Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Execute a Research Spec step by step, respecting context protocol, deviation limits, and backtrack rules. Supports multi-session recovery.
.claude/skills/yogsoth-ai-executing-specs/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | -18% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -30% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 32% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 85% | 0% |
You are executing a Research Spec produced by writing-specs. Your job is to follow the spec stage by stage, invoking the named campaigns and strategies, while strictly respecting the context protocol and execution rules.
The user provides a path to a spec file (e.g., docs/de-anthropocentric/specs/2026-05-20-topic-spec.md). Read it in full before starting.
Before executing, scan the spec for:
If concerns exist, present them to the user before proceeding. Do NOT silently start execution with known issues.
For each Stage in the spec (in order):
Read the context files referenced in "Expected Input". If they don't exist or are incomplete, STOP and ask the user.
Invoke context-init. CC names the file: YYYY-MM-DD-HH-MM-<topic-slug>
For each strategy listed in "Execution Steps":
context-checkpointAfter all strategies in the Stage complete, invoke one final context-checkpoint as a safety net.
Check whether the Stage's completion criteria are met:
[x] in the spec file, proceed to next StageCheck the Stage's backtrack condition:
When entering a new session to continue execution:
[ ] stepcontext/INDEX.md — identify the latest context fileNo special "resume" command needed. The spec's checkbox state IS the progress tracker.
You have limited autonomy to adjust within bounds:
Allowed:
NOT Allowed:
Documentation requirement: Every deviation must be recorded in the next context-checkpoint:
## Deviation from Spec
**Stage**: N
**Prescribed**: [what spec said]
**Actual**: [what you did]
**Rationale**: [why]When a backtrack condition evaluates to true:
"Backtrack condition triggered: condition text]. Recommendation: Return to Stage N (campaign name]). Options: (A) Approve backtrack (B) Continue forward (C) Other"
When all Stages are complete, compile the research results into a summary:
Inform the user: "Research spec execution complete. All context files are in context/."
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 4,457 | 3,504 | -21% | 1 | 1 | 0% | 347 | 1,377 | +297% | 0 | 0 | — |
case-02 | fail→fail | 5,535 | 7,509 | +36% | 1 | 1 | 0% | 371 | 1,183 | +219% | 0 | 0 | — |
case-03 | fail→fail | 6,349 | 5,900 | -7% | 1 | 1 | 0% | 437 | 1,142 | +161% | 0 | 0 | — |
case-04 | pass→pass | 7,478 | 11,759 | +57% | 1 | 1 | 0% | 1,258 | 2,325 | +85% | 0 | 0 | — |
case-05 | pass→pass | 4,164 | 2,985 | -28% | 1 | 1 | 0% | 694 | 1,378 | +99% | 0 | 0 | — |
case-06 | fail→pass | 10,161 | 2,961 | -71% | 1 | 1 | 0% | 1,788 | 1,463 | -18% | 0 | 0 | — |
case-07 | fail→pass | 9,963 | 4,478 | -55% | 1 | 1 | 0% | 1,513 | 1,630 | +8% | 0 | 0 | — |
case-08 | fail→fail | 9,540 | 4,757 | -50% | 1 | 1 | 0% | 1,759 | 1,672 | -5% | 0 | 0 | — |
case-09 | pass→pass | 11,944 | 4,625 | -61% | 1 | 1 | 0% | 1,787 | 1,600 | -10% | 0 | 0 | — |
case-10 | fail→pass | 12,556 | 2,956 | -76% | 1 | 1 | 0% | 1,866 | 1,305 | -30% | 0 | 0 | — |
case-11 | fail→pass | 8,308 | 5,426 | -35% | 1 | 1 | 0% | 1,324 | 1,744 | +32% | 0 | 0 | — |
case-12 | pass→pass | 10,784 | 2,728 | -75% | 1 | 1 | 0% | 1,569 | 1,326 | -15% | 0 | 0 | — |
case-13 | fail→pass | 8,259 | 9,918 | +20% | 1 | 1 | 0% | 1,287 | 2,379 | +85% | 0 | 0 | — |
case-14 | fail→fail | 7,098 | 5,939 | -16% | 1 | 1 | 0% | 1,131 | 1,771 | +57% | 0 | 0 | — |
case-15 | fail→pass | 10,800 | 2,415 | -78% | 1 | 1 | 0% | 1,691 | 1,249 | -26% | 0 | 0 | — |
case-16 | pass→pass | 9,437 | 3,065 | -68% | 1 | 1 | 0% | 1,510 | 1,380 | -9% | 0 | 0 | — |
case-17 | pass→pass | 7,492 | 2,617 | -65% | 1 | 1 | 0% | 1,026 | 1,271 | +24% | 0 | 0 | — |
case-18 | fail→pass | 9,916 | 3,120 | -69% | 1 | 1 | 0% | 1,454 | 1,356 | -7% | 0 | 0 | — |
case-19 | pass→pass | 1,957 | 2,657 | +36% | 1 | 1 | 0% | 357 | 1,332 | +273% | 0 | 0 | — |
case-20 | fail→pass | 9,363 | 3,504 | -63% | 1 | 1 | 0% | 1,502 | 1,457 | -3% | 0 | 0 | — |
case-21 | pass→pass | 20,186 | 34,200 | +69% | 1 | 1 | 0% | 2,241 | 5,910 | +164% | 0 | 0 | — |
case-22 | pass→pass | 8,939 | 6,391 | -29% | 1 | 1 | 0% | 1,759 | 2,050 | +17% | 0 | 0 | — |
case-23 | pass→pass | 11,358 | 11,801 | +4% | 1 | 1 | 0% | 1,965 | 3,081 | +57% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 20 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +35 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.