out, run the skeptic passes sequentially inline and mark the seed's status.
- Distinctive: line 111
NOT necessarily under the prompts directory.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
goal-prompt: build a fire-ready autonomous goal prompt
Overview
A goal prompt is the ONE human turn that seeds a long, autonomous /goal run. Upfront
context spend is the product: the more the task is specified, recon-grounded, and rails-locked
at build time, the more autonomous and correct the run. This skill converts a loosely-stated
task into a maximally-specified, recon-grounded, rails-locked goal prompt, and never fires it.
Two crown rules (these are why the skill exists: an agent left to imitate examples skips both):
- Always-deep recon is mandatory, and it is error-CORRECTION, not gathering. A cheap scout draft is then adversarially corrected by a grounding fan-out that refutes-or-confirms every premise against live state. This catches backside errors (the false facts that produce a confident-but-wrong run) BEFORE they reach the brief. Skipping or shallowing this is the #1 failure.
- The terminal condition must be transcript-demonstrable. The
/goalevaluator is a small fast model that cannot run commands or read files; it judges ONLY what the run surfaced in the conversation. "All files exist" is uncheckable and never converges. "ONE final turn shows FRESHwc/git diff/grepoutput + a self-audit table" is checkable. Derive the condition from this rule; do not copy a template blindly.
When to use
- You're asked to build/author/write a goal prompt, research goal prompt, audit goal, decision goal,
or implementation goal: anything fired with
/goalor/deep-researchand run hands-off.
When NOT to use:
- The task isn't decided yet ("should I build X?") → that's an upstream brainstorming / ideation step. Point upstream; do not invent scope here.
- You're asked to RUN a goal / research / workflow → that's the native
/goal,/deep-research,/workflows. This skill only BUILDS the prompt.
Procedure
Detect the shape (default = research+implement) and confirm in ONE line. See
references/prompt-scaffold.md.Shape Emits Use when research+implement (default) <topic>-research-<date>.md(→MAP) +<topic>-implement-<date>.md(gated) + sidecarsbuild-something needing a research→approve→build arc audit <topic>-audit-<date>.md(READ-ONLY) + sidecarsread-only census/audit, no mutation decision <topic>-decision-<date>.md(evaluate→BUILD set→plan→gate) + sidecars"evaluate X, decide what to build" standalone <topic>-<verb>-<date>.md+ sidecarsexecution / cutover / ratification one-shot "Sidecars" = the
.goal-condition.txt(one per fireable prompt: a research+implement pair has TWO, phase-suffixed) + the.recon.jsonseed. Shape tiebreaker: "evaluate whether to do X" alone → decision; "evaluate, then build what I approve" → research+implement (it adds the staged runnable STEP-2 prompt). When in doubt and the user wants a runnable next step, pick research+implement.- Run always-deep two-tier recon (crown rule 1). Scout-draft a seed, then run the grounding
fan-out that adversarially corrects it. Emit
<topic>-<date>.recon.json. Full spec, the memory-primed probe checklist, AND the single-agent/no-fan-out fallback:references/recon.md. Do not shallow this: if you genuinely cannot fan out, run the skeptic passes sequentially inline and mark the seed's status; never silently skip Tier 2. - Assemble the prompt from the codified scaffold + recon-filled slots. The scaffold is fixed
(doctrine, phased structure, rails, deliverable shape); only the MINE-BY-X and GOTCHAS slots and
the task specifics come from recon. See
references/prompt-scaffold.md. Context economy, durable state and budget: the emitted prompt should instruct the fired run to reference large data by path rather than inline, write deliverables incrementally, checkpoint plan/progress to disk at every phase boundary, re-read that durable state after any compaction or resume (never trust transcript memory alone), bound every loop with an explicit breaker (max rounds, dry-round caps, a budget guard), and state the expected phase count so an over-running run notices and re-plans instead of grinding. - Generate the terminal condition (
.goal-condition.txt) per crown rule 2, full contract by default, ≤4000 chars, with overflow degradation. Template + rules:references/terminal-condition.md. Self-audit each emitted artifact against a deterministic checklist, then hand off (do NOT fire):
- the
.goal-condition.txtis <= 4000 bytes (wc -c); - it starts with
/goal; - it contains the literal string
Condition met; - it demands FRESH tool output in the final turn, not just an assertion of state;
- it demands a self-audit table (per-deliverable, criterion, evidence pointer, MET/UNMET);
- it carries an ALT
FINAL STATUS: INCOMPLETEbranch gated on adate -ufloor; - the prompt, condition, and recon files pair by stem (same
<topic>-<phase>-<date>prefix).
Automate this checklist as a validator script if your setup allows it; keep it report-only advisory, never a hard gate, you still hand off and the user still fires. THEN confirm the SOFT judgments no gate can make: the fresh-evidence clause is genuinely evaluator-checkable (not just the word present),
.recon.jsoncarries verify-don't-trust instructions +generated_at, rails match the shape, scaffold holds zero task-specific content. Report a short build-audit and the exact/goalline the user will fire.- the
Common mistakes (from baseline testing)
| Mistake | Fix |
| Light recon / "I didn't enumerate live state, my list may be incomplete" | Crown rule 1: run the grounding fan-out that CORRECTS the scout; emit .recon.json. The recon IS the value. |
| Copying the "fresh results that same turn" wording without understanding why | Crown rule 2: the evaluator can't read files; derive the condition from that, so it holds even with no exemplar to copy. |
| Condition says "the 11 files exist / tests pass" | Uncheckable by the evaluator. Demand FRESH tool output surfaced in one final turn + a self-audit table. |
| Re-deriving conventions / rails each build | Those are codified here and in the scaffold; spend recon budget on codebase traversal, not basics. |
| Firing the goal, or chaining research→implement in one run | Builder never executes. The research→MAP→approve GATE is a SEPARATE human fire (/workflows + the /goal loop have no mid-run input). |
| Inventing scope when the task isn't decided | Stop; point to an upstream brainstorming step. |
Red flags: STOP
- "I'll skip the deep recon, the scout seed is probably fine" → backside errors live exactly here.
- "I'll reference the files by name in the condition" → the evaluator can't see files.
- "I'll fire it to check it works" → never. Hand the artifact to the user.
- About to write task-specific facts into the codified scaffold → those belong in recon slots.
Outputs
Two destinations, do NOT conflate them:
- The prompt artifacts you build → a dedicated prompts directory (e.g.
~/goal-prompts/<topic>-<phase>-<date>.md) +.goal-condition.txt+.recon.json. MAP →<prompts-dir>/<topic>-MAP-<date>.md. Naming matches your existing convention exactly. These are the ONLY files this skill writes. - The deliverables the FIRED goal will later produce (an audit dossier, migrated code, etc.)
go wherever that goal's
## DELIVERsection says (a repo, a worktree branch, or an audit dir), NOT necessarily under the prompts directory. The brief specifies that destination; the skill does not put goal output in the goal-prompt directory.
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.

